WIP templating

.gitignore

@@ -11,7 +11,6 @@ __pycache__/
 # tests and logs
 tests/fixtures
 logs/
-lightning_logs/
 
 # Distribution / packaging
 .Python
@@ -140,7 +139,6 @@ runs
 /wandb
 /examples/runs
 /examples/**/*.args
-/examples/rag/sweep
 
 # data
 /data

examples/lightning_base.py

@@ -366,8 +366,6 @@ def generic_train(
     if args.gpus > 1:
         train_params["distributed_backend"] = "ddp"
 
-    train_params["accumulate_grad_batches"] = args.accumulate_grad_batches
-
     trainer = pl.Trainer.from_argparse_args(
         args,
         weights_summary=None,

examples/rag/README.md

@@ -1,85 +0,0 @@
-# Intro
-RAG (for Retrieval Augmented Generation) is a seq2seq model which encapsulates two core components: a question encoder and a generator. During a forward pass, we encode the input with the question encoder and pass it
-to the retriever to extract relevant context documents. The documents are then prepended to the input. Such contextualized input is passed to the generator. See [the paper](https://arxiv.org/pdf/2005.11401.pdf) for mored details.
-
-We implement two variants of the model, both presented in the paper - `RagSequenceForGeneration. and `RagTokenForGeneration`. In both cases we use `DPRQuestionEncoder` as the question encoder. As for the generator, two compatible architectures have been tested: `BartForConditionalGeneration`  and `T5ForConditionalGeneration`.
-
-Key files:
-- `modeling_rag.py`, `tokenization_rag.py`, `configuration_rag.py` the core model implementation
-- `retrieval_rag.py` - a distributed retriever built on top of the `torch.distributed` communication package. The retriever is an interface between the model and the faiss index of the encoded documents. During training, all workers initialize their own instance of the retriever, however, only the main worker loads the index into memory, which prevents OOMs on machines with multiple GPUs (we store the index in RAM). The index itself is based on the `nlp.Datasets`. We also implement a variant compatible with indices built using the original DPR implementation (https://github.com/facebookresearch/DPR)
-- `eval_rag.py` - an evaluation script which allows to perform the evaluation end to end (measures the exact match and F1 on the downstream task) as well as the evaluation of the retrieval component alone (measures precision@k).
-- `finetune.py` - a training script for finetuning RAG models.
-
-
-# Finetuning
-Our finetuning logic is based on scripts from [`examples/seq2seq`](https://github.com/huggingface/transformers/tree/master/examples/seq2seq).
-Follow instructions there regarding data preprocessing. A sample finetuning command:
-
-```
-python examples/rag/finetune.py \
-    --data_dir $DATA_DIR \
-    --output_dir $OUTPUT_DIR \
-    --model_name_or_path $MODEL_NAME_OR_PATH \
-    --model_type rag_sequence \
-    --fp16 \
-    --gpus 8
-```
-
-
-# Evaluation
-Apart for parameters specifying the model that's being evaluated and some extra parameters, the evaluation script expects paths to two files:
-- `evaluation_set` - a path file specifying the input dataset for evaluation, a single datapoint per line, e.g.
-```who is the owner of reading football club```
-- `gold_data_path` - a path to a file contaning ground truth answers for samples from the `evaluation_set`.
-
-We expect the following formats of the gold data file:
-
-- for e2e evaluation, we support two formats of gold files:
-    - `qa` - where a single line in the following format: input [tab] output_list, e.g.:
-    ```
-    who is the owner of reading football club	['Xiu Li Dai', 'Dai Yongge', 'Dai Xiuli', 'Yongge Dai']
-    ```
-    - `ans` - where a single line of the gold file contains the expected output string,
-    ```
-    Xiu Li Dai
-    ```
-
-- for retrieval evaluation, we expect a tab-separated list of Wikipedia page titles constituting positive contexts for a given query, e.g. given a question `who sings does he love me with reba`, a line with ground truth retrieval data could look as follows:
-```
-Does He Love You	Does He Love You	Red Sandy Spika dress of Reba McEntire	Greatest Hits Volume Two (Reba McEntire album)	Shoot for the Moon (album)
-```
-
-## Retrieval evaluation
-
-We demonstrate how to evaluate retrieval against DPR evaluation data. You can download respective files from links listed [here](https://github.com/facebookresearch/DPR/blob/master/data/download_data.py#L39-L45).
-
-1. Download and unzip the gold data file. We use the `biencoder-nq-dev` from https://dl.fbaipublicfiles.com/dpr/data/retriever/biencoder-nq-dev.json.gz.
-2. Parse the unziped file using the `parse_dpr_relevance_data.py`
-```
-python examples/rag/parse_dpr_relevance_data.py --src_path path/to/unziped/biencoder-nq-dev.json --evaluation_set path/to/output/biencoder-nq-dev.questions --gold_data_path path/to/output/biencoder-nq-dev.pages
-```
-3. Run evaluation:
-```
-python examples/rag/eval_rag.py \
-    --model_name_or_path $MODEL_NAME_OR_PATH \ # model name or path of the model we're evaluating
-    --model_type rag_sequence \ # RAG model type (rag_token or rag_sequence)
-    --evaluation_set path/to/output/biencoder-nq-dev.questions \ # an input dataset for evaluation
-    --gold_data_path path/to/output/biencoder-nq-dev.pages \ # a dataset containing ground truth answers for samples from the evaluation_set
-    --predictions_filename retrieval_preds.tsv  \ # name of file in which predictions will be stored
-    --eval_mode retrieval  \ # indicates whether we're performing retrieval evaluation or e2e evaluation
-    --recalculate # if predictions_filename already exists, and this option is set - we regenerate the answers, otherwise we reuse the predicsion file to calculate metrics.
-```
-
-
-## End-to-end evaluation
-```
-python examples/rag/eval_rag.py \
-    --model_name_or_path /private/home/piktus/rag_huggingface/data/repro-rag-sequence-63/ \
-    --model_type rag_sequence \
-    --evaluation_set path/to/test.source \
-    --gold_data_path path/to/gold_data \
-    --predictions_filename e2e_preds.txt \
-    --eval_mode e2e  \ # indicates whether we're performing retrieval evaluation or e2e evaluation (default)
-    --n_docs 5 \ # You can experiment with retrieving different number of documents at evaluation time
-    --print_predictions
-```

examples/rag/callbacks.py

@@ -1,30 +0,0 @@
-import logging
-import os
-
-from pytorch_lightning.callbacks import ModelCheckpoint
-
-
-logger = logging.getLogger(__name__)
-
-
-def get_checkpoint_callback(output_dir, metric):
-    """Saves the best model by validation ROUGE2 score."""
-    if metric == "rouge2":
-        exp = "{val_avg_rouge2:.4f}-{step_count}"
-    elif metric == "bleu":
-        exp = "{val_avg_bleu:.4f}-{step_count}"
-    elif metric == "em":
-        exp = "{val_avg_em:.4f}-{step_count}"
-    else:
-        raise NotImplementedError(
-            f"seq2seq callbacks only support rouge2 and bleu, got {metric}, You can make your own by adding to this function."
-        )
-
-    checkpoint_callback = ModelCheckpoint(
-        filepath=os.path.join(output_dir, exp),
-        monitor=f"val_{metric}",
-        mode="max",
-        save_top_k=10,
-        period=0,  # maybe save a checkpoint every time val is run, not just end of epoch.
-    )
-    return checkpoint_callback

examples/rag/eval_rag.py

@@ -1,317 +0,0 @@
-""" Evaluation script for RAG models."""
-
-import argparse
-import ast
-import logging
-import os
-import sys
-
-import pandas as pd
-import torch
-from tqdm import tqdm
-
-from transformers import (
-    BartForConditionalGeneration,
-    BartTokenizer,
-    RagRetriever,
-    RagSequenceForGeneration,
-    RagTokenForGeneration,
-)
-from transformers import logging as transformers_logging
-
-
-sys.path.append(os.path.join(os.getcwd()))  # noqa: E402 # isort:skip
-from examples.rag.utils import exact_match_score, f1_score  # noqa: E402 # isort:skip
-
-
-logger = logging.getLogger(__name__)
-logging.basicConfig(level=logging.INFO)
-
-transformers_logging.set_verbosity_info()
-
-
-def infer_model_type(model_name_or_path):
-    if "token" in model_name_or_path:
-        return "rag_token"
-    if "sequence" in model_name_or_path:
-        return "rag_sequence"
-    if "bart" in model_name_or_path:
-        return "bart"
-    return None
-
-
-def metric_max_over_ground_truths(metric_fn, prediction, ground_truths):
-    scores_for_ground_truths = []
-    for ground_truth in ground_truths:
-        score = metric_fn(prediction, ground_truth)
-        scores_for_ground_truths.append(score)
-    return max(scores_for_ground_truths)
-
-
-def get_scores(args, preds_path, gold_data_path):
-    hypos = [line.strip() for line in open(preds_path, "r").readlines()]
-    answers = []
-
-    if args.gold_data_mode == "qa":
-        data = pd.read_csv(gold_data_path, sep="\t", header=None)
-        for answer_list in data[1]:
-            ground_truths = ast.literal_eval(answer_list)
-            answers.append(ground_truths)
-    else:
-        references = [line.strip() for line in open(gold_data_path, "r").readlines()]
-        answers = [[reference] for reference in references]
-
-    f1 = em = total = 0
-    for prediction, ground_truths in zip(hypos, answers):
-        total += 1
-        em += metric_max_over_ground_truths(exact_match_score, prediction, ground_truths)
-        f1 += metric_max_over_ground_truths(f1_score, prediction, ground_truths)
-
-    em = 100.0 * em / total
-    f1 = 100.0 * f1 / total
-
-    logger.info("F1: {}".format(f1))
-    logger.info("EM: {}".format(em))
-
-
-def get_precision_at_k(args, preds_path, gold_data_path):
-    k = args.k
-    hypos = [line.strip() for line in open(preds_path, "r").readlines()]
-    references = [line.strip() for line in open(gold_data_path, "r").readlines()]
-
-    em = total = 0
-    for hypo, reference in zip(hypos, references):
-        hypo_provenance = set(hypo.split("\t")[:k])
-        ref_provenance = set(reference.split("\t")[1 : (k + 1)])
-        total += 1
-        em += len(hypo_provenance & ref_provenance) / k
-
-    em = 100.0 * em / total
-    logger.info("Precision@{}: {}".format(k, em))
-
-
-def evaluate_batch_retrieval(args, rag_model, tokenizer, retriever, questions):
-    def strip_title(title):
-        if title.startswith('"'):
-            title = title[1:]
-        if title.endswith('"'):
-            title = title[:-1]
-        return title
-
-    retriever_inputs = tokenizer.batch_encode_plus(
-        questions,
-        return_tensors="pt",
-        padding=True,
-        truncation=True,
-    )
-    retriever_input_embs = rag_model.model.question_encoder(retriever_inputs["input_ids"].to(args.device))[0]
-
-    _, all_docs = retriever.retrieve(retriever_input_embs.numpy(), rag_model.config.n_docs)
-
-    provenance_strings = []
-    for docs in all_docs:
-        provenance = [strip_title(title) for title in docs["title"]]
-        provenance_strings.append("\t".join(provenance))
-    return provenance_strings
-
-
-def evaluate_batch_e2e(args, rag_model, tokenizer, retriever, questions):
-    with torch.no_grad():
-        input_ids = tokenizer.batch_encode_plus(questions, return_tensors="pt", padding=True, truncation=True)[
-            "input_ids"
-        ].to(args.device)
-        outputs = rag_model.generate(
-            input_ids,
-            retriever=retriever,
-            num_beams=args.num_beams,
-            min_length=args.min_length,
-            max_length=args.max_length,
-            early_stopping=False,
-            num_return_sequences=1,
-            bad_words_ids=[[0, 0]],  # BART likes to repeat BOS tokens, dont allow it to generate more than one
-            clean_up_tokenization=True,
-            print_docs=args.print_docs,
-        )
-        answers = tokenizer.batch_decode(outputs, skip_special_tokens=True)
-
-        if args.print_predictions:
-            for q, a in zip(questions, answers):
-                logger.info("Q: {} - A: {}".format(q, a))
-
-        return answers
-
-
-def get_args():
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--model_type",
-        choices=["rag_sequence", "rag_token", "bart"],
-        type=str,
-        help="RAG model type: rag_sequence, rag_token or bart, if none specified, the type is inferred from the model_name_or_path",
-    )
-    parser.add_argument(
-        "--retriever_type",
-        default=None,
-        choices=["hf_retriever", "legacy_retriever"],
-        type=str,
-        help="RAG model retriever type",
-    )
-    parser.add_argument(
-        "--index_path",
-        default=None,
-        type=str,
-        help="Path to the retrieval index",
-    )
-    parser.add_argument("--n_docs", default=5, type=int, help="Number of retrieved docs")
-    parser.add_argument(
-        "--model_name_or_path",
-        default=None,
-        type=str,
-        required=True,
-        help="Path to pretrained checkpoints or model identifier from huggingface.co/models",
-    )
-    parser.add_argument(
-        "--eval_mode",
-        choices=["e2e", "retrieval"],
-        default="e2e",
-        type=str,
-        help="Evaluation mode, e2e calculates exact match and F1 of the downstream task, retrieval calulates precision@k.",
-    )
-    parser.add_argument("--k", default=1, type=int, help="k for the precision@k calculation")
-    parser.add_argument(
-        "--evaluation_set",
-        default=None,
-        type=str,
-        required=True,
-        help="Path to a file containing evaluation samples",
-    )
-    parser.add_argument(
-        "--gold_data_path",
-        default=None,
-        type=str,
-        required=True,
-        help="Path to a tab-separated file with gold samples",
-    )
-    parser.add_argument(
-        "--gold_data_mode",
-        default="qa",
-        type=str,
-        choices=["qa", "ans"],
-        help="Format of the gold data file"
-        "qa - a single line in the following format: question [tab] answer_list"
-        "ans - a single line of the gold file contains the expected answer string",
-    )
-    parser.add_argument(
-        "--predictions_filename",
-        type=str,
-        default="predictions.txt",
-        help="Name of the predictions file, to be stored in the checkpoints directry",
-    )
-    parser.add_argument(
-        "--eval_all_checkpoints",
-        action="store_true",
-        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
-    )
-    parser.add_argument(
-        "--eval_batch_size",
-        default=8,
-        type=int,
-        help="Batch size per GPU/CPU for evaluation.",
-    )
-    parser.add_argument(
-        "--recalculate",
-        help="Recalculate predictions even if the prediction file exists",
-        action="store_true",
-    )
-    parser.add_argument(
-        "--num_beams",
-        default=4,
-        type=int,
-        help="Number of beams to be used when generating answers",
-    )
-    parser.add_argument("--min_length", default=1, type=int, help="Min length of the generated answers")
-    parser.add_argument("--max_length", default=50, type=int, help="Max length of the generated answers")
-
-    parser.add_argument(
-        "--print_predictions",
-        action="store_true",
-        help="If True, prints predictions while evaluating.",
-    )
-    parser.add_argument(
-        "--print_docs",
-        action="store_true",
-        help="If True, prints docs retried while generating.",
-    )
-    args = parser.parse_args()
-    args.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-    return args
-
-
-def main(args):
-    model_kwargs = {}
-    if args.model_type is None:
-        args.model_type = infer_model_type(args.model_name_or_path)
-        assert args.model_type is not None
-    if args.model_type.startswith("rag"):
-        model_class = RagTokenForGeneration if args.model_type == "rag_token" else RagSequenceForGeneration
-        model_kwargs["n_docs"] = args.n_docs
-        if args.retriever_type is not None:
-            model_kwargs["retriever_type"] = args.retriever_type
-        if args.index_path is not None:
-            model_kwargs["index_path"] = args.index_path
-    else:
-        model_class = BartForConditionalGeneration
-
-    checkpoints = (
-        [f.path for f in os.scandir(args.model_name_or_path) if f.is_dir()]
-        if args.eval_all_checkpoints
-        else [args.model_name_or_path]
-    )
-
-    logger.info("Evaluate the following checkpoints: %s", checkpoints)
-
-    score_fn = get_scores if args.eval_mode == "e2e" else get_precision_at_k
-    evaluate_batch_fn = evaluate_batch_e2e if args.eval_mode == "e2e" else evaluate_batch_retrieval
-
-    for checkpoint in checkpoints:
-        predictions_path = os.path.join(checkpoint, args.predictions_filename)
-        if os.path.exists(predictions_path) and (not args.recalculate):
-            logger.info("Calculating metrics based on an existing predictions file: {}".format(predictions_path))
-            score_fn(args, predictions_path, args.gold_data_path)
-            continue
-
-        logger.info("***** Running evaluation for {} *****".format(checkpoint))
-        logger.info("  Batch size = %d", args.eval_batch_size)
-        logger.info("  Predictions will be stored under {}".format(predictions_path))
-
-        model = model_class.from_pretrained(checkpoint, **model_kwargs)
-        model.to(args.device)
-        retriever = RagRetriever(model.config)  # TODO: add tokenizers
-        tokenizer = (
-            retriever.generator_tokenizer
-            if args.model_type != "bart" and args.eval_mode == "e2e"
-            else retriever.question_encoder_tokenizer
-            if args.model_type != "bart" and args.eval_mode == "retrieval"
-            else BartTokenizer.from_pretrained("facebook/bart-large")
-        )
-
-        with open(args.evaluation_set, "r") as eval_file, open(predictions_path, "w") as preds_file:
-            questions = []
-            for line in tqdm(eval_file):
-                questions.append(line.strip())
-                if len(questions) == args.eval_batch_size:
-                    answers = evaluate_batch_fn(args, model, tokenizer, retriever, questions)
-                    preds_file.write("\n".join(answers) + "\n")
-                    preds_file.flush()
-                    questions = []
-            if len(questions) > 0:
-                answers = evaluate_batch_fn(args, model, tokenizer, retriever, questions)
-                preds_file.write("\n".join(answers))
-                preds_file.flush()
-
-            score_fn(args, predictions_path, args.gold_data_path)
-
-
-if __name__ == "__main__":
-    args = get_args()
-    main(args)

examples/rag/finetune.py

@@ -1,476 +0,0 @@
-"""Finetuning script for RAG models. Adapted from examples.seq2seq.finetune.py"""
-
-import argparse
-import glob
-import logging
-import os
-import sys
-import time
-import warnings
-from collections import defaultdict
-from pathlib import Path
-from typing import Any, Dict, List, Tuple
-
-import numpy as np
-import pytorch_lightning as pl
-import torch
-import torch.distributed as dist
-from torch.utils.data import DataLoader
-
-from transformers import (
-    AutoConfig,
-    AutoTokenizer,
-    BartForConditionalGeneration,
-    RagConfig,
-    RagPyTorchDistributedRetriever,
-    RagSequenceForGeneration,
-    RagTokenForGeneration,
-    T5ForConditionalGeneration,
-    get_linear_schedule_with_warmup,
-)
-from transformers import logging as transformers_logging
-
-
-sys.path.append(os.path.join(os.getcwd()))  # noqa: E402 # noqa: E402 # isort:skip
-
-from examples.lightning_base import BaseTransformer, add_generic_args, generic_train  # noqa: E402 # isort:skip
-from examples.rag.callbacks import get_checkpoint_callback  # noqa: E402 # isort:skip
-from examples.rag.utils import (  # noqa: E402 # isort:skip
-    Seq2SeqDataset,
-    calculate_exact_match,
-    is_rag_model,
-    set_extra_model_params,
-)
-from examples.seq2seq.callbacks import Seq2SeqLoggingCallback, get_early_stopping_callback  # noqa: E402 # isort:skip
-from examples.seq2seq.utils import (  # noqa: E402 # isort:skip
-    flatten_list,
-    get_git_info,
-    lmap,
-    pickle_save,
-    save_git_info,
-    save_json,
-)
-
-logging.basicConfig(level=logging.INFO)
-logger = logging.getLogger(__name__)
-
-transformers_logging.set_verbosity_info()
-
-
-class AttrDict(dict):
-    def __init__(self, *args, **kwargs):
-        super(AttrDict, self).__init__(*args, **kwargs)
-        self.__dict__ = self
-
-
-class GenerativeQAModule(BaseTransformer):
-    mode = "generative_qa"
-    loss_names = ["loss"]
-    metric_names = ["em"]
-    val_metric = "em"
-
-    def __init__(self, hparams, **kwargs):
-        # when loading from a pytorch lightning checkpoint, hparams are passed as dict
-        if isinstance(hparams, dict):
-            hparams = AttrDict(hparams)
-        if hparams.model_type == "rag_sequence":
-            self.model_class = RagSequenceForGeneration
-        elif hparams.model_type == "rag_token":
-            self.model_class = RagTokenForGeneration
-        elif hparams.model_type == "bart":
-            self.model_class = BartForConditionalGeneration
-        else:
-            self.model_class = T5ForConditionalGeneration
-        self.is_rag_model = is_rag_model(hparams.model_type)
-
-        config_class = RagConfig if self.is_rag_model else AutoConfig
-        config = config_class.from_pretrained(hparams.model_name_or_path)
-
-        # set extra_model_params for generator configs and load_model
-        extra_model_params = ("encoder_layerdrop", "decoder_layerdrop", "attention_dropout", "dropout")
-        if self.is_rag_model:
-            generator_config = AutoConfig.from_pretrained(
-                config.pretrained_generator_name_or_path, prefix=config.prefix
-            )
-            hparams, generator_config = set_extra_model_params(extra_model_params, hparams, generator_config)
-            model = self.model_class.from_pretrained(
-                hparams.model_name_or_path, config=config, generator_config=generator_config
-            )
-        else:
-            if args.prefix is not None:
-                setattr(config, "prefix", args.prefix)
-            hparams, config = set_extra_model_params(extra_model_params, hparams, config)
-            model = self.model_class.from_pretrained(hparams.model_name_or_path, config=config)
-            generator_config = config
-
-        tokenizer = (
-            AutoTokenizer.from_pretrained(config.pretrained_generator_tokenizer_name_or_path)
-            if self.is_rag_model
-            else AutoTokenizer.from_pretrained(hparams.model_name_or_path)
-        )
-
-        super().__init__(hparams, config=config, tokenizer=tokenizer, model=model)
-
-        self.retriever = (
-            RagPyTorchDistributedRetriever(self.model.config) if self.is_rag_model else None
-        )  # TODO add tokenizers
-
-        save_git_info(self.hparams.output_dir)
-        self.output_dir = Path(self.hparams.output_dir)
-        self.metrics_save_path = Path(self.output_dir) / "metrics.json"
-        self.hparams_save_path = Path(self.output_dir) / "hparams.pkl"
-        pickle_save(self.hparams, self.hparams_save_path)
-        self.step_count = 0
-        self.metrics = defaultdict(list)
-
-        self.dataset_kwargs: dict = dict(
-            data_dir=self.hparams.data_dir,
-            max_source_length=self.hparams.max_source_length,
-            prefix=generator_config.prefix or "",
-        )
-        n_observations_per_split = {
-            "train": self.hparams.n_train,
-            "val": self.hparams.n_val,
-            "test": self.hparams.n_test,
-        }
-        self.n_obs = {k: v if v >= 0 else None for k, v in n_observations_per_split.items()}
-
-        self.target_lens = {
-            "train": self.hparams.max_target_length,
-            "val": self.hparams.val_max_target_length,
-            "test": self.hparams.test_max_target_length,
-        }
-        assert self.target_lens["train"] <= self.target_lens["val"], f"target_lens: {self.target_lens}"
-        assert self.target_lens["train"] <= self.target_lens["test"], f"target_lens: {self.target_lens}"
-
-        self.hparams.git_sha = get_git_info()["repo_sha"]
-        self.num_workers = hparams.num_workers
-        self.distributed_port = self.hparams.distributed_port
-
-    def init_ddp_connection(self, global_rank: int, world_size: int, is_slurm_managing_tasks: bool = True):
-        logger.info("Custom init_ddp_connection.")
-        os.environ["MASTER_PORT"] = str(self.distributed_port)
-        super().init_ddp_connection(global_rank, world_size, is_slurm_managing_tasks)
-        if self.is_rag_model:
-            self.retriever.init_retrieval(self.distributed_port)
-
-    def forward(self, input_ids, **kwargs):
-        return self.model(input_ids, **kwargs)
-
-    def ids_to_clean_text(self, generated_ids: List[int]):
-        gen_text = self.tokenizer.batch_decode(
-            generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True
-        )
-        return lmap(str.strip, gen_text)
-
-    def _step(self, batch: dict) -> Tuple:
-        source_ids, source_mask, target_ids = batch["input_ids"], batch["attention_mask"], batch["decoder_input_ids"]
-
-        if isinstance(self.model, T5ForConditionalGeneration):
-            decoder_input_ids = self.model._shift_right(target_ids)
-            lm_labels = target_ids
-        elif isinstance(self.model, BartForConditionalGeneration):
-            decoder_input_ids = target_ids[:, :-1].contiguous()
-            lm_labels = target_ids[:, 1:].clone()
-        else:
-            assert self.is_rag_model
-            generator = self.model.model.generator
-            if isinstance(generator, T5ForConditionalGeneration):
-                decoder_start_token_id = generator.config.decoder_start_token_id
-                decoder_input_ids = (
-                    torch.cat(
-                        [torch.Tensor([[decoder_start_token_id]] * target_ids.shape[0]).to(target_ids), target_ids],
-                        dim=1,
-                    )
-                    if target_ids.shape[0] < self.target_lens["train"]
-                    else generator._shift_right(target_ids)
-                )
-            elif isinstance(generator, BartForConditionalGeneration):
-                decoder_input_ids = target_ids
-            lm_labels = None
-
-        assert decoder_input_ids is not None
-
-        if lm_labels is not None:
-            outputs = self(
-                source_ids,
-                attention_mask=source_mask,
-                decoder_input_ids=decoder_input_ids,
-                use_cache=False,
-                labels=lm_labels,
-                return_dict=True,
-            )
-        else:  # RAG models
-            outputs = self(
-                source_ids,
-                retriever=self.retriever,
-                attention_mask=source_mask,
-                decoder_input_ids=decoder_input_ids,
-                use_cache=False,
-                return_loss=True,
-                reduce=True,
-                label_smoothing=self.hparams.label_smoothing,
-            )
-
-        loss = outputs["loss"]
-        return (loss,)
-
-    @property
-    def pad(self) -> int:
-        return self.tokenizer.pad_token_id
-
-    def training_step(self, batch, batch_idx) -> Dict:
-        loss_tensors = self._step(batch)
-
-        logs = {name: loss for name, loss in zip(self.loss_names, loss_tensors)}
-        # tokens per batch
-        logs["tpb"] = batch["input_ids"].ne(self.pad).sum() + batch["decoder_input_ids"].ne(self.pad).sum()
-
-        return {"loss": loss_tensors[0], "log": logs}
-
-    def validation_step(self, batch, batch_idx) -> Dict:
-        return self._generative_step(batch)
-
-    def validation_epoch_end(self, outputs, prefix="val") -> Dict:
-        self.step_count += 1
-        losses = {k: torch.stack([x[k] for x in outputs]).mean() for k in self.loss_names}
-        loss = losses["loss"]
-        gen_metrics = {
-            k: np.array([x[k] for x in outputs]).mean() for k in self.metric_names + ["gen_time", "gen_len"]
-        }
-        metrics_tensor: torch.FloatTensor = torch.tensor(gen_metrics[self.val_metric]).type_as(loss)
-        gen_metrics.update({k: v.item() for k, v in losses.items()})
-
-        # fix for https://github.com/PyTorchLightning/pytorch-lightning/issues/2424
-        if dist.is_initialized():
-            dist.all_reduce(metrics_tensor, op=dist.ReduceOp.SUM)
-            metrics_tensor = metrics_tensor / dist.get_world_size()
-            gen_metrics.update({self.val_metric: metrics_tensor.item()})
-
-        losses.update(gen_metrics)
-        metrics = {f"{prefix}_avg_{k}": x for k, x in losses.items()}
-        metrics["step_count"] = self.step_count
-        self.save_metrics(metrics, prefix)  # writes to self.metrics_save_path
-        preds = flatten_list([x["preds"] for x in outputs])
-        return {"log": metrics, "preds": preds, f"{prefix}_loss": loss, f"{prefix}_{self.val_metric}": metrics_tensor}
-
-    def save_metrics(self, latest_metrics, type_path) -> None:
-        self.metrics[type_path].append(latest_metrics)
-        save_json(self.metrics, self.metrics_save_path)
-
-    def calc_generative_metrics(self, preds, target) -> Dict:
-        return calculate_exact_match(preds, target)
-
-    def _generative_step(self, batch: dict) -> dict:
-        start_time = time.time()
-        generated_ids = self.model.generate(
-            batch["input_ids"],
-            retriever=self.retriever,
-            dedup=False,  # rag specific parameter
-            attention_mask=batch["attention_mask"],
-            use_cache=True,
-            min_length=1,
-            max_length=self.target_lens["val"],
-        )
-
-        gen_time = (time.time() - start_time) / batch["input_ids"].shape[0]
-        preds: List[str] = self.ids_to_clean_text(generated_ids)
-        target: List[str] = self.ids_to_clean_text(batch["decoder_input_ids"])
-        loss_tensors = self._step(batch)
-        base_metrics = {name: loss for name, loss in zip(self.loss_names, loss_tensors)}
-        gen_metrics: Dict = self.calc_generative_metrics(preds, target)
-
-        summ_len = np.mean(lmap(len, generated_ids))
-        base_metrics.update(gen_time=gen_time, gen_len=summ_len, preds=preds, target=target, **gen_metrics)
-        return base_metrics
-
-    def test_step(self, batch, batch_idx):
-        return self._generative_step(batch)
-
-    def test_epoch_end(self, outputs):
-        return self.validation_epoch_end(outputs, prefix="test")
-
-    def get_dataset(self, type_path) -> Seq2SeqDataset:
-        n_obs = self.n_obs[type_path]
-        max_target_length = self.target_lens[type_path]
-        dataset = Seq2SeqDataset(
-            self.tokenizer,
-            type_path=type_path,
-            n_obs=n_obs,
-            max_target_length=max_target_length,
-            **self.dataset_kwargs,
-        )
-        return dataset
-
-    def get_dataloader(self, type_path: str, batch_size: int, shuffle: bool = False) -> DataLoader:
-        dataset = self.get_dataset(type_path)
-        sampler = None
-        if self.hparams.sortish_sampler and type_path == "train":
-            assert self.hparams.gpus <= 1  # TODO: assert earlier
-            sampler = dataset.make_sortish_sampler(batch_size)
-            shuffle = False
-
-        dataloader = DataLoader(
-            dataset,
-            batch_size=batch_size,
-            collate_fn=dataset.collate_fn,
-            shuffle=shuffle,
-            num_workers=self.num_workers,
-            sampler=sampler,
-        )
-        return dataloader
-
-    def train_dataloader(self) -> DataLoader:
-        dataloader = self.get_dataloader("train", batch_size=self.hparams.train_batch_size, shuffle=True)
-        t_total = (
-            (len(dataloader.dataset) // (self.hparams.train_batch_size * max(1, self.hparams.gpus)))
-            // self.hparams.accumulate_grad_batches
-            * float(self.hparams.max_epochs)
-        )
-        scheduler = get_linear_schedule_with_warmup(
-            self.opt, num_warmup_steps=self.hparams.warmup_steps, num_training_steps=t_total
-        )
-        if max(scheduler.get_last_lr()) > 0:
-            warnings.warn("All learning rates are 0")
-        self.lr_scheduler = scheduler
-        return dataloader
-
-    def val_dataloader(self) -> DataLoader:
-        return self.get_dataloader("val", batch_size=self.hparams.eval_batch_size)
-
-    def test_dataloader(self) -> DataLoader:
-        return self.get_dataloader("test", batch_size=self.hparams.eval_batch_size)
-
-    @pl.utilities.rank_zero_only
-    def on_save_checkpoint(self, checkpoint: Dict[str, Any]) -> None:
-        save_path = self.output_dir.joinpath("checkpoint{}".format(self.step_count))
-        self.model.config.save_step = self.step_count
-        self.model.save_pretrained(save_path)
-        self.tokenizer.save_pretrained(save_path)
-
-    @staticmethod
-    def add_model_specific_args(parser, root_dir):
-        BaseTransformer.add_model_specific_args(parser, root_dir)
-        add_generic_args(parser, root_dir)
-        parser.add_argument(
-            "--max_source_length",
-            default=128,
-            type=int,
-            help="The maximum total input sequence length after tokenization. Sequences longer "
-            "than this will be truncated, sequences shorter will be padded.",
-        )
-        parser.add_argument(
-            "--max_target_length",
-            default=25,
-            type=int,
-            help="The maximum total input sequence length after tokenization. Sequences longer "
-            "than this will be truncated, sequences shorter will be padded.",
-        )
-        parser.add_argument(
-            "--val_max_target_length",
-            default=25,
-            type=int,
-            help="The maximum total input sequence length after tokenization. Sequences longer "
-            "than this will be truncated, sequences shorter will be padded.",
-        )
-        parser.add_argument(
-            "--test_max_target_length",
-            default=25,
-            type=int,
-            help="The maximum total input sequence length after tokenization. Sequences longer "
-            "than this will be truncated, sequences shorter will be padded.",
-        )
-        parser.add_argument("--sortish_sampler", action="store_true", default=False)
-        parser.add_argument("--logger_name", type=str, choices=["default", "wandb", "wandb_shared"], default="default")
-        parser.add_argument("--n_train", type=int, default=-1, required=False, help="# examples. -1 means use all.")
-        parser.add_argument("--n_val", type=int, default=-1, required=False, help="# examples. -1 means use all.")
-        parser.add_argument("--n_test", type=int, default=-1, required=False, help="# examples. -1 means use all.")
-        parser.add_argument("--label_smoothing", type=float, default=0.0, required=False)
-        parser.add_argument(
-            "--prefix",
-            type=str,
-            default=None,
-            help="Prefix added at the beginning of each text, typically used with T5-based models.",
-        )
-        parser.add_argument(
-            "--early_stopping_patience",
-            type=int,
-            default=-1,
-            required=False,
-            help="-1 means never early stop. early_stopping_patience is measured in validation checks, not epochs. So val_check_interval will effect it.",
-        )
-        parser.add_argument(
-            "--distributed-port", type=int, default=-1, required=False, help="Port number for distributed training."
-        )
-        parser.add_argument(
-            "--model_type",
-            choices=["rag_sequence", "rag_token", "bart", "t5"],
-            type=str,
-            help="RAG model type: sequence or token, if none specified, the type is inferred from the model_name_or_path",
-        )
-        return parser
-
-
-def main(args, model=None) -> GenerativeQAModule:
-    Path(args.output_dir).mkdir(exist_ok=True)
-    if model is None:
-        model: GenerativeQAModule = GenerativeQAModule(args)
-
-    dataset = Path(args.data_dir).name
-    if (
-        args.logger_name == "default"
-        or args.fast_dev_run
-        or str(args.output_dir).startswith("/tmp")
-        or str(args.output_dir).startswith("/var")
-    ):
-        logger = True  # don't pollute wandb logs unnecessarily
-    elif args.logger_name == "wandb":
-        from pytorch_lightning.loggers import WandbLogger
-
-        project = os.environ.get("WANDB_PROJECT", dataset)
-        logger = WandbLogger(name=model.output_dir.name, project=project)
-
-    elif args.logger_name == "wandb_shared":
-        from pytorch_lightning.loggers import WandbLogger
-
-        logger = WandbLogger(name=model.output_dir.name, project=f"hf_{dataset}")
-
-    es_callback = (
-        get_early_stopping_callback(model.val_metric, args.early_stopping_patience)
-        if args.early_stopping_patience >= 0
-        else False
-    )
-    trainer: pl.Trainer = generic_train(
-        model,
-        args,
-        logging_callback=Seq2SeqLoggingCallback(),
-        checkpoint_callback=get_checkpoint_callback(args.output_dir, model.val_metric),
-        early_stopping_callback=es_callback,
-        logger=logger,
-    )
-    pickle_save(model.hparams, model.output_dir / "hparams.pkl")
-
-    if not args.do_predict:
-        return model
-
-    model.hparams.test_checkpoint = ""
-    checkpoints = list(sorted(glob.glob(os.path.join(args.output_dir, "*.ckpt"), recursive=True)))
-    if checkpoints:
-        model.hparams.test_checkpoint = checkpoints[-1]
-        trainer.resume_from_checkpoint = checkpoints[-1]  # best checkpoint
-    trainer.logger.log_hyperparams(model.hparams)
-
-    # test() without a model tests using the best checkpoint automatically
-    trainer.test()
-
-    return model
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser = pl.Trainer.add_argparse_args(parser)
-    parser = GenerativeQAModule.add_model_specific_args(parser, os.getcwd())
-
-    args = parser.parse_args()
-
-    main(args)

examples/rag/finetune.sh

@@ -1,34 +0,0 @@
-# Add parent directory to python path to access lightning_base.py
-export PYTHONPATH="../":"${PYTHONPATH}"
-
-# A sample finetuning run, you need to specify data_dir, output_dir and model_name_or_path
-# run ./examples/rag/finetune.sh --help to see all the possible options
-
-python examples/rag/finetune.py \
-    --data_dir $DATA_DIR \
-    --output_dir $OUTPUT_DIR \
-    --model_name_or_path $MODLE_NAME_OR_PATH \
-    --model_type rag_sequence \
-    --fp16 \
-    --gpus 8 \
-    --do_train \
-    --do_predict \
-    --n_val -1 \
-    --val_check_interval 0.25 \
-    --train_batch_size 8 \
-    --eval_batch_size 1 \
-    --max_source_length 128 \
-    --max_target_length 25 \
-    --val_max_target_length 25 \
-    --test_max_target_length 25 \
-    --label_smoothing 0.1 \
-    --dropout 0.1 \
-    --attention_dropout 0.1 \
-    --weight_decay 0.001 \
-    --adam_epsilon 1e-08 \
-    --max_grad_norm 0.1 \
-    --lr_scheduler polynomial \
-    --learning_rate 3e-05 \
-    --num_train_epochs 100 \
-    --warmup_steps 500 \
-    --gradient_accumulation_steps 1

examples/rag/parse_dpr_relevance_data.py

@@ -1,47 +0,0 @@
-"""
-This script reads DPR retriever training data and parses each datapoint. We save a line per datapoint.
-Each line consists of the query followed by a tab-separated list of Wikipedia page titles constituting
-positive contexts for a given query.
-"""
-
-import argparse
-import json
-
-from tqdm import tqdm
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    # Required parameters
-    parser.add_argument(
-        "--src_path",
-        type=str,
-        default="biencoder-nq-dev.json",
-        help="Path to raw DPR training data",
-    )
-    parser.add_argument(
-        "--evaluation_set",
-        type=str,
-        help="where to store parsed evaluation_set file",
-    )
-    parser.add_argument(
-        "--gold_data_path",
-        type=str,
-        help="where to store parsed gold_data_path file",
-    )
-    args = parser.parse_args()
-
-    with open(args.src_path, "r") as src_file, open(args.evaluation_set, "w") as eval_file, open(
-        args.gold_data_path, "w"
-    ) as gold_file:
-        dpr_records = json.load(src_file)
-        for dpr_record in tqdm(dpr_records):
-            question = dpr_record["question"]
-            contexts = [context["title"] for context in dpr_record["positive_ctxs"]]
-            eval_file.write(question + "\n")
-            gold_file.write("\t".join(contexts) + "\n")
-
-
-if __name__ == "__main__":
-    main()

examples/rag/utils.py

@@ -1,174 +0,0 @@
-import linecache
-import re
-import string
-from collections import Counter
-from logging import getLogger
-from pathlib import Path
-from typing import Dict, List
-
-import torch
-from torch.utils.data import Dataset
-
-from examples.seq2seq.utils import SortishSampler, trim_batch
-from transformers import BartTokenizer, T5Tokenizer
-
-
-def encode_line(tokenizer, line, max_length, padding_side, pad_to_max_length=True, return_tensors="pt"):
-    extra_kw = {"add_prefix_space": True} if isinstance(tokenizer, BartTokenizer) else {}
-    tokenizer.padding_side = padding_side
-    return tokenizer(
-        [line],
-        max_length=max_length,
-        padding="max_length" if pad_to_max_length else None,
-        truncation=True,
-        return_tensors=return_tensors,
-        add_special_tokens=True,
-        **extra_kw,
-    )
-
-
-class Seq2SeqDataset(Dataset):
-    def __init__(
-        self,
-        tokenizer,
-        data_dir,
-        max_source_length,
-        max_target_length,
-        type_path="train",
-        n_obs=None,
-        src_lang=None,
-        tgt_lang=None,
-        prefix="",
-    ):
-        super().__init__()
-        self.src_file = Path(data_dir).joinpath(type_path + ".source")
-        self.tgt_file = Path(data_dir).joinpath(type_path + ".target")
-        self.src_lens = self.get_char_lens(self.src_file)
-        self.max_source_length = max_source_length
-        self.max_target_length = max_target_length
-        assert min(self.src_lens) > 0, f"found empty line in {self.src_file}"
-        self.tokenizer = tokenizer
-        self.prefix = prefix
-        if n_obs is not None:
-            self.src_lens = self.src_lens[:n_obs]
-        self.pad_token_id = self.tokenizer.pad_token_id
-        self.src_lang = src_lang
-        self.tgt_lang = tgt_lang
-
-    def __len__(self):
-        return len(self.src_lens)
-
-    def __getitem__(self, index) -> Dict[str, torch.Tensor]:
-        index = index + 1  # linecache starts at 1
-        source_line = self.prefix + linecache.getline(str(self.src_file), index).rstrip("\n")
-        tgt_line = linecache.getline(str(self.tgt_file), index).rstrip("\n")
-        assert source_line, f"empty source line for index {index}"
-        assert tgt_line, f"empty tgt line for index {index}"
-
-        # Need to add eos token manually for T5
-        if isinstance(self.tokenizer, T5Tokenizer):
-            source_line += self.tokenizer.eos_token
-            tgt_line += self.tokenizer.eos_token
-
-        # Pad source to the left and target to the right
-        source_inputs = encode_line(self.tokenizer, source_line, self.max_source_length, "right")  # "left")
-        target_inputs = encode_line(self.tokenizer, tgt_line, self.max_target_length, "right")
-
-        source_ids = source_inputs["input_ids"].squeeze()
-        target_ids = target_inputs["input_ids"].squeeze()
-        src_mask = source_inputs["attention_mask"].squeeze()
-        return {
-            "input_ids": source_ids,
-            "attention_mask": src_mask,
-            "decoder_input_ids": target_ids,
-        }
-
-    @staticmethod
-    def get_char_lens(data_file):
-        return [len(x) for x in Path(data_file).open().readlines()]
-
-    def collate_fn(self, batch) -> Dict[str, torch.Tensor]:
-        input_ids = torch.stack([x["input_ids"] for x in batch])
-        masks = torch.stack([x["attention_mask"] for x in batch])
-        target_ids = torch.stack([x["decoder_input_ids"] for x in batch])
-        pad_token_id = self.pad_token_id
-        y = trim_batch(target_ids, pad_token_id)
-        source_ids, source_mask = trim_batch(input_ids, pad_token_id, attention_mask=masks)
-        batch = {
-            "input_ids": source_ids,
-            "attention_mask": source_mask,
-            "decoder_input_ids": y,
-        }
-        return batch
-
-    def make_sortish_sampler(self, batch_size):
-        return SortishSampler(self.src_lens, batch_size)
-
-
-logger = getLogger(__name__)
-
-
-def normalize_answer(s):
-    """Lower text and remove punctuation, articles and extra whitespace."""
-
-    def remove_articles(text):
-        return re.sub(r"\b(a|an|the)\b", " ", text)
-
-    def white_space_fix(text):
-        return " ".join(text.split())
-
-    def remove_punc(text):
-        exclude = set(string.punctuation)
-        return "".join(ch for ch in text if ch not in exclude)
-
-    def lower(text):
-        return text.lower()
-
-    return white_space_fix(remove_articles(remove_punc(lower(s))))
-
-
-def f1_score(prediction, ground_truth):
-    prediction_tokens = normalize_answer(prediction).split()
-    ground_truth_tokens = normalize_answer(ground_truth).split()
-    common = Counter(prediction_tokens) & Counter(ground_truth_tokens)
-    num_same = sum(common.values())
-    if num_same == 0:
-        return 0
-    precision = 1.0 * num_same / len(prediction_tokens)
-    recall = 1.0 * num_same / len(ground_truth_tokens)
-    f1 = (2 * precision * recall) / (precision + recall)
-    return f1
-
-
-def exact_match_score(prediction, ground_truth):
-    return normalize_answer(prediction) == normalize_answer(ground_truth)
-
-
-def calculate_exact_match(output_lns: List[str], reference_lns: List[str]) -> Dict:
-    assert len(output_lns) == len(reference_lns)
-    em = 0
-    for hypo, pred in zip(output_lns, reference_lns):
-        em += exact_match_score(hypo, pred)
-    if len(output_lns) > 0:
-        em /= len(output_lns)
-    return {"em": em}
-
-
-def is_rag_model(model_prefix):
-    return model_prefix.startswith("rag")
-
-
-def set_extra_model_params(extra_params, hparams, config):
-    equivalent_param = {p: p for p in extra_params}
-    # T5 models don't have `dropout` param, they have `dropout_rate` instead
-    equivalent_param["dropout"] = "dropout_rate"
-    for p in extra_params:
-        if getattr(hparams, p, None):
-            if not hasattr(config, p) and not hasattr(config, equivalent_param[p]):
-                logger.info("config doesn't have a `{}` attribute".format(p))
-                delattr(hparams, p)
-                continue
-            set_p = p if hasattr(config, p) else equivalent_param[p]
-            setattr(config, set_p, getattr(hparams, p))
-            delattr(hparams, p)
-    return hparams, config

examples/seq2seq/aggregate_distributed_results.py

@@ -0,0 +1,46 @@
+from pathlib import Path
+
+import fire
+
+
+try:
+    from .utils import calculate_bleu, calculate_rouge, load_json, save_json, write_txt_file
+except ImportError:
+    from utils import calculate_bleu, calculate_rouge, load_json, save_json, write_txt_file
+
+
+def combine_partial_results(
+    result_dir: str, save_dir: str = None, save_prefix=None, calc_bleu=False, just_metrics=False
+):
+    """Write first n lines of each file f in src_dir to dest_dir/f """
+    src_dir = Path(result_dir)
+    save_dir = Path(save_dir)
+    save_dir.mkdir(exist_ok=True)
+    paths_to_combine = list(src_dir.glob("rank*.json"))
+    records = []
+    for partial_result in paths_to_combine:
+        records.extend(load_json(partial_result))
+    preds = [x["pred"] for x in records]
+    labels = [x["label"] for x in records]
+    score_fn = calculate_bleu if calc_bleu else calculate_rouge
+    metrics = score_fn(preds, labels)
+    save_json(metrics, save_dir.joinpath("metrics.json"))  # better would be be {prefix}_{rouge|bleu}.json
+    print(metrics)
+    if just_metrics:
+        return
+
+    if save_prefix is None:
+        save_prefix = "generated"
+        print("using generated as prefix")
+
+    tgt_path = save_dir.joinpath(f"{save_prefix}.target")
+    write_txt_file(labels, tgt_path)
+    pred_path = save_dir.joinpath(f"{save_prefix}.pred_target")
+    write_txt_file(preds, pred_path)
+    if "source" in records[0]:
+        src_path = save_dir.joinpath(f"{save_prefix}.source")
+        write_txt_file([x["source"] for x in records], src_path)
+
+
+if __name__ == "__main__":
+    fire.Fire(combine_partial_results)

examples/seq2seq/distillation.py

@@ -472,6 +472,7 @@ LAYERS_TO_COPY = {
         6: [0, 3, 6, 9, 12, 15],
         8: [0, 2, 4, 6, 8, 10, 12, 15],
         9: [0, 1, 3, 5, 7, 9, 11, 13, 15],
+        12: [0, 1, 2, 3, 4, 5, 6, 7, 9, 11, 13, 15],
         16: list(range(16)),
     },
     6: {1: [0], 2: [0, 5], 3: [0, 2, 5], 4: [0, 1, 3, 5], 6: list(range(6))},

examples/seq2seq/run_distributed_eval.py

@@ -0,0 +1,142 @@
+import argparse
+from logging import getLogger
+from pathlib import Path
+from typing import Dict
+
+import torch
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+
+from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
+
+
+logger = getLogger(__name__)
+
+try:
+    from .utils import Seq2SeqDataset, parse_numeric_cl_kwargs, save_json, use_task_specific_params
+except ImportError:
+    from utils import Seq2SeqDataset, parse_numeric_cl_kwargs, save_json, use_task_specific_params
+
+
+DEFAULT_DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+
+
+def eval_data_dir(
+    data_dir,
+    save_dir: str,
+    model_name: str,
+    bs: int = 8,
+    max_source_length: int = 1024,
+    type_path="val",
+    n_obs=None,
+    fp16=False,
+    save_source=False,
+    num_beams: int = 4,
+    task="summarization",
+    local_rank=None,
+    **generate_kwargs,
+) -> Dict:
+    """Run evaluation on part of the data for one gpu and save to {save_dir}/rank_{rank}_output.json"""
+    model_name = str(model_name)
+    assert local_rank is not None
+    torch.distributed.init_process_group(backend="nccl", rank=local_rank)
+
+    save_dir = Path(save_dir)
+    save_path = save_dir.joinpath(f"rank_{local_rank}_output.json")
+    torch.cuda.set_device(local_rank)
+    model = AutoModelForSeq2SeqLM.from_pretrained(model_name).cuda()
+    if fp16:
+        model = model.half()
+
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    logger.info(f"Inferred tokenizer type: {tokenizer.__class__}")  # if this is wrong, check config.model_type.
+    use_task_specific_params(model, task)  # update config with task specific params
+    if max_source_length is None:
+        max_source_length = tokenizer.model_max_length
+    ds = Seq2SeqDataset(
+        tokenizer,
+        data_dir,
+        max_source_length,
+        max_target_length=1024,
+        type_path=type_path,
+        n_obs=n_obs,
+        prefix=model.config.prefix,
+    )
+    sampler = ds.make_sortish_sampler(bs, distributed=True)
+    data_loader = DataLoader(ds, sampler=sampler, batch_size=bs, collate_fn=ds.collate_fn)
+    dec_kwargs = dict(skip_special_tokens=True, clean_up_tokenization_spaces=False)  # tokenizer.decode
+    results = []
+    for batch in tqdm(data_loader):
+        summaries = model.generate(
+            input_ids=batch["input_ids"].to(model.device),
+            attention_mask=batch["attention_mask"].to(model.device),
+            num_beams=num_beams,
+            **generate_kwargs,
+        )
+        preds = tokenizer.batch_decode(summaries, **dec_kwargs)
+        labels = tokenizer.batch_decode(batch["labels"], **dec_kwargs)
+        if save_source:
+            docs = tokenizer.batch_decode(batch["input_ids"], **dec_kwargs)
+        for i in range(len(labels)):
+            label, pred = labels[i], preds[i]
+            if save_source:
+                results.append(dict(pred=pred, label=label, source=docs[i]))
+            else:
+                results.append(dict(pred=pred, label=label))
+    save_json(results, save_path)
+    return results
+
+
+def run_generate():
+    parser = argparse.ArgumentParser(
+        epilog="Unspecified args like --num_beams=2 --decoder_start_token_id=4 are passed to model.generate"
+    )
+    parser.add_argument("--input_path", type=str, help="like cnn_dm/test.source")
+    parser.add_argument(
+        "--model_name",
+        type=str,
+        help="like facebook/bart-large-cnn,t5-base, etc.",
+        default="sshleifer/distilbart-xsum-12-3",
+    )
+    parser.add_argument("--save_dir", type=str, help="where to save", default="tmp_gen")
+    parser.add_argument("--max_source_length", type=int, default=None)
+    parser.add_argument(
+        "--type_path", type=str, default="test", help="which subset to evaluate typically train/val/test"
+    )
+    parser.add_argument("--reference_path", type=str, required=False, help="like cnn_dm/test.target")
+    parser.add_argument("--task", type=str, default="summarization", help="used for task_specific_params + metrics")
+    parser.add_argument("--bs", type=int, default=8, required=False, help="batch size")
+    parser.add_argument(
+        "--local_rank", type=int, default=-1, required=False, help="should be passed by distributed.launch"
+    )
+
+    parser.add_argument(
+        "--n_obs", type=int, default=None, required=False, help="How many observations. Defaults to all."
+    )
+    parser.add_argument("--fp16", action="store_true")
+    parser.add_argument("--save_source", action="store_true")
+
+    args, rest = parser.parse_known_args()
+    generate_kwargs = parse_numeric_cl_kwargs(rest)
+    if generate_kwargs:
+        print(f"parsed the following generate kwargs: {generate_kwargs}")
+    Path(args.save_dir).mkdir(exist_ok=True)
+    eval_data_dir(
+        args.input_path,
+        args.save_dir,
+        args.model_name,
+        type_path=args.type_path,
+        batch_size=args.bs,
+        fp16=args.fp16,
+        task=args.task,
+        local_rank=args.local_rank,
+        n_obs=args.n_obs,
+        save_source=args.save_source,
+        max_source_length=args.max_source_length,
+        **generate_kwargs,
+    )
+
+
+if __name__ == "__main__":
+    # Usage for MT:
+    run_generate()

examples/seq2seq/utils.py

@@ -98,7 +98,8 @@ class AbstractSeq2SeqDataset(Dataset):
         self.max_target_length = max_target_length
         assert min(self.src_lens) > 0, f"found empty line in {self.src_file}"
         self.tokenizer = tokenizer
-        self.prefix = prefix
+        self.prefix = prefix if prefix is not None else ""
+
         if n_obs is not None:
             self.src_lens = self.src_lens[:n_obs]
         self.pad_token_id = self.tokenizer.pad_token_id
@@ -387,3 +388,10 @@ def parse_numeric_cl_kwargs(unparsed_args: List[str]) -> Dict[str, Union[int, fl
 
         result[unparsed_args[i][2:]] = value
     return result
+
+
+def write_txt_file(ordered_tgt, path):
+    f = Path(path).open("w")
+    for ln in ordered_tgt:
+        f.write(ln + "\n")
+        f.flush()

setup.py

@@ -89,7 +89,7 @@ extras["onnxruntime"] = ["onnxruntime>=1.4.0", "onnxruntime-tools>=1.4.2"]
 extras["serving"] = ["pydantic", "uvicorn", "fastapi", "starlette"]
 extras["all"] = extras["serving"] + ["tensorflow", "torch"]
 
-extras["testing"] = ["pytest", "pytest-xdist", "timeout-decorator", "psutil", "parameterized", "faiss", "datasets"]
+extras["testing"] = ["pytest", "pytest-xdist", "timeout-decorator", "psutil", "parameterized"]
 # sphinx-rtd-theme==0.5.0 introduced big changes in the style.
 extras["docs"] = ["recommonmark", "sphinx", "sphinx-markdown-tables", "sphinx-rtd-theme==0.4.3", "sphinx-copybutton"]
 extras["quality"] = ["black >= 20.8b1", "isort >= 5", "flake8 >= 3.8.3"]

src/transformers/__init__.py

@@ -84,7 +84,6 @@ from .file_utils import (
     cached_path,
     is_apex_available,
     is_datasets_available,
-    is_faiss_available,
     is_psutil_available,
     is_py3nvml_available,
     is_tf_available,
@@ -713,13 +712,6 @@ if is_tf_available():
     from .trainer_tf import TFTrainer
 
 
-if is_torch_available() and is_datasets_available() and is_faiss_available() and is_psutil_available():
-    from .configuration_rag import RagConfig
-    from .modeling_rag import RagModel, RagSequenceForGeneration, RagTokenForGeneration
-    from .retrieval_rag import RagRetriever
-    from .tokenization_rag import RagTokenizer
-
-
 if not is_tf_available() and not is_torch_available():
     logger.warning(
         "Neither PyTorch nor TensorFlow >= 2.0 have been found."

src/transformers/configuration_auto.py

@@ -24,7 +24,6 @@ from .configuration_bert_generation import BertGenerationConfig
 from .configuration_camembert import CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, CamembertConfig
 from .configuration_ctrl import CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP, CTRLConfig
 from .configuration_distilbert import DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, DistilBertConfig
-from .configuration_dpr import DPR_PRETRAINED_CONFIG_ARCHIVE_MAP, DPRConfig
 from .configuration_electra import ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP, ElectraConfig
 from .configuration_encoder_decoder import EncoderDecoderConfig
 from .configuration_flaubert import FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, FlaubertConfig
@@ -72,7 +71,6 @@ ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict(
         RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
         FUNNEL_PRETRAINED_CONFIG_ARCHIVE_MAP,
         LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
-        DPR_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ]
     for key, value, in pretrained_map.items()
 )
@@ -107,7 +105,6 @@ CONFIG_MAPPING = OrderedDict(
         ("encoder-decoder", EncoderDecoderConfig),
         ("funnel", FunnelConfig),
         ("lxmert", LxmertConfig),
-        ("dpr", DPRConfig),
     ]
 )
 
@@ -140,7 +137,6 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("encoder-decoder", "Encoder decoder"),
         ("funnel", "Funnel Transformer"),
         ("lxmert", "LXMERT"),
-        ("dpr", "DPR"),
     ]
 )
 

src/transformers/configuration_dpr.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 """ DPR model configuration """
 
-from .configuration_utils import PretrainedConfig
+from .configuration_bert import BertConfig
 from .utils import logging
 
 
@@ -27,7 +27,7 @@ DPR_PRETRAINED_CONFIG_ARCHIVE_MAP = {
 }
 
 
-class DPRConfig(PretrainedConfig):
+class DPRConfig(BertConfig):
     r"""
     :class:`~transformers.DPRConfig` is the configuration class to store the configuration of a
     `DPRModel`.
@@ -36,73 +36,12 @@ class DPRConfig(PretrainedConfig):
     It is used to instantiate the components of the DPR model.
 
     Args:
-        vocab_size (:obj:`int`, optional, defaults to 30522):
-            Vocabulary size of the DPR model. Defines the different tokens that
-            can be represented by the `inputs_ids` passed to the forward method of :class:`~transformers.BertModel`.
-        hidden_size (:obj:`int`, optional, defaults to 768):
-            Dimensionality of the encoder layers and the pooler layer.
-        num_hidden_layers (:obj:`int`, optional, defaults to 12):
-            Number of hidden layers in the Transformer encoder.
-        num_attention_heads (:obj:`int`, optional, defaults to 12):
-            Number of attention heads for each attention layer in the Transformer encoder.
-        intermediate_size (:obj:`int`, optional, defaults to 3072):
-            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
-        hidden_act (:obj:`str` or :obj:`function`, optional, defaults to "gelu"):
-            The non-linear activation function (function or string) in the encoder and pooler.
-            If string, "gelu", "relu", "swish" and "gelu_new" are supported.
-        hidden_dropout_prob (:obj:`float`, optional, defaults to 0.1):
-            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
-        attention_probs_dropout_prob (:obj:`float`, optional, defaults to 0.1):
-            The dropout ratio for the attention probabilities.
-        max_position_embeddings (:obj:`int`, optional, defaults to 512):
-            The maximum sequence length that this model might ever be used with.
-            Typically set this to something large just in case (e.g., 512 or 1024 or 2048).
-        type_vocab_size (:obj:`int`, optional, defaults to 2):
-            The vocabulary size of the `token_type_ids` passed into :class:`~transformers.BertModel`.
-        initializer_range (:obj:`float`, optional, defaults to 0.02):
-            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
-        layer_norm_eps (:obj:`float`, optional, defaults to 1e-12):
-            The epsilon used by the layer normalization layers.
-        gradient_checkpointing (:obj:`bool`, optional, defaults to :obj:`False`):
-            If True, use gradient checkpointing to save memory at the expense of slower backward pass.
         projection_dim (:obj:`int`, optional, defaults to 0):
             Dimension of the projection for the context and question encoders.
             If it is set to zero (default), then no projection is done.
     """
     model_type = "dpr"
 
-    def __init__(
-        self,
-        vocab_size=30522,
-        hidden_size=768,
-        num_hidden_layers=12,
-        num_attention_heads=12,
-        intermediate_size=3072,
-        hidden_act="gelu",
-        hidden_dropout_prob=0.1,
-        attention_probs_dropout_prob=0.1,
-        max_position_embeddings=512,
-        type_vocab_size=2,
-        initializer_range=0.02,
-        layer_norm_eps=1e-12,
-        pad_token_id=0,
-        gradient_checkpointing=False,
-        projection_dim: int = 0,
-        **kwargs
-    ):
-        super().__init__(pad_token_id=pad_token_id, **kwargs)
-
-        self.vocab_size = vocab_size
-        self.hidden_size = hidden_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_attention_heads = num_attention_heads
-        self.hidden_act = hidden_act
-        self.intermediate_size = intermediate_size
-        self.hidden_dropout_prob = hidden_dropout_prob
-        self.attention_probs_dropout_prob = attention_probs_dropout_prob
-        self.max_position_embeddings = max_position_embeddings
-        self.type_vocab_size = type_vocab_size
-        self.initializer_range = initializer_range
-        self.layer_norm_eps = layer_norm_eps
-        self.gradient_checkpointing = gradient_checkpointing
+    def __init__(self, projection_dim: int = 0, **kwargs):  # projection of the encoders, 0 for no projection
+        super().__init__(**kwargs)
         self.projection_dim = projection_dim

src/transformers/configuration_pegasus.py

@@ -49,7 +49,7 @@ DEFAULTS = dict(
 # Config values that vary between checkpoints: for testing and conversion
 task_specific_params = {
     # These are task specific params for pegasus-large and normal params for finetuned checkpoints
-    "summarization_xsum": {"length_penalty": 0.8, "max_length": 64, "max_position_embeddings": 512},
+    "summarization_xsum": {"length_penalty": 0.6, "max_length": 64, "max_position_embeddings": 512},
     "summarization_cnn_dailymail": {"length_penalty": 0.8, "max_length": 128, "max_position_embeddings": 1024},
     "summarization_newsroom": {"length_penalty": 0.8, "max_length": 128, "max_position_embeddings": 512},
     "summarization_wikihow": {"length_penalty": 0.6, "max_length": 256, "max_position_embeddings": 512},

src/transformers/configuration_rag.py

@@ -1,202 +0,0 @@
-# coding=utf-8
-# Copyright 2020, The RAG Authors and The HuggingFace Inc. team.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-""" RAG model configuration """
-
-import copy
-
-from .configuration_utils import PretrainedConfig
-from .file_utils import add_start_docstrings_to_callable
-
-
-RAG_PRETRAINED_CONFIG_ARCHIVE_MAP = {
-    "facebook/rag-sequence-nq": "TBA",
-    "facebook/rag-token-nq": "TBA",
-}
-
-RAG_CONFIG_DOC = r"""
-    :class:`~transformers.RagConfig` is the configuration class to store the configuration of a `RagModel`.
-
-    Args:
-        vocab_size (:obj:`int`, optional, defaults to ``None``):
-            Vocabulary size of the underlying generator model.
-        is_encoder_decoder (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Whether the model is used as an encoder/decoder or not.
-        title_sep (:obj:`str`, optional, defaults to  ``" / "``):
-            Separator inserted between the title and the text of the retrieved document when running
-            `:func:`~transformers.RagModel.contextualize``.
-        doc_sep (:obj:`str`, optional, defaults to  ``" // "``):
-            Separator inserted between the the text of the retrieved document and the original input when running
-            `:func:`~transformers.RagModel.contextualize``.
-        n_docs (:obj:`int`, optional, defaults to ``5``):
-            Number of retrieved docs.
-        max_combined_length (:int:`bool`, optional, defaults to ``300``):
-            Max length of contextualized input returned by `:func:`~transformers.RagModel.contextualize``.
-        retrieval_vector_size (:obj:`int`, optional, defaults to ``768``):
-            Dimensionality of the document embeddings indexed by the ``retriever``.
-        retrieval_batch_size (:obj:`int`, optional, defaults to ``8``):
-            Retrieval batch size - the number of queries issues concurrently to the faiss index excapsulated
-            by the ``retriever``.
-        retriever_type (:obj:`str`, optional, defaults to ``hf_retriever``):
-            A type of index encapsulated by the ``retriever``. Possible options include:
-
-                - ``hf_retriever`` - and index build for an instance of :class:`~nlp.Datasets`
-                - ``legacy_retriever`` - an index build with the native DPR implementation (see https://github.com/facebookresearch/DPR for details).
-        dataset (:obj:`str`, optional, defaults to ``wiki_dpr``):
-            A datatset identifier of the indexed dataset on HuggingFace AWS bucket (list all available datasets and ids with ``nlp.list_datasets()``).
-        dataset_split (:obj:`str`, optional, defaults to ``train``)
-            Which split of the ``dataset`` to load.
-        index_name (:obj:`str`, optional, defaults to ``train``)
-            The index_name of the index associated with the ``dataset``.
-        index_path (:obj:`str`, optional, defaults to ``None``)
-            The path to the serialized faiss index on disk.
-        passages_path: (:obj:`str`, optional, defaults to ``None``):
-            A path to text passages compatible with the faiss index. Required if using :class:`~transformers.retrieval_rag.LegacyIndex`
-        dummy (:obj:`bool`, optional, defaults to ``False``)
-            Whether to load a ``dummy`` variant of the dataset specified by ``dataset`` argument.
-        pretrained_question_encoder_tokenizer_name_or_path: (:obj:`str`, optional, defaults to ``facebook/dpr-question_encoder-single-nq-base``):
-            A string specifying the ``question_encoder`` tokenizer to be loaded.
-        pretrained_question_encoder_name_or_path: (:obj:`str`, optional, defaults to ``facebook/dpr-question_encoder-single-nq-base``):
-            A string specifying the ``question_encoder`` model to be loaded.
-        pretrained_generator_tokenizer_name_or_path: (:obj:`str`, optional, defaults to ``facebook/bart-large``):
-            A string specifying the ``generator`` tokenizer to be loaded.
-        pretrained_generator_name_or_path: (:obj:`str`, optional, defaults to ``facebook/bart-large``):
-            A string specifying the ``generator`` model to be loaded.
-
-    Args linked to the tokenizer - they have to be compatible with equivalent parameters of the ``generator``:
-        prefix (:obj:`str`, `optional`):
-            A specific prompt that should be added at the beginning of each text before calling the model.
-        bos_token_id (:obj:`int`, `optional`):
-            The id of the `beginning-of-stream` token.
-        pad_token_id (:obj:`int`, `optional`):
-            The id of the `padding` token.
-        eos_token_id (:obj:`int`, `optional`)"
-            The id of the `end-of-stream` token.
-        decoder_start_token_id** (:obj:`int`, `optional`):
-            If an encoder-decoder model starts decoding with a different token than `bos`, the id of that token.
-        marginalize (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            If :obj:`True`, `logits`, returned as part of :class:`~transformers.file_utils.Seq2SeqLMOutputWithDocs` are marginalized, yielding
-            the shape of :obj:`(batch_size, sequence_length, hidden_size)`. Otherwise we return raw, non-marginalized logits of shape
-            :obj:`(batch_size * n_docs, sequence_length, hidden_size)`. ``marginalize`` is set to :obj:`True` during generation. The parameter is
-            ignored if ``return_loss`` is set to :obj:`True`.
-
-"""
-
-
-@add_start_docstrings_to_callable(RAG_CONFIG_DOC)
-class RagConfig(PretrainedConfig):
-    model_type = "rag"
-
-    def __init__(
-        self,
-        vocab_size=None,
-        is_encoder_decoder=True,
-        prefix=None,
-        bos_token_id=None,
-        pad_token_id=None,
-        eos_token_id=None,
-        decoder_start_token_id=None,
-        title_sep=" / ",
-        doc_sep=" // ",
-        n_docs=5,
-        max_combined_length=300,
-        retrieval_vector_size=768,
-        retrieval_batch_size=8,
-        dataset="wiki_dpr",
-        dataset_split="train",
-        index_name="compressed",
-        index_path=None,
-        passages_path=None,
-        use_dummy_dataset=False,
-        reduce_loss=False,
-        label_smoothing=0.0,
-        deduplicate=True,  # defaults to True
-        num_doc_return_sequences=1,  # defaults to 1
-        num_doc_beams=1,  # defaults to 1
-        exclude_bos_score=False,
-        do_marginalize=False,
-        output_retrieved=False,
-        **kwargs
-    ):
-        super().__init__(**kwargs)
-        assert (
-            "question_encoder" in kwargs and "generator" in kwargs
-        ), "Config has to be initialized with question_encoder and generator config"
-        question_encoder_config = kwargs.pop("question_encoder")
-        question_encoder_model_type = question_encoder_config.pop("model_type")
-        decoder_config = kwargs.pop("generator")
-        decoder_model_type = decoder_config.pop("model_type")
-
-        from .configuration_auto import AutoConfig
-
-        self.question_encoder = AutoConfig.for_model(question_encoder_model_type, **question_encoder_config)
-        self.generator = AutoConfig.for_model(decoder_model_type, **decoder_config)
-
-        self.reduce_loss = reduce_loss
-        self.label_smoothing = label_smoothing
-        self.exclude_bos_score = exclude_bos_score
-        self.do_marginalize = do_marginalize
-        self.vocab_size = vocab_size
-        self.is_encoder_decoder = is_encoder_decoder
-        self.prefix = prefix
-        self.bos_token_id = bos_token_id
-        self.pad_token_id = pad_token_id
-        self.eos_token_id = eos_token_id
-        self.decoder_start_token_id = decoder_start_token_id
-
-        self.title_sep = title_sep
-        self.doc_sep = doc_sep
-        self.n_docs = n_docs
-        self.max_combined_length = max_combined_length
-
-        self.dataset = dataset
-        self.dataset_split = dataset_split
-        self.index_name = index_name
-
-        self.retrieval_vector_size = retrieval_vector_size
-        self.retrieval_batch_size = retrieval_batch_size
-        self.passages_path = passages_path
-        self.index_path = index_path
-        self.use_dummy_dataset = use_dummy_dataset
-
-        self.output_retrieved = output_retrieved
-
-        self.deduplicate = deduplicate
-        self.num_doc_return_sequences = num_doc_return_sequences
-        self.num_doc_beams = num_doc_beams
-
-    @classmethod
-    def from_question_encoder_generator_configs(
-        cls, question_encoder_config: PretrainedConfig, generator_config: PretrainedConfig, **kwargs
-    ) -> PretrainedConfig:
-        r"""
-        Instantiate a :class:`~transformers.EncoderDecoderConfig` (or a derived class) from a pre-trained encoder model configuration and decoder model configuration.
-
-        Returns:
-            :class:`EncoderDecoderConfig`: An instance of a configuration object
-        """
-        return cls(question_encoder=question_encoder_config.to_dict(), generator=generator_config.to_dict(), **kwargs)
-
-    def to_dict(self):
-        """
-        Serializes this instance to a Python dictionary. Override the default `to_dict()` from `PretrainedConfig`.
-
-        Returns:
-            :obj:`Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
-        """
-        output = copy.deepcopy(self.__dict__)
-        output["question_encoder"] = self.question_encoder.to_dict()
-        output["generator"] = self.generator.to_dict()
-        output["model_type"] = self.__class__.model_type
-        return output

src/transformers/convert_pegasus_tf_to_pytorch.py

@@ -47,8 +47,8 @@ PATTERNS = [
 
 def rename_state_dict_key(k):
 
-    for pegasus_name, bart_name in PATTERNS:
-        k = k.replace(pegasus_name, bart_name)
+    for pegasus_name, hf_name in PATTERNS:
+        k = k.replace(pegasus_name, hf_name)
     return k
 
 
@@ -57,13 +57,12 @@ def rename_state_dict_key(k):
 # TODO(SS): one constant
 
 
-def convert_pegasus_to_bart(tf_weights: dict, cfg_updates: dict) -> PegasusForConditionalGeneration:
+def convert_pegasus(tf_weights: dict, cfg_updates: dict) -> PegasusForConditionalGeneration:
     cfg_kwargs = DEFAULTS.copy()
     cfg_kwargs.update(cfg_updates)
-
-    cfg = PegasusConfig(**cfg_updates)
-    bart = PegasusForConditionalGeneration(cfg)
-    sd = bart.model.state_dict()
+    cfg = PegasusConfig(**cfg_kwargs)
+    torch_model = PegasusForConditionalGeneration(cfg)
+    sd = torch_model.model.state_dict()
     mapping = {}
     for k, v in tf_weights.items():
         new_k = rename_state_dict_key(k)
@@ -80,13 +79,13 @@ def convert_pegasus_to_bart(tf_weights: dict, cfg_updates: dict) -> PegasusForCo
     mapping["decoder.embed_tokens.weight"] = mapping["shared.weight"]
     empty_biases = {k: torch.zeros_like(v) for k, v in sd.items() if k.endswith("bias") and k not in mapping}
     mapping.update(**empty_biases)
-    missing, extra = bart.model.load_state_dict(mapping, strict=False)
+    missing, extra = torch_model.model.load_state_dict(mapping, strict=False)
     unexpected_missing = [
         k for k in missing if k not in ["encoder.embed_positions.weight", "decoder.embed_positions.weight"]
     ]
     assert unexpected_missing == [], f"no matches found for the following torch keys {unexpected_missing}"
     assert extra == [], f"no matches found for the following tf keys {extra}"
-    return bart
+    return torch_model
 
 
 def get_tf_weights_as_numpy(path="./ckpt/aeslc/model.ckpt-32000") -> Dict:
@@ -115,7 +114,7 @@ def convert_pegasus_ckpt_to_pytorch(ckpt_path: str, save_dir: str):
     cfg_updates = task_specific_params[f"summarization_{dataset}"]
     if dataset == "large":
         cfg_updates["task_specific_params"] = task_specific_params
-    torch_model = convert_pegasus_to_bart(tf_weights, cfg_updates)
+    torch_model = convert_pegasus(tf_weights, cfg_updates)
     torch_model.save_pretrained(save_dir)
     sd = torch_model.state_dict()
     sd.pop("model.decoder.embed_positions.weight")

src/transformers/file_utils.py

@@ -69,7 +69,6 @@ try:
     import datasets  # noqa: F401
 
     _datasets_available = True
-    logger.debug(f"Succesfully imported datasets version {datasets.__version__}")
 
 except ImportError:
     _datasets_available = False
@@ -120,16 +119,6 @@ try:
 except ImportError:
     _has_apex = False
 
-
-try:
-    import faiss  # noqa: F401
-
-    _faiss_available = True
-    logger.debug(f"Succesfully imported faiss version {faiss.__version__}")
-except ImportError:
-    _faiss_available = False
-
-
 default_cache_path = os.path.join(torch_cache_home, "transformers")
 
 
@@ -182,10 +171,6 @@ def is_apex_available():
     return _has_apex
 
 
-def is_faiss_available():
-    return _faiss_available
-
-
 def add_start_docstrings(*docstr):
     def docstring_decorator(fn):
         fn.__doc__ = "".join(docstr) + (fn.__doc__ if fn.__doc__ is not None else "")

src/transformers/modeling_auto.py

@@ -27,7 +27,6 @@ from .configuration_auto import (
     CamembertConfig,
     CTRLConfig,
     DistilBertConfig,
-    DPRConfig,
     ElectraConfig,
     EncoderDecoderConfig,
     FlaubertConfig,
@@ -95,7 +94,6 @@ from .modeling_distilbert import (
     DistilBertForTokenClassification,
     DistilBertModel,
 )
-from .modeling_dpr import DPRQuestionEncoder
 from .modeling_electra import (
     ElectraForMaskedLM,
     ElectraForMultipleChoice,
@@ -219,7 +217,6 @@ MODEL_MAPPING = OrderedDict(
         (FunnelConfig, FunnelModel),
         (LxmertConfig, LxmertModel),
         (BertGenerationConfig, BertGenerationEncoder),
-        (DPRConfig, DPRQuestionEncoder),
     ]
 )
 

src/transformers/modeling_dpr.py

@@ -272,7 +272,6 @@ class DPRPretrainedContextEncoder(PreTrainedModel):
     config_class = DPRConfig
     load_tf_weights = None
     base_model_prefix = "ctx_encoder"
-    authorized_missing_keys = [r"position_ids"]
 
     def init_weights(self):
         self.ctx_encoder.init_weights()
@@ -286,7 +285,6 @@ class DPRPretrainedQuestionEncoder(PreTrainedModel):
     config_class = DPRConfig
     load_tf_weights = None
     base_model_prefix = "question_encoder"
-    authorized_missing_keys = [r"position_ids"]
 
     def init_weights(self):
         self.question_encoder.init_weights()
@@ -300,7 +298,6 @@ class DPRPretrainedReader(PreTrainedModel):
     config_class = DPRConfig
     load_tf_weights = None
     base_model_prefix = "span_predictor"
-    authorized_missing_keys = [r"position_ids"]
 
     def init_weights(self):
         self.span_predictor.encoder.init_weights()

src/transformers/modeling_encoder_decoder.py

@@ -250,7 +250,7 @@ class EncoderDecoderModel(PreTrainedModel):
                     encoder_config.is_decoder = False
                     encoder_config.add_cross_attention = False
 
-                kwargs_encoder["config"] = encoder_config
+                    kwargs_encoder["config"] = encoder_config
 
             encoder = AutoModel.from_pretrained(encoder_pretrained_model_name_or_path, *model_args, **kwargs_encoder)
 

src/transformers/retrieval_rag.py

@@ -1,585 +0,0 @@
-# coding=utf-8
-# Copyright 2020, The RAG Authors and The HuggingFace Inc. team.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""RAG Retriever model implementation."""
-
-import os
-import pickle
-import time
-from typing import Iterable, List, Optional, Tuple
-
-import faiss
-import numpy as np
-import psutil
-import torch
-import torch.distributed as dist
-from datasets import load_dataset
-
-from .configuration_rag import RagConfig
-from .file_utils import cached_path, is_remote_url
-from .tokenization_rag import RagTokenizer
-from .tokenization_t5 import T5Tokenizer
-from .tokenization_utils_base import BatchEncoding
-from .utils import logging
-
-
-logger = logging.get_logger(__name__)
-
-
-LEGACY_INDEX_PATH = "https://storage.googleapis.com/huggingface-nlp/datasets/wiki_dpr/"
-
-
-class Index(object):
-    """
-    A base class for the Indices encapsulated by the :class:`~transformers.RagRetriever`.
-    """
-
-    def __init__(self, *args, **kwargs):
-        pass
-
-    def get_doc_dicts(self, doc_ids: np.ndarray) -> List[dict]:
-        """
-        Returns a list of dictionaries, containing titles and text of the retrieved documents.
-
-        Args:
-            doc_ids (:obj:`np.ndarray` of shape :obj:`(batch_size, n_docs)`):
-                A tensor of document indices.
-        """
-        pass
-
-    def get_top_docs(self, question_hidden_states: np.ndarray, n_docs=5) -> Tuple[np.ndarray, np.ndarray]:
-        """
-        For each query in the batch, retrieves ``n_docs`` documents.
-
-        Args:
-            question_hidden_states (:obj:`np.ndarray` of shape :obj:`(batch_size, vector_size):
-                An array of query vectors.
-            n_docs (:obj:`int`):
-                The number of docs retrieved per query.
-
-        Returns:
-            :obj:`np.ndarray` of shape :obj:`(batch_size, n_docs)`: A tensor of indices of retrieved documents.
-            :obj:`np.ndarray` of shape :obj:`(batch_size, vector_size)`: A tensor of vector representations of retrieved documents.
-        """
-        raise NotImplementedError
-
-    def is_initialized(self):
-        """
-        Returns :obj:`True` if index is already initialized.
-        """
-        raise NotImplementedError
-
-    def init_index(self):
-        """
-        A function responsible for loading the index into memory. Should be called only once per training run of a RAG model.
-        E.g. if the model is trained on multiple GPUs in a distributed setup, only one of the workers will load the index.
-        """
-        raise NotImplementedError
-
-
-class LegacyIndex(Index):
-    """
-    An index which can be deserialized from the files built using https://github.com/facebookresearch/DPR.
-    We use default faiss index parameters as specified in that repository.
-
-    Args:
-        vector_size (:obj:`int`):
-            The dimension of indexed vectors.
-        index_path (:obj:`str`):
-            A path to a `directory` containing index files compatible with
-            :class:`~transformers.retrieval_rag.LegacyIndex`
-    """
-
-    INDEX_FILENAME = "hf_bert_base.hnswSQ8_correct_phi_128.c_index"
-    PASSAGE_FILENAME = "psgs_w100.tsv.pkl"
-
-    def __init__(self, vector_size, index_path):
-        self.index_id_to_db_id = []
-        self.index_path = index_path
-        self.passages = self._load_passages()
-        self.vector_size = vector_size
-        self.index = None
-        self._index_initialize = False
-
-    def _resolve_path(self, index_path, filename):
-        assert os.path.isdir(index_path) or is_remote_url(index_path), "Please specify a valid ``index_path``."
-        archive_file = os.path.join(index_path, filename)
-        try:
-            # Load from URL or cache if already cached
-            resolved_archive_file = cached_path(archive_file)
-            if resolved_archive_file is None:
-                raise EnvironmentError
-        except EnvironmentError:
-            msg = (
-                f"Can't load '{archive_file}'. Make sure that:\n\n"
-                f"- '{index_path}' is a correct remote path to a directory containing a file named {filename}"
-                f"- or '{index_path}' is the correct path to a directory containing a file named {filename}.\n\n"
-            )
-            raise EnvironmentError(msg)
-        if resolved_archive_file == archive_file:
-            logger.info("loading file {}".format(archive_file))
-        else:
-            logger.info("loading file {} from cache at {}".format(archive_file, resolved_archive_file))
-        return resolved_archive_file
-
-    def _load_passages(self):
-        logger.info("Loading passages from {}".format(self.index_path))
-        passages_path = self._resolve_path(self.index_path, self.PASSAGE_FILENAME)
-        with open(passages_path, "rb") as passages_file:
-            passages = pickle.load(passages_file)
-        return passages
-
-    def _deserialize_index(self):
-        logger.info("Loading index from {}".format(self.index_path))
-        resolved_index_path = self._resolve_path(self.index_path, self.INDEX_FILENAME + ".index.dpr")
-        self.index = faiss.read_index(resolved_index_path)
-        resolved_meta_path = self._resolve_path(self.index_path, self.INDEX_FILENAME + ".index_meta.dpr")
-        with open(resolved_meta_path, "rb") as metadata_file:
-            self.index_id_to_db_id = pickle.load(metadata_file)
-        assert (
-            len(self.index_id_to_db_id) == self.index.ntotal
-        ), "Deserialized index_id_to_db_id should match faiss index size"
-
-    def is_initialized(self):
-        return self._index_initialize
-
-    def init_index(self):
-        index = faiss.IndexHNSWFlat(self.vector_size + 1, 512)
-        index.hnsw.efSearch = 128
-        index.hnsw.efConstruction = 200
-        self.index = index
-        self._deserialize_index()
-        self._index_initialize = True
-
-    def get_doc_dicts(self, doc_ids):
-        doc_list = []
-        for doc_ids_i in doc_ids:
-            ids = [str(int(doc_id)) for doc_id in doc_ids_i]
-            docs = [self.passages[doc_id] for doc_id in ids]
-            doc_list.append(docs)
-        doc_dicts = []
-        for docs in doc_list:
-            doc_dict = {}
-            doc_dict["title"] = [doc[1] for doc in docs]
-            doc_dict["text"] = [doc[0] for doc in docs]
-            doc_dicts.append(doc_dict)
-        return doc_dicts
-
-    def get_top_docs(self, question_hidden_states: np.ndarray, n_docs=5) -> Tuple[np.ndarray, np.ndarray]:
-        aux_dim = np.zeros(len(question_hidden_states), dtype="float32").reshape(-1, 1)
-        query_nhsw_vectors = np.hstack((question_hidden_states, aux_dim))
-        _, docs_ids = self.index.search(query_nhsw_vectors, n_docs)
-        vectors = [[self.index.reconstruct(int(doc_id))[:-1] for doc_id in doc_ids] for doc_ids in docs_ids]
-        ids = [[int(self.index_id_to_db_id[doc_id]) for doc_id in doc_ids] for doc_ids in docs_ids]
-        return np.array(ids), np.array(vectors)
-
-
-class HFIndex(Index):
-    """
-    A wrapper around an instance of :class:`~datasets.Datasets`. If ``index_path`` is set to ``None``,
-    we load the pre-computed index available with the :class:`~datasets.arrow_dataset.Dataset`, otherwise, we load the index from the indicated path on disk.
-
-    Args:
-        dataset (:obj:`str`, optional, defaults to ``wiki_dpr``):
-            A datatset identifier of the indexed dataset on HuggingFace AWS bucket (list all available datasets and ids with ``datasets.list_datasets()``).
-        dataset_split (:obj:`str`, optional, defaults to ``train``)
-            Which split of the ``dataset`` to load.
-        index_name (:obj:`str`, optional, defaults to ``train``)
-            The index_name of the index associated with the ``dataset``. The index loaded from ``index_path`` will be saved under this name.
-        index_path (:obj:`str`, optional, defaults to ``None``)
-            The path to the serialized faiss index on disk.
-    """
-
-    def __init__(
-        self,
-        dataset_name: str,
-        dataset_split: str,
-        index_name: str,
-        index_path: Optional[str] = None,
-        use_dummy_dataset=False,
-    ):
-        super().__init__()
-        self.dataset_name = dataset_name
-        self.dataset_split = dataset_split
-        self.index_name = index_name
-        self.index_path = index_path
-        self.use_dummy_dataset = use_dummy_dataset
-        self._index_initialize = False
-
-        logger.info("Loading passages from {}".format(self.dataset_name))
-        self.dataset = load_dataset(
-            self.dataset_name, with_index=False, split=self.dataset_split, dummy=self.use_dummy_dataset
-        )
-
-    def is_initialized(self):
-        return self._index_initialize
-
-    def init_index(self):
-        if self.index_path is not None:
-            logger.info("Loading index from {}".format(self.index_path))
-            self.index.load_faiss_index(index_name=self.index_name, file=self.index_path)
-        else:
-            logger.info("Loading index from {}".format(self.dataset_name + " with index name " + self.index_name))
-            self.dataset = load_dataset(
-                self.dataset_name,
-                with_embeddings=True,
-                with_index=True,
-                split=self.dataset_split,
-                index_name=self.index_name,
-                dummy=self.use_dummy_dataset,
-            )
-        self._index_initialize = True
-
-    def get_doc_dicts(self, doc_ids: np.ndarray) -> List[dict]:
-        return [self.dataset[doc_ids[i].tolist()] for i in range(doc_ids.shape[0])]
-
-    def get_top_docs(self, question_hidden_states: np.ndarray, n_docs=5) -> Tuple[np.ndarray, np.ndarray]:
-        _, docs = self.dataset.get_nearest_examples_batch("embeddings", question_hidden_states, n_docs)
-        ids = [[int(i) for i in doc["id"]] for doc in docs]
-        vectors = [doc["embeddings"] for doc in docs]
-        return np.array(ids), np.array(vectors)  # shapes (batch_size, n_docs) and (batch_size, n_docs, d)
-
-
-class RagRetriever(object):
-    """
-    A distributed retriever built on top of the ``torch.distributed`` communication package. During training all workers
-    initalize their own instance of the retriever, however, only the main worker loads the index into memory. The index is stored
-    in cpu memory. The index will also work well in a non-distributed setup.
-
-    Args:
-        config (:class:`~transformers.RagConfig`):
-            The configuration of the RAG model this Retriever is used with. Contains parameters indicating which ``Index`` to build.
-    """
-
-    _init_retrieval = True
-
-    def __init__(self, config, question_encoder_tokenizer, generator_tokenizer):
-        super().__init__()
-        self.index = (
-            LegacyIndex(
-                config.retrieval_vector_size,
-                config.index_path or LEGACY_INDEX_PATH,
-            )
-            if config.index_name == "legacy"
-            else HFIndex(
-                config.dataset, config.dataset_split, config.index_name, config.index_path, config.use_dummy_dataset
-            )
-        )
-        self.generator_tokenizer = generator_tokenizer
-        self.question_encoder_tokenizer = question_encoder_tokenizer
-
-        self.n_docs = config.n_docs
-        self.batch_size = config.retrieval_batch_size
-
-        if torch.cuda.is_available():
-            self.batch_size *= torch.cuda.device_count()
-
-        self.config = config
-        if self._init_retrieval:
-            self.init_retrieval()
-
-    @classmethod
-    def from_pretrained(cls, retriever_name_or_path, **kwargs):
-        config = RagConfig.from_pretrained(retriever_name_or_path, **kwargs)
-        rag_tokenizer = RagTokenizer.from_pretrained(retriever_name_or_path, config=config)
-        question_encoder_tokenizer = rag_tokenizer.question_encoder
-        generator_tokenizer = rag_tokenizer.generator
-        return cls(
-            config, question_encoder_tokenizer=question_encoder_tokenizer, generator_tokenizer=generator_tokenizer
-        )
-
-    def save_pretrained(self, save_directory):
-        self.config.save_pretrained(save_directory)
-        rag_tokenizer = RagTokenizer(
-            question_encoder_tokenizer=self.question_encoder_tokenizer,
-            generator_tokenizer=self.generator_tokenizer,
-        )
-        rag_tokenizer.save_pretrained(save_directory)
-
-    def init_retrieval(self):
-        """
-        Retriever initalization function. It loads the index into memory.
-        """
-
-        logger.info("initializing retrieval")
-        self.index.init_index()
-
-    def postprocess_docs(self, docs, input_strings, prefix, n_docs, return_tensors=None):
-        r"""
-        Postprocessing retrieved ``docs`` and combining them with ``input_strings``.
-
-        Args:
-            doc_scores (:obj:`np.ndarray` of shape :obj:`(batch_size, n_docs)`):
-                Retrieval scores of respective docs - passed for logging.
-            docs  (:obj:`dict`):
-                Retrieved documents.
-            input_strings (:obj:`str`):
-                Input strings decoded by ``preprocess_query``.
-            prefix (:obj:`str`):
-                Prefix added at the beginning of each input, typically used with T5-based models.
-
-        Return:
-            :obj:`tuple(tensors)`:
-                a tuple consisting of two elements: contextualized ``input_ids`` and a compatible ``attention_mask``.
-        """
-
-        def cat_input_and_doc(doc_title, doc_text, input_string, prefix):
-            # TODO(Patrick): if we train more RAG models, I want to put the input first to take advantage of effortless truncation
-            # TODO(piktus): better handling of truncation
-            if doc_title.startswith('"'):
-                doc_title = doc_title[1:]
-            if doc_title.endswith('"'):
-                doc_title = doc_title[:-1]
-            if prefix is None:
-                prefix = ""
-            # TODO(Patrick, piktus, quention) with current master, T5Tokenizer should add eos token => so `add_eos` is not needed anymore
-            #            suffix = self.generator_tokenizer.eos_token if add_eos else ""
-            out = (prefix + doc_title + self.config.title_sep + doc_text + self.config.doc_sep + input_string).replace(
-                "  ", " "
-            )
-            return out
-
-        rag_input_strings = [
-            cat_input_and_doc(
-                docs[i]["title"][j],
-                docs[i]["text"][j],
-                input_strings[i],
-                prefix,
-            )
-            for i in range(len(docs))
-            for j in range(n_docs)
-        ]
-
-        contextualized_inputs = self.generator_tokenizer.batch_encode_plus(
-            rag_input_strings,
-            max_length=self.config.max_combined_length,
-            return_tensors=return_tensors,
-            padding="max_length",
-            truncation=True,
-        )
-
-        return contextualized_inputs["input_ids"], contextualized_inputs["attention_mask"]
-
-    def _chunk_tensor(self, t: Iterable, chunk_size: int) -> List[Iterable]:
-        return [t[i : i + chunk_size] for i in range(0, len(t), chunk_size)]
-
-    def _main_retrieve(self, question_hidden_states: np.ndarray, n_docs: int) -> Tuple[np.ndarray, np.ndarray]:
-        question_hidden_states_batched = self._chunk_tensor(question_hidden_states, self.batch_size)
-        ids_batched = []
-        vectors_batched = []
-        for question_hidden_states in question_hidden_states_batched:
-            start_time = time.time()
-            ids, vectors = self.index.get_top_docs(question_hidden_states, n_docs)
-            logger.debug(
-                "index search time: {} sec, batch size {}".format(
-                    time.time() - start_time, question_hidden_states.shape
-                )
-            )
-            ids_batched.extend(ids)
-            vectors_batched.extend(vectors)
-        return np.array(ids_batched), np.array(
-            vectors_batched
-        )  # shapes (batch_size, n_docs) and (batch_size, n_docs, d)
-
-    def retrieve(self, question_hidden_states: np.ndarray, n_docs: int) -> Tuple[np.ndarray, List[dict]]:
-        """
-        Retrieves documents for specified ``question_hidden_states``.
-
-        Args:
-            question_hidden_states (:obj:`np.ndarray` of shape :obj:`(batch_size, vector_size)`:
-                A batch of query vectors to retrieve with.
-            n_docs (:obj:`int`):
-                The number of docs retrieved per query.
-
-        Ouput:
-            retrieved_doc_embeds (:obj:`np.ndarray` of shape :obj:`(batch_size, n_docs, dim)`
-                The retrieval embeddings of the retrieved docs per query.
-            doc_ids (:obj:`np.ndarray` of shape :obj:`batch_size, n_docs`)
-                The ids of the documents in the index
-            doc_dicts (:obj:`List[dict]`):
-                The retrieved_doc_embeds examples per query.
-        """
-
-        doc_ids, retrieved_doc_embeds = self._main_retrieve(question_hidden_states, n_docs)
-        return retrieved_doc_embeds, doc_ids, self.index.get_doc_dicts(doc_ids)
-
-    def __call__(
-        self,
-        question_input_ids: List[List[int]],
-        question_hidden_states: np.ndarray,
-        prefix: str,
-        n_docs=None,
-        return_tensors=None,
-    ) -> BatchEncoding:
-        """
-        Retrieves documents for specified ``question_hidden_states``.
-
-        Args:
-            question_input_ids: (:obj:`List[List[int]]`) batch of input ids
-            question_hidden_states (:obj:`np.ndarray` of shape :obj:`(batch_size, vector_size)`:
-                A batch of query vectors to retrieve with.
-            n_docs (:obj:`int`):
-                The number of docs retrieved per query.
-
-        Output:
-            :class:`~transformers.BatchEncoding`: A :class:`~transformers.BatchEncoding` with the following fields:
-
-            - **context_input_ids** -- List of token ids to be fed to a model.
-
-              `What are input IDs? <../glossary.html#input-ids>`__
-            - **context_attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
-              :obj:`return_attention_mask=True` or if `"attention_mask"` is in :obj:`self.model_input_names`).
-
-              `What are attention masks? <../glossary.html#attention-mask>`__
-            - **retrieved_doc_embeds** -- List of embeddings of the retrieved documents
-            - **doc_ids** -- List of ids of the retrieved documents
-        """
-
-        n_docs = n_docs if n_docs is not None else self.n_docs
-        retrieved_doc_embeds, doc_ids, docs = self.retrieve(question_hidden_states, n_docs)
-
-        input_strings = self.question_encoder_tokenizer.batch_decode(question_input_ids, skip_special_tokens=True)
-        context_input_ids, context_attention_mask = self.postprocess_docs(
-            docs, input_strings, prefix, n_docs, return_tensors=return_tensors
-        )
-
-        return BatchEncoding(
-            {
-                "context_input_ids": context_input_ids,
-                "context_attention_mask": context_attention_mask,
-                "retrieved_doc_embeds": retrieved_doc_embeds,
-                "doc_ids": doc_ids,
-            },
-            tensor_type=return_tensors,
-        )
-
-
-class RagPyTorchDistributedRetriever(RagRetriever):
-    """
-    A distributed retriever built on top of the ``torch.distributed`` communication package. During training all workers
-    initalize their own instance of the retriever, however, only the main worker loads the index into memory. The index is stored
-    in cpu memory. The index will also work well in a non-distributed setup.
-
-    Args:
-        config (:class:`~transformers.RagConfig`):
-            The configuration of the RAG model this Retriever is used with. Contains parameters indicating which ``Index`` to build.
-        distributed_port (:obj:`int`):
-            The port on which the main communication of the training run is carried out. We set the port for retrieval-related
-            communication as ``distributed_port + 1``.
-    """
-
-    _init_retrieval = False
-
-    def __init__(self, config, question_encoder_tokenizer, generator_tokenizer):
-        super().__init__(
-            config, question_encoder_tokenizer=question_encoder_tokenizer, generator_tokenizer=generator_tokenizer
-        )
-
-        self.process_group = None
-
-    def init_retrieval(self, distributed_port: int):
-        """
-        Retriever initalization function, needs to be called from the training process. The function sets some common parameters
-        and environment variables. On top of that, (only) the main process in the process group loads the index into memory.
-
-        Args:
-            distributed_port (:obj:`int`):
-                The port on which the main communication of the training run is carried out. We set the port for retrieval-related
-                communication as ``distributed_port + 1``.
-        """
-
-        logger.info("initializing retrieval")
-
-        # initializing a separate process group for retrievel as the default
-        # nccl backend doesn't support gather/scatter operations while gloo
-        # is too slow to replace nccl for the core gpu communication
-        if dist.is_initialized():
-            logger.info("dist initialized")
-            # needs to be set manually
-            os.environ["GLOO_SOCKET_IFNAME"] = self._infer_socket_ifname()
-            # avoid clash with the NCCL port
-            os.environ["MASTER_PORT"] = str(distributed_port + 1)
-            self.process_group = dist.new_group(ranks=None, backend="gloo")
-
-        # initialize retriever only on the main worker
-        if not dist.is_initialized() or self._is_main():
-            logger.info("dist not initialized / main")
-            self.index.init_index()
-
-        # all processes wait untill the retriever is initialized by the main process
-        if dist.is_initialized():
-            torch.distributed.barrier(group=self.process_group)
-
-    def _is_main(self):
-        return dist.get_rank(group=self.process_group) == 0
-
-    def _scattered(self, scatter_list, target_shape, target_type=torch.float32):
-        target_tensor = torch.empty(target_shape, dtype=target_type)
-        dist.scatter(target_tensor, src=0, scatter_list=scatter_list, group=self.process_group)
-        return target_tensor
-
-    def _infer_socket_ifname(self):
-        addrs = psutil.net_if_addrs()
-        # a hacky way to deal with varying network interface names
-        ifname = next((addr for addr in addrs if addr.startswith("e")), None)
-        return ifname
-
-    def retrieve(self, question_hidden_states: np.ndarray, n_docs: int) -> Tuple[np.ndarray, List[dict]]:
-        """
-        Retrieves documents for specified ``question_hidden_states``. The main process, which has the access to the index stored in memory, gathers queries
-        from all the processes in the main training process group, performs the retrieval and scatters back the results.
-
-        Args:
-            question_hidden_states (:obj:`np.ndarray` of shape :obj:`(batch_size, vector_size)`:
-                A batch of query vectors to retrieve with.
-            n_docs (:obj:`int`):
-                The number of docs retrieved per query.
-
-        Ouput:
-            retrieved_doc_embeds (:obj:`np.ndarray` of shape :obj:`(batch_size, n_docs, dim)`
-                The retrieval embeddings of the retrieved docs per query.
-            doc_ids (:obj:`np.ndarray` of shape :obj:`batch_size, n_docs`)
-                The ids of the documents in the index
-            doc_dicts (:obj:`List[dict]`):
-                The retrieved_doc_embeds examples per query.
-        """
-
-        # single GPU training
-        if not dist.is_initialized():
-            doc_ids, retrieved_doc_embeds = self._main_retrieve(question_hidden_states, n_docs)
-            return retrieved_doc_embeds, doc_ids, self.index.get_doc_dicts(doc_ids)
-
-        # distributed training
-        world_size = dist.get_world_size(group=self.process_group)
-
-        # gather logic
-        gather_list = None
-        if self._is_main():
-            gather_list = [torch.empty(question_hidden_states.shape, dtype=torch.float32) for _ in range(world_size)]
-        dist.gather(torch.tensor(question_hidden_states), dst=0, gather_list=gather_list, group=self.process_group)
-
-        # scatter logic
-        n_queries = question_hidden_states.shape[0]
-        scatter_ids = []
-        scatter_vectors = []
-        if self._is_main():
-            assert len(gather_list) == world_size
-            ids, vectors = self._main_retrieve(torch.cat(gather_list).numpy(), n_docs)
-            ids, vectors = torch.tensor(ids), torch.tensor(vectors)
-            scatter_ids = self._chunk_tensor(ids, n_queries)
-            scatter_vectors = self._chunk_tensor(vectors, n_queries)
-        doc_ids = self._scattered(scatter_ids, [n_queries, n_docs], target_type=torch.int64)
-        retrieved_doc_embeds = self._scattered(scatter_vectors, [n_queries, n_docs, question_hidden_states.shape[1]])
-
-        return retrieved_doc_embeds.numpy(), doc_ids.numpy(), self.index.get_doc_dicts(doc_ids)

src/transformers/testing_utils.py

@@ -10,7 +10,7 @@ from distutils.util import strtobool
 from io import StringIO
 from pathlib import Path
 
-from .file_utils import _datasets_available, _faiss_available, _tf_available, _torch_available, _torch_tpu_available
+from .file_utils import _tf_available, _torch_available, _torch_tpu_available
 
 
 SMALL_MODEL_IDENTIFIER = "julien-c/bert-xsmall-dummy"
@@ -161,30 +161,6 @@ def require_torch_and_cuda(test_case):
         return test_case
 
 
-def require_datasets(test_case):
-    """
-    Decorator marking a test that requires TensorFlow.
-
-    These tests are skipped when TensorFlow isn't installed.
-
-    """
-    if not _datasets_available:
-        test_case = unittest.skip("test requires Datasets")(test_case)
-    return test_case
-
-
-def require_faiss(test_case):
-    """
-    Decorator marking a test that requires TensorFlow.
-
-    These tests are skipped when TensorFlow isn't installed.
-
-    """
-    if not _faiss_available:
-        test_case = unittest.skip("test requires Faiss")(test_case)
-    return test_case
-
-
 def get_tests_dir():
     """
     returns the full path to the `tests` dir, so that the tests can be invoked from anywhere

src/transformers/tokenization_auto.py

@@ -26,7 +26,6 @@ from .configuration_auto import (
     CamembertConfig,
     CTRLConfig,
     DistilBertConfig,
-    DPRConfig,
     ElectraConfig,
     EncoderDecoderConfig,
     FlaubertConfig,
@@ -58,7 +57,6 @@ from .tokenization_bert_japanese import BertJapaneseTokenizer
 from .tokenization_camembert import CamembertTokenizer
 from .tokenization_ctrl import CTRLTokenizer
 from .tokenization_distilbert import DistilBertTokenizer, DistilBertTokenizerFast
-from .tokenization_dpr import DPRQuestionEncoderTokenizer, DPRQuestionEncoderTokenizerFast
 from .tokenization_electra import ElectraTokenizer, ElectraTokenizerFast
 from .tokenization_flaubert import FlaubertTokenizer
 from .tokenization_funnel import FunnelTokenizer, FunnelTokenizerFast
@@ -112,7 +110,6 @@ TOKENIZER_MAPPING = OrderedDict(
         (XLMConfig, (XLMTokenizer, None)),
         (CTRLConfig, (CTRLTokenizer, None)),
         (BertGenerationConfig, (BertGenerationTokenizer, None)),
-        (DPRConfig, (DPRQuestionEncoderTokenizer, DPRQuestionEncoderTokenizerFast)),
     ]
 )
 

src/transformers/tokenization_rag.py

@@ -1,60 +0,0 @@
-# coding=utf-8
-# Copyright 2020, The RAG Authors and The HuggingFace Inc. team.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""Tokenization classes for RAG."""
-import os
-
-from .configuration_rag import RagConfig
-from .tokenization_auto import AutoTokenizer
-from .utils import logging
-
-
-logger = logging.get_logger(__name__)
-
-
-class RagTokenizer:
-    def __init__(self, question_encoder, generator):
-        self.question_encoder = question_encoder
-        self.generator = generator
-
-    def save_pretrained(self, save_directory):
-        if os.path.isfile(save_directory):
-            logger.error("Provided path ({}) should be a directory, not a file".format(save_directory))
-            return
-        os.makedirs(save_directory, exist_ok=True)
-        question_encoder_path = os.path.join(save_directory, "question_encoder")
-        generator_path = os.path.join(save_directory, "generator")
-        self.question_encoder.save_pretrained(question_encoder_path)
-        self.generator.save_pretrained(generator_path)
-
-    @classmethod
-    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
-        config = kwargs.pop("config", None)
-
-        if config is None:
-            config = RagConfig.from_pretrained(pretrained_model_name_or_path)
-
-        question_encoder_path = os.path.join(pretrained_model_name_or_path, "question_encoder_tokenizer")
-        generator_path = os.path.join(pretrained_model_name_or_path, "generator_tokenizer")
-        question_encoder = AutoTokenizer.from_pretrained(question_encoder_path, config=config.question_encoder)
-        generator = AutoTokenizer.from_pretrained(generator_path, config=config.generator)
-        return cls(question_encoder=question_encoder, generator=generator)
-
-    def __call__(self, *args, **kwargs):
-        return self.question_encoder(*args, **kwargs)
-
-    def batch_decode(self, *args, **kwargs):
-        return self.generator.batch_decode(*args, **kwargs)
-
-    # TODO(Patrick) add prepare_seq2seq_batch function

templates/examples/type/scripts/explicit_example.py

@@ -0,0 +1,63 @@
+import torch
+from torch.utils.data import DataLoader
+from datasets import load_dataset, load_metric
+from transformers import AutoTokenizer, AutoModelForSequenceClassification, AdamW, logging
+
+logger = logging.get_logger(__name__)
+
+model_name = 'bert-base-cased'
+dataset_name, dataset_subset = 'glue', 'mrpc'
+device = 'cuda' if torch.cuda.is_available() else 'cpu'
+
+tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
+model = AutoModelForSequenceClassification.from_pretrained(model_name)
+dataset = load_dataset(dataset_name, dataset_subset)
+metric = load_metric(dataset_name, dataset_subset)
+
+
+# Prepare the dataset: encode it as input to our model and truncate to max model length
+def encode(examples):
+    output = tokenizer(examples['sentence1'], examples['sentence2'], truncation=True)
+    output['labels'] = examples['label']
+    return output
+
+encoded_dataset = dataset.map(encode, batched=True)
+encoded_dataset.set_format(columns=['input_ids', 'token_type_ids', 'labels'])
+
+def train_epoch(model, optimizer, train_dataloader, eval_dataloader, device):
+    model.to(device)
+    model.train()
+    for batch in train_dataloader:
+        batch.to(device)
+        model(**batch)
+        loss = model.losss
+        loss.backward()
+        optimizer.step()
+
+    model.eval()
+    for batch in eval_dataloader:
+        batch.to(device)
+        outputs = model(**batch)
+        predictions = outputs.logits.argmax(dim=-1)
+        metric.add_batch(predictions=predictions, references=batch['labels'])
+    scores = metric.compute()
+    return scores
+
+learning_rate = 1e-5
+train_batch_size = 16
+seed = 42
+epochs = 2
+
+# Prepare pytorch dataloader (with dynamic batch i.e. pad the sequences to the longest in the batch)
+def collate_fn(examples):
+  return tokenizer.pad(examples, padding='longest', return_tensors='pt')
+
+train_dataloader = DataLoader(dataset['train'], shuffle=True, collate_fn=collate_fn, batch_size=32)
+valid_dataloader = DataLoader(dataset['validation'], shuffle=False, collate_fn=collate_fn, batch_size=16)
+
+optimizer = AdamW(params=model.parameters(), lr=learning_rate)
+
+for epoch in range(epochs):
+    eval_metrics = train_epoch(model, optimizer, train_dataloader, valid_dataloader)
+    print(f"eval_metrics for trial {trial.number} epoch {epoch}:", eval_metrics)
+    sum_metrics = sum(eval_metrics.values())

templates/examples/type/scripts/{{cookiecutter.application}}_{{cookiecutter.style}}.py

@@ -0,0 +1,127 @@
+from transformers import AutoTokenizer, AdamW, get_constant_schedule_with_warmup, get_linear_schedule_with_warmup, {{cookiecutter.model_class_name}}
+from datasets import load_dataset, load_metric, temp_seed
+from torch.utils.data import DataLoader
+import argparse
+import os
+
+MAX_GPU_BATCH_SIZE = 8
+EVAL_BATCH_SIZE = 16
+
+def prepare_dataset(args):
+    """ Load a dataset, a tokenizer and encode the dataset """
+    tokenizer = AutoTokenizer.from_pretrained(args.model_name)
+
+    dataset = load_dataset(args.dataset_name, args.dataset_config_name)
+    # If you have a custom dataset in a CSV/JSON/TEXT file
+    # dataset = load_dataset("csv", data_files={'train': "train_data.csv", 'validation': "valid_data.csv"})
+
+    # Prepare the dataset (cached)
+    def encode(example):
+        output = tokenizer(example['sentence1'], example['sentence2'], truncation=True)
+        output['labels'] = example['label']
+        return output
+
+    dataset = dataset.map(encode)
+    dataset.set_format(columns=['attention_mask', 'input_ids', 'token_type_ids', 'labels'])
+
+    return dataset, tokenizer
+
+
+def train_and_evaluate(args):
+    # Load dataset, tokenizer and metric
+    dataset, tokenizer = prepare_dataset(args)
+    metric = load_metric(args.dataset_name, args.dataset_config_name)
+
+    # Instantiate train and evaluation dataloaders
+    def collate_fn(examples):
+        """ Collate function which pads the batches to the longest sequence in the batch and returns PyTorch tensors """
+        return tokenizer.pad(examples, padding='longest', return_tensors='pt')
+
+    train_dataloader = DataLoader(dataset['train'], shuffle=True, collate_fn=collate_fn, batch_size=args.train_batch_size)
+    eval_dataloader = DataLoader(dataset['validation'], shuffle=False, collate_fn=collate_fn, batch_size=EVAL_BATCH_SIZE)
+
+    # Instantiate model
+    model = {{cookiecutter.model_class_name}}.from_pretrained(args.model_name_or_path, return_dict=True)
+
+    # Instantiate optimizer and learning rate schedule
+    optimizer = AdamW(params=model.parameters(), lr=args.lr, correct_bias=args.correct_bias)
+    if args.lr_schedule == 'constant':
+        lr_scheduler = get_constant_schedule_with_warmup(optimizer=optimizer, num_warmup_steps=100)
+    else:
+        lr_scheduler = get_linear_schedule_with_warmup(optimizer=optimizer, num_warmup_steps=100, num_training_steps=len(train_dataloader)*args.num_epochs)
+
+    # Train the model
+    eval_metrics = {}
+    for epoch in range(args.num_epochs):
+        model.train()
+        model.to(args.device)
+        for step, batch in enumerate(train_dataloader):
+            batch.to(args.device)
+            outputs = model(**batch)
+            loss = outputs.loss
+            loss = loss / args.gradient_accumulation_steps
+            loss.backward()
+            if step % args.gradient_accumulation_steps == 0:
+                optimizer.step()
+                lr_scheduler.step()
+                optimizer.zero_grad()
+
+        model.eval()
+        for step, batch in enumerate(eval_dataloader):
+            batch.to(args.device)
+            outputs = model(**batch)
+            predictions = outputs.logits.argmax(dim=-1)
+            metric.add_batch(predictions=predictions, references=batch['labels'])
+
+        eval_metrics = metric.compute()
+        print(f"eval_metric for epoch {epoch}:", eval_metrics)
+
+    return eval_metrics
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_name_or_path", default=None, type=str, required=True, help="Path to pre-trained model or shortcut name")
+    parser.add_argument("--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name")
+    parser.add_argument("--tokenizer_name", default="", type=str, help="Pretrained tokenizer name or path if not the same as model_name")
+    parser.add_argument("--task_name", default=None, type=str, required=True, help="The name of the task to train.")
+    parser.add_argument("--max_seq_length", default=128, type=int, help="The maximum total input sequence length after tokenization. Sequences longer than this will be truncated, sequences shorter will be padded.")
+    parser.add_argument("--output_dir", default=None, type=str, required=True, help="The output directory where the model predictions and checkpoints will be written.")
+    parser.add_argument("--overwrite_output_dir", default=False, type=bool, help="Overwrite the content of the output directory. Use this to continue training if output_dir points to a checkpoint directory.")
+    parser.add_argument("--do_train", action='store_true', help="Whether to run training.")
+    parser.add_argument("--do_eval", action='store_true', help="Whether to run eval on the dev set.")
+    parser.add_argument("--evaluate_during_training", action='store_true', help="Run evaluation during training at each logging step.")
+    parser.add_argument("--prediction_loss_only", action='store_true', help="When performing evaluation and predictions, only returns the loss.")
+    parser.add_argument("--per_device_train_batch_size", default=8, type=int, help="Batch size per GPU/TPU core/CPU for training.")
+    parser.add_argument("--per_device_eval_batch_size", default=8, type=int, help="Batch size per GPU/TPU core/CPU for evaluation.")
+    parser.add_argument('--gradient_accumulation_steps', type=int, default=1, help="Number of updates steps to accumulate before performing a backward/update pass.")     
+    parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
+    parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
+    parser.add_argument("--adam_beta1", default=0.9, type=float, help="Beta1 for Adam optimizer.")
+    parser.add_argument("--adam_beta2", default=0.999, type=float, help="Beta2 for Adam optimizer.")
+    parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
+    parser.add_argument("--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform.")
+    parser.add_argument("--max_steps", default=-1, type=int, help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
+    parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
+    parser.add_argument("--logging_dir", type=str, default_factory=default_logdir, help="Tensorboard log dir.")
+    parser.add_argument("--logging_first_step", action='store_true', help="Log and eval the first global_step")
+    parser.add_argument('--logging_steps', type=int, default=500, help="Log every X updates steps.")
+    parser.add_argument('--save_steps', type=int, default=500, help="Save checkpoint every X updates steps.")
+    parser.add_argument('--save_total_limit', type=int, default=None, help="Limit the total amount of checkpoints. Deletes the older checkpoints in the output_dir. Default is unlimited checkpoints")
+    parser.add_argument('--seed', type=int, default=42, help="random seed for initialization")
+    parser.add_argument("--no_cuda", action='store_true', help="Avoid using CUDA when available")
+    parser.add_argument('--fp16', action='store_true', help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
+    parser.add_argument('--fp16_opt_level', type=str, default='O1', help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']. See details at https://nvidia.github.io/apex/amp.html")
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+    parser.add_argument("--dataloader_drop_last", action='store_true', help="Drop the last incomplete batch if it is not divisible by the batch size.")
+    parser.add_argument("--eval_steps", type=int, default=1000, help="Run an evaluation every X steps.")
+    parser.add_argument("--past_index", type=int, default=-1, help="If >=0, uses the corresponding part of the output as the past state for next step.")
+    parser.add_argument("--run_name", type=str, default=None, help="An optional descriptor for the run. Notably used for wandb logging.")
+    parser.add_argument("--disable_tqdm", action='store_true', help="Whether or not to disable the tqdm progress bars.")
+    args = parser.parse_args()
+
+    with temp_seed(args.seed, set_pytorch=True):
+        return train_and_evaluate(args)
+
+if __name__ == "__main__":
+    main()

tests/test_modeling_rag.py

@@ -1,823 +0,0 @@
-# coding=utf-8
-# Copyright 2020, The RAG Authors and The HuggingFace Inc. team.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-
-import json
-import os
-import shutil
-import tempfile
-import unittest
-from unittest.mock import patch
-
-import numpy as np
-
-from transformers.file_utils import (
-    cached_property,
-    is_datasets_available,
-    is_faiss_available,
-    is_psutil_available,
-    is_torch_available,
-)
-from transformers.modeling_outputs import BaseModelOutput
-from transformers.testing_utils import require_torch, slow, torch_device
-from transformers.tokenization_bart import BartTokenizer
-from transformers.tokenization_bert import VOCAB_FILES_NAMES as DPR_VOCAB_FILES_NAMES
-from transformers.tokenization_dpr import DPRQuestionEncoderTokenizer
-from transformers.tokenization_roberta import VOCAB_FILES_NAMES as BART_VOCAB_FILES_NAMES
-
-from .test_modeling_bart import ModelTester as BartModelTester
-from .test_modeling_common import ids_tensor
-from .test_modeling_dpr import DPRModelTester
-
-
-TOLERANCE = 1e-4
-
-
-if is_torch_available() and is_datasets_available() and is_faiss_available() and is_psutil_available():
-    import faiss
-    import torch
-    from datasets import Dataset
-
-    from transformers import (
-        AutoConfig,
-        AutoModel,
-        AutoModelForSeq2SeqLM,
-        BartConfig,
-        DPRConfig,
-        RagConfig,
-        RagModel,
-        RagRetriever,
-        RagSequenceForGeneration,
-        RagTokenForGeneration,
-    )
-
-
-def _assert_tensors_equal(a, b, atol=1e-12, prefix=""):
-    """If tensors not close, or a and b arent both tensors, raise a nice Assertion error."""
-    if a is None and b is None:
-        return True
-    try:
-        if torch.allclose(a, b, atol=atol):
-            return True
-        raise
-    except Exception:
-        msg = "{} != {}".format(a, b)
-        if prefix:
-            msg = prefix + ": " + msg
-        raise AssertionError(msg)
-
-
-def require_retrieval(test_case):
-    """
-    Decorator marking a test that requires a set of dependencies necessary for pefrorm retrieval with
-    :class:`~transformers.RagRetriever`.
-
-    These tests are skipped when respective libraries are not installed.
-
-    """
-    if not (is_torch_available() and is_datasets_available() and is_faiss_available() and is_psutil_available()):
-        test_case = unittest.skip("test requires PyTorch")(test_case)
-    return test_case
-
-
-class RagModelTester:
-    def __init__(
-        self,
-        parent,
-    ):
-        # Global params
-        self.parent = parent
-        self.batch_size = 13
-        self.seq_length = 7
-
-        # RAG params
-        self.n_docs = 3
-        self.vocab_size = 50265
-        self.bos_token_id = 0
-        self.pad_token_id = 1
-        self.eos_token_id = 2
-        self.decoder_start_token_id = 2
-        self.max_combined_length = 123
-        self.retrieval_vector_size = 768
-        self.retrieval_batch_size = 8
-
-        self.rag_config = RagConfig(
-            n_docs=self.n_docs,
-            vocab_size=self.vocab_size,
-            bos_token_id=self.bos_token_id,
-            pad_token_id=self.pad_token_id,
-            eos_token_id=self.eos_token_id,
-            decoder_start_token_id=self.decoder_start_token_id,
-            max_combined_length=self.max_combined_length,
-            retrieval_vector_size=self.retrieval_vector_size,
-            retrieval_batch_size=self.retrieval_batch_size,
-        )
-
-        # BART params
-        self.hidden_size = 16
-        self.num_hidden_layers = 2
-        self.num_attention_heads = 4
-        self.intermediate_size = 4
-        self.hidden_dropout_prob = 0.1
-        self.attention_probs_dropout_prob = 0.1
-        self.max_position_embeddings = 20
-
-        self.bart_config = BartConfig(
-            vocab_size=self.vocab_size,
-            d_model=self.hidden_size,
-            encoder_layers=self.num_hidden_layers,
-            decoder_layers=self.num_hidden_layers,
-            encoder_attention_heads=self.num_attention_heads,
-            decoder_attention_heads=self.num_attention_heads,
-            encoder_ffn_dim=self.intermediate_size,
-            decoder_ffn_dim=self.intermediate_size,
-            dropout=self.hidden_dropout_prob,
-            attention_dropout=self.attention_probs_dropout_prob,
-            max_position_embeddings=self.max_position_embeddings,
-            eos_token_id=self.eos_token_id,
-            bos_token_id=self.bos_token_id,
-            pad_token_id=self.pad_token_id,
-            decoder_start_token_id=self.decoder_start_token_id,
-        )
-
-        # DPR params
-        self.dpr_vocab_size = 51
-        self.hidden_size = 20
-        self.num_hidden_layers = 3
-        self.num_attention_heads = 5
-        self.intermediate_size = 5
-        self.hidden_act = "gelu"
-        self.hidden_dropout_prob = 0.2
-        self.attention_probs_dropout_prob = 0.2
-        self.max_position_embeddings = 19
-        self.type_vocab_size = 17
-        self.initializer_range = 0.02
-        self.projection_dim = 0
-
-        self.dpr_config = DPRConfig(
-            projection_dim=self.projection_dim,
-            vocab_size=self.dpr_vocab_size,
-            hidden_size=self.hidden_size,
-            num_hidden_layers=self.num_hidden_layers,
-            num_attention_heads=self.num_attention_heads,
-            intermediate_size=self.intermediate_size,
-            hidden_act=self.hidden_act,
-            hidden_dropout_prob=self.hidden_dropout_prob,
-            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
-            max_position_embeddings=self.max_position_embeddings,
-            type_vocab_size=self.type_vocab_size,
-            is_decoder=False,
-            initializer_range=self.initializer_range,
-        )
-
-    def prepare_inputs(self):
-        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp(
-            3,
-        )
-        input_ids[:, -1] = self.eos_token_id
-        attention_mask = input_ids.ne(self.pad_token_id)
-
-        return input_ids, attention_mask
-
-
-@require_torch
-@require_retrieval
-class RagTestMixin:
-
-    all_model_classes = (
-        (RagModel, RagTokenForGeneration, RagSequenceForGeneration)
-        if is_torch_available() and is_datasets_available() and is_faiss_available() and is_psutil_available()
-        else ()
-    )
-
-    retrieval_vector_size = 32
-    n_docs = 2
-    max_combined_length = 16
-
-    def setUp(self):
-        self.tmpdirname = tempfile.mkdtemp()
-
-        # DPR tok
-        vocab_tokens = [
-            "[UNK]",
-            "[CLS]",
-            "[SEP]",
-            "[PAD]",
-            "[MASK]",
-            "want",
-            "##want",
-            "##ed",
-            "wa",
-            "un",
-            "runn",
-            "##ing",
-            ",",
-            "low",
-            "lowest",
-        ]
-        dpr_tokenizer_path = os.path.join(self.tmpdirname, "dpr_tokenizer")
-        os.makedirs(dpr_tokenizer_path, exist_ok=True)
-        self.vocab_file = os.path.join(dpr_tokenizer_path, DPR_VOCAB_FILES_NAMES["vocab_file"])
-        with open(self.vocab_file, "w", encoding="utf-8") as vocab_writer:
-            vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
-
-        # BART tok
-        vocab = [
-            "l",
-            "o",
-            "w",
-            "e",
-            "r",
-            "s",
-            "t",
-            "i",
-            "d",
-            "n",
-            "\u0120",
-            "\u0120l",
-            "\u0120n",
-            "\u0120lo",
-            "\u0120low",
-            "er",
-            "\u0120lowest",
-            "\u0120newer",
-            "\u0120wider",
-            "<unk>",
-        ]
-        vocab_tokens = dict(zip(vocab, range(len(vocab))))
-        merges = ["#version: 0.2", "\u0120 l", "\u0120l o", "\u0120lo w", "e r", ""]
-        self.special_tokens_map = {"unk_token": "<unk>"}
-
-        bart_tokenizer_path = os.path.join(self.tmpdirname, "bart_tokenizer")
-        os.makedirs(bart_tokenizer_path, exist_ok=True)
-        self.vocab_file = os.path.join(bart_tokenizer_path, BART_VOCAB_FILES_NAMES["vocab_file"])
-        self.merges_file = os.path.join(bart_tokenizer_path, BART_VOCAB_FILES_NAMES["merges_file"])
-        with open(self.vocab_file, "w", encoding="utf-8") as fp:
-            fp.write(json.dumps(vocab_tokens) + "\n")
-        with open(self.merges_file, "w", encoding="utf-8") as fp:
-            fp.write("\n".join(merges))
-
-    @cached_property
-    def dpr_tokenizer(self) -> DPRQuestionEncoderTokenizer:
-        return DPRQuestionEncoderTokenizer.from_pretrained(os.path.join(self.tmpdirname, "dpr_tokenizer"))
-
-    @cached_property
-    def bart_tokenizer(self) -> BartTokenizer:
-        return BartTokenizer.from_pretrained(os.path.join(self.tmpdirname, "bart_tokenizer"))
-
-    def tearDown(self):
-        shutil.rmtree(self.tmpdirname)
-
-    def get_retriever(self, config):
-        dataset = Dataset.from_dict(
-            {
-                "id": ["0", "1"],
-                "text": ["foo", "bar"],
-                "title": ["Foo", "Bar"],
-                "embeddings": [np.ones(self.retrieval_vector_size), 2 * np.ones(self.retrieval_vector_size)],
-            }
-        )
-        dataset.add_faiss_index("embeddings", string_factory="Flat", metric_type=faiss.METRIC_INNER_PRODUCT)
-        with patch("transformers.retrieval_rag.load_dataset") as mock_load_dataset:
-            mock_load_dataset.return_value = dataset
-            retriever = RagRetriever(
-                config,
-                question_encoder_tokenizer=self.dpr_tokenizer,
-                generator_tokenizer=self.bart_tokenizer,
-            )
-        return retriever
-
-    def check_model_with_retriever(
-        self, config, input_ids, attention_mask, decoder_input_ids, decoder_attention_mask, **kwargs
-    ):
-        self.assertIsNotNone(config.question_encoder)
-        self.assertIsNotNone(config.generator)
-
-        for model_class in self.all_model_classes:
-            model = model_class(config, retriever=self.get_retriever(config)).to(torch_device)
-            model.eval()
-
-            self.assertTrue(model.config.is_encoder_decoder)
-
-            outputs = model(
-                input_ids=input_ids,
-                attention_mask=attention_mask,
-                decoder_input_ids=decoder_input_ids,
-                decoder_attention_mask=decoder_attention_mask,
-            )
-
-            # logits
-            self.assertEqual(
-                outputs.logits.shape,
-                (self.n_docs * decoder_input_ids.shape[0], decoder_input_ids.shape[1], config.generator.vocab_size),
-            )
-            # generator encoder last hidden states
-            self.assertEqual(
-                outputs.generator_enc_last_hidden_state.shape,
-                (self.n_docs * decoder_input_ids.shape[0], self.max_combined_length, config.generator.hidden_size),
-            )
-            # doc scores
-            self.assertEqual(outputs.doc_scores.shape, (input_ids.shape[0], self.n_docs))
-
-    def check_model_generate(
-        self, config, input_ids, attention_mask, decoder_input_ids, decoder_attention_mask, **kwargs
-    ):
-        self.assertIsNotNone(config.question_encoder)
-        self.assertIsNotNone(config.generator)
-
-        for model_class in self.all_model_classes[1:]:
-            model = model_class(config, retriever=self.get_retriever(config)).to(torch_device)
-            model.eval()
-
-            self.assertTrue(model.config.is_encoder_decoder)
-
-            outputs = model.generate(
-                input_ids=input_ids,
-                num_beams=2,
-                num_return_sequences=2,
-                decoder_start_token_id=config.generator.eos_token_id,
-            )
-
-    def check_model_without_retriever(
-        self, config, input_ids, attention_mask, decoder_input_ids, decoder_attention_mask, **kwargs
-    ):
-        self.assertIsNotNone(config.question_encoder)
-        self.assertIsNotNone(config.generator)
-
-        retriever = self.get_retriever(config)
-
-        for model_class in self.all_model_classes:
-            model = model_class(config).to(torch_device)
-            model.eval()
-            self.assertTrue(model.config.is_encoder_decoder)
-
-            question_hidden_states = model.question_encoder(input_ids, attention_mask=attention_mask)[0]
-
-            out = retriever(
-                input_ids,
-                question_hidden_states.cpu().detach().to(torch.float32).numpy(),
-                prefix=config.generator.prefix,
-                return_tensors="pt",
-            )
-
-            context_input_ids, context_attention_mask, retrieved_doc_embeds = (
-                out["context_input_ids"],
-                out["context_attention_mask"],
-                out["retrieved_doc_embeds"],
-            )
-
-            # cast
-            retrieved_doc_embeds = retrieved_doc_embeds.to(question_hidden_states)
-            context_input_ids = context_input_ids.to(input_ids)
-            context_attention_mask = context_attention_mask.to(input_ids)
-
-            # compute doc_scores
-            doc_scores = torch.bmm(question_hidden_states.unsqueeze(1), retrieved_doc_embeds.transpose(1, 2)).squeeze(
-                1
-            )
-
-            outputs = model(
-                context_input_ids=context_input_ids,
-                context_attention_mask=context_attention_mask,
-                doc_scores=doc_scores,
-                decoder_input_ids=decoder_input_ids,
-                decoder_attention_mask=decoder_attention_mask,
-            )
-
-            # logits
-            self.assertEqual(
-                outputs.logits.shape,
-                (self.n_docs * decoder_input_ids.shape[0], decoder_input_ids.shape[1], config.generator.vocab_size),
-            )
-            # generator encoder last hidden states
-            self.assertEqual(
-                outputs.generator_enc_last_hidden_state.shape,
-                (self.n_docs * decoder_input_ids.shape[0], self.max_combined_length, config.generator.hidden_size),
-            )
-            # doc scores
-            self.assertEqual(outputs.doc_scores.shape, (input_ids.shape[0], self.n_docs))
-
-    def check_model_with_encoder_outputs(
-        self, config, input_ids, attention_mask, decoder_input_ids, decoder_attention_mask, **kwargs
-    ):
-        self.assertIsNotNone(config.question_encoder)
-        self.assertIsNotNone(config.generator)
-
-        for model_class in self.all_model_classes:
-            model = model_class(config, retriever=self.get_retriever(config)).to(torch_device)
-            model.eval()
-
-            self.assertTrue(model.config.is_encoder_decoder)
-
-            outputs = model(
-                input_ids=input_ids,
-                attention_mask=attention_mask,
-                decoder_input_ids=decoder_input_ids,
-                decoder_attention_mask=decoder_attention_mask,
-            )
-
-            encoder_outputs = BaseModelOutput(outputs.generator_enc_last_hidden_state)
-
-            # run only generator
-            outputs = model(
-                encoder_outputs=encoder_outputs,
-                doc_scores=outputs.doc_scores,
-                decoder_input_ids=decoder_input_ids,
-                decoder_attention_mask=decoder_attention_mask,
-            )
-
-            # logits
-            self.assertEqual(
-                outputs.logits.shape,
-                (self.n_docs * decoder_input_ids.shape[0], decoder_input_ids.shape[1], config.generator.vocab_size),
-            )
-            # generator encoder last hidden states
-            self.assertEqual(
-                outputs.generator_enc_last_hidden_state.shape,
-                (self.n_docs * decoder_input_ids.shape[0], self.max_combined_length, config.generator.hidden_size),
-            )
-            # doc scores
-            self.assertEqual(outputs.doc_scores.shape, (input_ids.shape[0], self.n_docs))
-
-    def test_model_with_retriever(self):
-        inputs_dict = self.config_and_inputs
-        self.check_model_with_retriever(**inputs_dict)
-
-    def test_model_without_retriever(self):
-        inputs_dict = self.config_and_inputs
-        self.check_model_without_retriever(**inputs_dict)
-
-    def test_model_with_encoder_outputs(self):
-        inputs_dict = self.config_and_inputs
-        self.check_model_with_encoder_outputs(**inputs_dict)
-
-    def test_model_generate(self):
-        inputs_dict = self.config_and_inputs
-        self.check_model_generate(**inputs_dict)
-
-
-@require_torch
-@require_retrieval
-class RagDPRBartTest(RagTestMixin, unittest.TestCase):
-    @cached_property
-    def config_and_inputs(self):
-        question_encoder_tester = DPRModelTester(self)
-        dpr_config_and_inputs = question_encoder_tester.prepare_config_and_inputs()
-        generator_tester = BartModelTester(self)
-        bart_config_and_inputs = generator_tester.prepare_config_and_inputs_for_common()
-
-        (question_encoder_config, input_ids, _, input_mask, _, _, _) = dpr_config_and_inputs
-        (generator_config, bart_inputs_dict) = bart_config_and_inputs
-        decoder_input_ids, decoder_attention_mask = bart_inputs_dict["input_ids"], bart_inputs_dict["attention_mask"]
-
-        config = RagConfig.from_question_encoder_generator_configs(
-            question_encoder_config,
-            generator_config,
-            n_docs=self.n_docs,
-            retrieval_vector_size=self.retrieval_vector_size,
-            max_combined_length=self.max_combined_length,
-            use_cache=False,
-        )
-
-        return {
-            "config": config,
-            "input_ids": input_ids,
-            "attention_mask": input_mask,
-            "decoder_input_ids": decoder_input_ids,
-            "decoder_attention_mask": decoder_attention_mask,
-        }
-
-
-@require_torch
-@require_retrieval
-class RagModelIntegrationTests(unittest.TestCase):
-    @cached_property
-    def sequence_model(self):
-        return RagSequenceForGeneration.from_pretrained_question_encoder_generator(
-            "facebook/dpr-question_encoder-single-nq-base", "facebook/bart-large-cnn"
-        ).to(torch_device)
-
-    @cached_property
-    def token_model(self):
-        return RagTokenForGeneration.from_pretrained_question_encoder_generator(
-            "facebook/dpr-question_encoder-single-nq-base", "facebook/bart-large-cnn"
-        ).to(torch_device)
-
-    def get_rag_config(self):
-        question_encoder_config = AutoConfig.from_pretrained("facebook/dpr-question_encoder-single-nq-base")
-        generator_config = AutoConfig.from_pretrained("facebook/bart-large-cnn")
-        return RagConfig.from_question_encoder_generator_configs(
-            question_encoder_config,
-            generator_config,
-            bos_token_id=0,
-            decoder_start_token_id=2,
-            eos_token_id=2,
-            is_encoder_decoder=True,
-            pad_token_id=1,
-            vocab_size=50264,
-            title_sep=" / ",
-            doc_sep=" // ",
-            n_docs=5,
-            max_combined_length=300,
-            dataset="wiki_dpr",
-            dataset_split="train",
-            index_name="exact",
-            index_path=None,
-            use_dummy_dataset=True,
-            retrieval_vector_size=768,
-            retrieval_batch_size=8,
-        )
-
-    @slow
-    def test_rag_sequence_inference(self):
-        rag_config = self.get_rag_config()
-        rag_decoder_tokenizer = BartTokenizer.from_pretrained("facebook/bart-large-cnn")
-        rag_question_encoder_tokenizer = DPRQuestionEncoderTokenizer.from_pretrained(
-            "facebook/dpr-question_encoder-single-nq-base"
-        )
-        rag_retriever = RagRetriever(
-            rag_config,
-            question_encoder_tokenizer=rag_question_encoder_tokenizer,
-            generator_tokenizer=rag_decoder_tokenizer,
-        )
-
-        rag_sequence = self.sequence_model
-        rag_sequence.set_retriever(rag_retriever)
-
-        input_ids = rag_question_encoder_tokenizer(
-            "who sings does he love me with reba", return_tensors="pt"
-        ).input_ids
-        decoder_input_ids = rag_decoder_tokenizer("Linda Davis", return_tensors="pt").input_ids
-
-        input_ids = input_ids.to(torch_device)
-        decoder_input_ids = decoder_input_ids.to(torch_device)
-
-        with torch.no_grad():
-            output = rag_sequence(
-                input_ids,
-                labels=decoder_input_ids,
-            )
-
-        expected_shape = torch.Size([5, 5, 50264])
-        self.assertEqual(output.logits.shape, expected_shape)
-
-        expected_doc_scores = torch.tensor([[75.0286, 74.4998, 74.0804, 74.0306, 73.9504]])
-        _assert_tensors_equal(expected_doc_scores, output.doc_scores, atol=TOLERANCE)
-
-        expected_loss = torch.tensor([38.7446])
-        _assert_tensors_equal(expected_loss, output.loss, atol=TOLERANCE)
-
-    @slow
-    def test_rag_token_inference(self):
-        rag_config = self.get_rag_config()
-        rag_decoder_tokenizer = BartTokenizer.from_pretrained("facebook/bart-large-cnn")
-        rag_question_encoder_tokenizer = DPRQuestionEncoderTokenizer.from_pretrained(
-            "facebook/dpr-question_encoder-single-nq-base"
-        )
-        rag_retriever = RagRetriever(
-            rag_config,
-            question_encoder_tokenizer=rag_question_encoder_tokenizer,
-            generator_tokenizer=rag_decoder_tokenizer,
-        )
-
-        rag_token = self.token_model
-        rag_token.set_retriever(rag_retriever)
-
-        input_ids = rag_question_encoder_tokenizer(
-            "who sings does he love me with reba", return_tensors="pt"
-        ).input_ids
-        decoder_input_ids = rag_decoder_tokenizer("Linda Davis", return_tensors="pt").input_ids
-
-        input_ids = input_ids.to(torch_device)
-        decoder_input_ids = decoder_input_ids.to(torch_device)
-
-        with torch.no_grad():
-            output = rag_token(
-                input_ids,
-                labels=decoder_input_ids,
-            )
-
-        expected_shape = torch.Size([5, 5, 50264])
-        self.assertEqual(output.logits.shape, expected_shape)
-
-        expected_doc_scores = torch.tensor([[75.0286, 74.4998, 74.0804, 74.0306, 73.9504]])
-        _assert_tensors_equal(expected_doc_scores, output.doc_scores, atol=TOLERANCE)
-
-        expected_loss = torch.tensor([38.7045])
-        _assert_tensors_equal(expected_loss, output.loss, atol=TOLERANCE)
-
-    @slow
-    def test_rag_sequence_generate(self):
-        rag_config = self.get_rag_config()
-        rag_decoder_tokenizer = BartTokenizer.from_pretrained("facebook/bart-large-cnn")
-        rag_question_encoder_tokenizer = DPRQuestionEncoderTokenizer.from_pretrained(
-            "facebook/dpr-question_encoder-single-nq-base"
-        )
-        rag_retriever = RagRetriever(
-            rag_config,
-            question_encoder_tokenizer=rag_question_encoder_tokenizer,
-            generator_tokenizer=rag_decoder_tokenizer,
-        )
-
-        rag_sequence = self.sequence_model
-        rag_sequence.set_retriever(rag_retriever)
-
-        input_ids = rag_question_encoder_tokenizer(
-            "who sings does he love me with reba", return_tensors="pt"
-        ).input_ids
-
-        input_ids = input_ids.to(torch_device)
-
-        output_ids = rag_sequence.generate(
-            input_ids,
-        )
-        # sequence generate test
-        output_text = rag_decoder_tokenizer.decode(output_ids[0], skip_special_tokens=True)
-
-        EXPECTED_OUTPUT_TEXT = """The album showed a songwriting maturity and depth of feeling distinctly lacking from their earlier recordings. The album\'s title track refers to secret meetings held against the approval of totalitarian governments in Soviet-dominated states. The only major single release, "One of Us", proved to be the last of ABBA\'s nine number-one singles in Germany."""
-        self.assertEqual(output_text, EXPECTED_OUTPUT_TEXT)
-
-    @slow
-    def test_rag_token_generate(self):
-        rag_config = self.get_rag_config()
-        rag_decoder_tokenizer = BartTokenizer.from_pretrained("facebook/bart-large-cnn")
-        rag_question_encoder_tokenizer = DPRQuestionEncoderTokenizer.from_pretrained(
-            "facebook/dpr-question_encoder-single-nq-base"
-        )
-        rag_retriever = RagRetriever(
-            rag_config,
-            question_encoder_tokenizer=rag_question_encoder_tokenizer,
-            generator_tokenizer=rag_decoder_tokenizer,
-        )
-
-        rag_token = self.token_model
-        rag_token.set_retriever(rag_retriever)
-
-        input_ids = rag_question_encoder_tokenizer(
-            "who sings does he love me with reba", return_tensors="pt"
-        ).input_ids
-
-        input_ids = input_ids.to(torch_device)
-
-        output_ids = rag_token.generate(
-            input_ids, decoder_start_token_id=rag_token.generator.config.decoder_start_token_id
-        )
-        # sequence generate test
-        output_text = rag_decoder_tokenizer.decode(output_ids[0], skip_special_tokens=True)
-        EXPECTED_OUTPUT_TEXT = """. The song peaked at"""
-        self.assertEqual(output_text, EXPECTED_OUTPUT_TEXT)
-
-
-@require_torch
-@require_retrieval
-class RagModelSaveLoadTests(unittest.TestCase):
-    def get_rag_config(self):
-        question_encoder_config = AutoConfig.from_pretrained("facebook/dpr-question_encoder-single-nq-base")
-        generator_config = AutoConfig.from_pretrained("facebook/bart-large-cnn")
-        return RagConfig.from_question_encoder_generator_configs(
-            question_encoder_config,
-            generator_config,
-            bos_token_id=0,
-            decoder_start_token_id=2,
-            eos_token_id=2,
-            is_encoder_decoder=True,
-            pad_token_id=1,
-            vocab_size=50264,
-            title_sep=" / ",
-            doc_sep=" // ",
-            n_docs=5,
-            max_combined_length=300,
-            dataset="wiki_dpr",
-            dataset_split="train",
-            index_name="exact",
-            index_path=None,
-            use_dummy_dataset=True,
-            retrieval_vector_size=768,
-            retrieval_batch_size=8,
-        )
-
-    @slow
-    def test_rag_sequence_from_pretrained(self):
-        rag_config = self.get_rag_config()
-        rag_decoder_tokenizer = BartTokenizer.from_pretrained("facebook/bart-large-cnn")
-        rag_question_encoder_tokenizer = DPRQuestionEncoderTokenizer.from_pretrained(
-            "facebook/dpr-question_encoder-single-nq-base"
-        )
-        rag_retriever = RagRetriever(
-            rag_config,
-            question_encoder_tokenizer=rag_question_encoder_tokenizer,
-            generator_tokenizer=rag_decoder_tokenizer,
-        )
-
-        input_ids = rag_question_encoder_tokenizer(
-            "who sings does he love me with reba", return_tensors="pt"
-        ).input_ids
-        decoder_input_ids = rag_decoder_tokenizer("Linda Davis", return_tensors="pt").input_ids
-
-        input_ids = input_ids.to(torch_device)
-        decoder_input_ids = decoder_input_ids.to(torch_device)
-
-        with tempfile.TemporaryDirectory() as tmp_dirname:
-            rag_sequence = RagSequenceForGeneration.from_pretrained_question_encoder_generator(
-                "facebook/dpr-question_encoder-single-nq-base",
-                "facebook/bart-large-cnn",
-                retriever=rag_retriever,
-                config=rag_config,
-            )
-            # check that the from pretrained methods work
-            rag_sequence.save_pretrained(tmp_dirname)
-            rag_sequence.from_pretrained(tmp_dirname, retriever=rag_retriever)
-
-            with torch.no_grad():
-                output = rag_sequence(
-                    input_ids,
-                    labels=decoder_input_ids,
-                )
-
-            loss_pretrained = output.loss
-            del rag_sequence
-
-        question_encoder = AutoModel.from_pretrained("facebook/dpr-question_encoder-single-nq-base")
-        generator = AutoModelForSeq2SeqLM.from_pretrained("facebook/bart-large-cnn")
-        rag_sequence = RagSequenceForGeneration(
-            config=rag_config, question_encoder=question_encoder, generator=generator, retriever=rag_retriever
-        )
-
-        with torch.no_grad():
-            output = rag_sequence(
-                input_ids,
-                labels=decoder_input_ids,
-            )
-
-        loss_init = output.loss
-
-        self.assertAlmostEqual(loss_pretrained.item(), loss_init.item(), places=4)
-
-    @slow
-    def test_rag_token_from_pretrained(self):
-        rag_config = self.get_rag_config()
-        rag_decoder_tokenizer = BartTokenizer.from_pretrained("facebook/bart-large-cnn")
-        rag_question_encoder_tokenizer = DPRQuestionEncoderTokenizer.from_pretrained(
-            "facebook/dpr-question_encoder-single-nq-base"
-        )
-        rag_retriever = RagRetriever(
-            rag_config,
-            question_encoder_tokenizer=rag_question_encoder_tokenizer,
-            generator_tokenizer=rag_decoder_tokenizer,
-        )
-
-        input_ids = rag_question_encoder_tokenizer(
-            "who sings does he love me with reba", return_tensors="pt"
-        ).input_ids
-        decoder_input_ids = rag_decoder_tokenizer("Linda Davis", return_tensors="pt").input_ids
-
-        input_ids = input_ids.to(torch_device)
-        decoder_input_ids = decoder_input_ids.to(torch_device)
-
-        with tempfile.TemporaryDirectory() as tmp_dirname:
-            rag_token = RagTokenForGeneration.from_pretrained_question_encoder_generator(
-                "facebook/dpr-question_encoder-single-nq-base",
-                "facebook/bart-large-cnn",
-                retriever=rag_retriever,
-                config=rag_config,
-            )
-            # check that the from pretrained methods work
-            rag_token.save_pretrained(tmp_dirname)
-            rag_token.from_pretrained(tmp_dirname, retriever=rag_retriever)
-
-            with torch.no_grad():
-                output = rag_token(
-                    input_ids,
-                    labels=decoder_input_ids,
-                )
-
-            loss_pretrained = output.loss
-            del rag_token
-
-        question_encoder = AutoModel.from_pretrained("facebook/dpr-question_encoder-single-nq-base")
-        generator = AutoModelForSeq2SeqLM.from_pretrained("facebook/bart-large-cnn")
-        rag_token = RagTokenForGeneration(
-            config=rag_config, question_encoder=question_encoder, generator=generator, retriever=rag_retriever
-        )
-
-        with torch.no_grad():
-            output = rag_token(
-                input_ids,
-                labels=decoder_input_ids,
-            )
-
-        loss_init = output.loss
-
-        self.assertAlmostEqual(loss_pretrained.item(), loss_init.item(), places=4)

tests/test_retrieval_rag.py

@@ -1,264 +0,0 @@
-import json
-import os
-import pickle
-import shutil
-import tempfile
-from unittest import TestCase
-from unittest.mock import patch
-
-import faiss
-import numpy as np
-from datasets import Dataset
-
-from transformers.configuration_bart import BartConfig
-from transformers.configuration_dpr import DPRConfig
-from transformers.configuration_rag import RagConfig
-from transformers.retrieval_rag import RagPyTorchDistributedRetriever, RagRetriever
-from transformers.testing_utils import require_datasets, require_faiss, require_torch
-from transformers.tokenization_bart import BartTokenizer
-from transformers.tokenization_bert import VOCAB_FILES_NAMES as DPR_VOCAB_FILES_NAMES
-from transformers.tokenization_dpr import DPRQuestionEncoderTokenizer
-from transformers.tokenization_roberta import VOCAB_FILES_NAMES as BART_VOCAB_FILES_NAMES
-
-
-@require_faiss
-@require_datasets
-@require_torch
-class RagRetrieverTest(TestCase):
-    def setUp(self):
-        self.tmpdirname = tempfile.mkdtemp()
-        self.retrieval_vector_size = 8
-
-        # DPR tok
-        vocab_tokens = [
-            "[UNK]",
-            "[CLS]",
-            "[SEP]",
-            "[PAD]",
-            "[MASK]",
-            "want",
-            "##want",
-            "##ed",
-            "wa",
-            "un",
-            "runn",
-            "##ing",
-            ",",
-            "low",
-            "lowest",
-        ]
-        dpr_tokenizer_path = os.path.join(self.tmpdirname, "dpr_tokenizer")
-        os.makedirs(dpr_tokenizer_path, exist_ok=True)
-        self.vocab_file = os.path.join(dpr_tokenizer_path, DPR_VOCAB_FILES_NAMES["vocab_file"])
-        with open(self.vocab_file, "w", encoding="utf-8") as vocab_writer:
-            vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
-
-        # BART tok
-        vocab = [
-            "l",
-            "o",
-            "w",
-            "e",
-            "r",
-            "s",
-            "t",
-            "i",
-            "d",
-            "n",
-            "\u0120",
-            "\u0120l",
-            "\u0120n",
-            "\u0120lo",
-            "\u0120low",
-            "er",
-            "\u0120lowest",
-            "\u0120newer",
-            "\u0120wider",
-            "<unk>",
-        ]
-        vocab_tokens = dict(zip(vocab, range(len(vocab))))
-        merges = ["#version: 0.2", "\u0120 l", "\u0120l o", "\u0120lo w", "e r", ""]
-        self.special_tokens_map = {"unk_token": "<unk>"}
-
-        bart_tokenizer_path = os.path.join(self.tmpdirname, "bart_tokenizer")
-        os.makedirs(bart_tokenizer_path, exist_ok=True)
-        self.vocab_file = os.path.join(bart_tokenizer_path, BART_VOCAB_FILES_NAMES["vocab_file"])
-        self.merges_file = os.path.join(bart_tokenizer_path, BART_VOCAB_FILES_NAMES["merges_file"])
-        with open(self.vocab_file, "w", encoding="utf-8") as fp:
-            fp.write(json.dumps(vocab_tokens) + "\n")
-        with open(self.merges_file, "w", encoding="utf-8") as fp:
-            fp.write("\n".join(merges))
-
-    def get_dpr_tokenizer(self) -> DPRQuestionEncoderTokenizer:
-        return DPRQuestionEncoderTokenizer.from_pretrained(os.path.join(self.tmpdirname, "dpr_tokenizer"))
-
-    def get_bart_tokenizer(self) -> BartTokenizer:
-        return BartTokenizer.from_pretrained(os.path.join(self.tmpdirname, "bart_tokenizer"))
-
-    def tearDown(self):
-        shutil.rmtree(self.tmpdirname)
-
-    def get_dummy_hf_index_retriever(self) -> RagRetriever:
-        dataset = Dataset.from_dict(
-            {
-                "id": ["0", "1"],
-                "text": ["foo", "bar"],
-                "title": ["Foo", "Bar"],
-                "embeddings": [np.ones(self.retrieval_vector_size), 2 * np.ones(self.retrieval_vector_size)],
-            }
-        )
-        dataset.add_faiss_index("embeddings", string_factory="Flat", metric_type=faiss.METRIC_INNER_PRODUCT)
-        config = RagConfig(
-            retrieval_vector_size=self.retrieval_vector_size,
-            question_encoder=DPRConfig().to_dict(),
-            generator=BartConfig().to_dict(),
-        )
-        with patch("transformers.retrieval_rag.load_dataset") as mock_load_dataset:
-            mock_load_dataset.return_value = dataset
-            retriever = RagRetriever(
-                config,
-                question_encoder_tokenizer=self.get_dpr_tokenizer(),
-                generator_tokenizer=self.get_bart_tokenizer(),
-            )
-        return retriever
-
-    def get_dummy_pytorch_distributed_retriever(self, init_retrieval, port=12345) -> RagRetriever:
-        dataset = Dataset.from_dict(
-            {
-                "id": ["0", "1"],
-                "text": ["foo", "bar"],
-                "title": ["Foo", "Bar"],
-                "embeddings": [np.ones(self.retrieval_vector_size), 2 * np.ones(self.retrieval_vector_size)],
-            }
-        )
-        dataset.add_faiss_index("embeddings", string_factory="Flat", metric_type=faiss.METRIC_INNER_PRODUCT)
-        config = RagConfig(
-            retrieval_vector_size=self.retrieval_vector_size,
-            question_encoder=DPRConfig().to_dict(),
-            generator=BartConfig().to_dict(),
-        )
-        with patch("transformers.retrieval_rag.load_dataset") as mock_load_dataset:
-            mock_load_dataset.return_value = dataset
-            retriever = RagPyTorchDistributedRetriever(
-                config,
-                question_encoder_tokenizer=self.get_dpr_tokenizer(),
-                generator_tokenizer=self.get_bart_tokenizer(),
-            )
-            if init_retrieval:
-                retriever.init_retrieval(port)
-        return retriever
-
-    def get_dummy_legacy_index_retriever(self) -> RagRetriever:
-        dataset = Dataset.from_dict(
-            {
-                "id": ["0", "1"],
-                "text": ["foo", "bar"],
-                "title": ["Foo", "Bar"],
-                "embeddings": [np.ones(self.retrieval_vector_size + 1), 2 * np.ones(self.retrieval_vector_size + 1)],
-            }
-        )
-        dataset.add_faiss_index("embeddings", string_factory="Flat", metric_type=faiss.METRIC_INNER_PRODUCT)
-
-        index_file_name = os.path.join(self.tmpdirname, "hf_bert_base.hnswSQ8_correct_phi_128.c_index")
-        dataset.save_faiss_index("embeddings", index_file_name + ".index.dpr")
-        pickle.dump(dataset["id"], open(index_file_name + ".index_meta.dpr", "wb"))
-
-        passages_file_name = os.path.join(self.tmpdirname, "psgs_w100.tsv.pkl")
-        passages = {sample["id"]: [sample["text"], sample["title"]] for sample in dataset}
-        pickle.dump(passages, open(passages_file_name, "wb"))
-
-        config = RagConfig(
-            retrieval_vector_size=self.retrieval_vector_size,
-            question_encoder=DPRConfig().to_dict(),
-            generator=BartConfig().to_dict(),
-            index_name="legacy",
-            index_path=self.tmpdirname,
-            passages_path=self.tmpdirname,
-        )
-        retriever = RagRetriever(
-            config, question_encoder_tokenizer=self.get_dpr_tokenizer(), generator_tokenizer=self.get_bart_tokenizer()
-        )
-        return retriever
-
-    def test_hf_index_retriever_retrieve(self):
-        n_docs = 1
-        retriever = self.get_dummy_hf_index_retriever()
-        hidden_states = np.array(
-            [np.ones(self.retrieval_vector_size), -np.ones(self.retrieval_vector_size)], dtype=np.float32
-        )
-        retrieved_doc_embeds, doc_ids, doc_dicts = retriever.retrieve(hidden_states, n_docs=n_docs)
-        self.assertEqual(retrieved_doc_embeds.shape, (2, n_docs, self.retrieval_vector_size))
-        self.assertEqual(len(doc_dicts), 2)
-        self.assertEqual(sorted(doc_dicts[0]), ["embeddings", "id", "text", "title"])
-        self.assertEqual(len(doc_dicts[0]["id"]), n_docs)
-        self.assertEqual(doc_dicts[0]["id"][0], "1")  # max inner product is reached with second doc
-        self.assertEqual(doc_dicts[1]["id"][0], "0")  # max inner product is reached with first doc
-        self.assertListEqual(list(doc_ids), [1, 0])
-
-    def test_pytorch_distributed_retriever_retrieve(self):
-        n_docs = 1
-        retriever = self.get_dummy_pytorch_distributed_retriever(init_retrieval=True)
-        hidden_states = np.array(
-            [np.ones(self.retrieval_vector_size), -np.ones(self.retrieval_vector_size)], dtype=np.float32
-        )
-        retrieved_doc_embeds, doc_ids, doc_dicts = retriever.retrieve(hidden_states, n_docs=n_docs)
-        self.assertEqual(retrieved_doc_embeds.shape, (2, n_docs, self.retrieval_vector_size))
-        self.assertEqual(len(doc_dicts), 2)
-        self.assertEqual(sorted(doc_dicts[0]), ["embeddings", "id", "text", "title"])
-        self.assertEqual(len(doc_dicts[0]["id"]), n_docs)
-        self.assertEqual(doc_dicts[0]["id"][0], "1")  # max inner product is reached with second doc
-        self.assertEqual(doc_dicts[1]["id"][0], "0")  # max inner product is reached with first doc
-        self.assertListEqual(list(doc_ids), [1, 0])
-
-    def test_legacy_index_retriever_retrieve(self):
-        n_docs = 1
-        retriever = self.get_dummy_legacy_index_retriever()
-        hidden_states = np.array(
-            [np.ones(self.retrieval_vector_size), -np.ones(self.retrieval_vector_size)], dtype=np.float32
-        )
-        retrieved_doc_embeds, doc_ids, doc_dicts = retriever.retrieve(hidden_states, n_docs=n_docs)
-        self.assertEqual(retrieved_doc_embeds.shape, (2, n_docs, self.retrieval_vector_size))
-        self.assertEqual(len(doc_dicts), 2)
-        self.assertEqual(sorted(doc_dicts[0]), ["text", "title"])
-        self.assertEqual(len(doc_dicts[0]["text"]), n_docs)
-        self.assertEqual(doc_dicts[0]["text"][0], "bar")  # max inner product is reached with second doc
-        self.assertEqual(doc_dicts[1]["text"][0], "foo")  # max inner product is reached with first doc
-        self.assertListEqual(list(doc_ids), [1, 0])
-
-    def test_hf_index_retriever_call(self):
-        import torch
-
-        n_docs = 1
-        retriever = self.get_dummy_hf_index_retriever()
-        question_input_ids = [[5, 7], [10, 11]]
-        hidden_states = np.array(
-            [np.ones(self.retrieval_vector_size), -np.ones(self.retrieval_vector_size)], dtype=np.float32
-        )
-        out = retriever(question_input_ids, hidden_states, prefix=retriever.config.generator.prefix, n_docs=n_docs)
-        context_input_ids, context_attention_mask, retrieved_doc_embeds = (
-            out["context_input_ids"],
-            out["context_attention_mask"],
-            out["retrieved_doc_embeds"],
-        )
-        self.assertEqual(retrieved_doc_embeds.shape, (2, n_docs, self.retrieval_vector_size))
-        self.assertIsInstance(context_input_ids, list)
-        self.assertIsInstance(context_attention_mask, list)
-        self.assertIsInstance(retrieved_doc_embeds, np.ndarray)
-
-        out = retriever(
-            question_input_ids,
-            hidden_states,
-            prefix=retriever.config.generator.prefix,
-            n_docs=n_docs,
-            return_tensors="pt",
-        )
-        context_input_ids, context_attention_mask, retrieved_doc_embeds, doc_ids = (  # noqa: F841
-            out["context_input_ids"],
-            out["context_attention_mask"],
-            out["retrieved_doc_embeds"],
-            out["doc_ids"],
-        )
-        self.assertEqual(retrieved_doc_embeds.shape, (2, n_docs, self.retrieval_vector_size))
-        self.assertIsInstance(context_input_ids, torch.Tensor)
-        self.assertIsInstance(context_attention_mask, torch.Tensor)
-        self.assertIsInstance(retrieved_doc_embeds, torch.Tensor)