0 Fork 0

Files

T

thomwolf 886cb49792 updating readme and notebooks

2018-11-16 14:31:15 +01:00

61 KiB

Raw Blame History

Comparing TensorFlow (original) and PyTorch models

You can use this small notebook to check the conversion of the model's weights from the TensorFlow model to the PyTorch model. In the following, we compare the weights of the last layer on a simple example (in input.txt) but both models returns all the hidden layers so you can check every stage of the model.

To run this notebook, follow these instructions:

make sure that your Python environment has both TensorFlow and PyTorch installed,
download the original TensorFlow implementation,
download a pre-trained TensorFlow model as indicaded in the TensorFlow implementation readme,
run the script convert_tf_checkpoint_to_pytorch.py as indicated in the README to convert the pre-trained TensorFlow model to PyTorch.

If needed change the relative paths indicated in this notebook (at the beggining of Sections 1 and 2) to point to the relevent models and code.

In [1]:

import os
os.chdir('../')

1/ TensorFlow code

In [2]:

original_tf_inplem_dir = "./tensorflow_code/"
model_dir = "../google_models/uncased_L-12_H-768_A-12/"

vocab_file = model_dir + "vocab.txt"
bert_config_file = model_dir + "bert_config.json"
init_checkpoint = model_dir + "bert_model.ckpt"

input_file = "./samples/input.txt"
max_seq_length = 128

In [6]:

import importlib.util
import sys

spec = importlib.util.spec_from_file_location('*', original_tf_inplem_dir + '/extract_features_tensorflow.py')
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
sys.modules['extract_features_tensorflow'] = module

from extract_features_tensorflow import *

[0;31m---------------------------------------------------------------------------[0m
[0;31mDuplicateFlagError[0m                        Traceback (most recent call last)
[0;32m<ipython-input-6-86ecffb49060>[0m in [0;36m<module>[0;34m[0m
[1;32m      4[0m [0mspec[0m [0;34m=[0m [0mimportlib[0m[0;34m.[0m[0mutil[0m[0;34m.[0m[0mspec_from_file_location[0m[0;34m([0m[0;34m'*'[0m[0;34m,[0m [0moriginal_tf_inplem_dir[0m [0;34m+[0m [0;34m'/extract_features_tensorflow.py'[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[1;32m      5[0m [0mmodule[0m [0;34m=[0m [0mimportlib[0m[0;34m.[0m[0mutil[0m[0;34m.[0m[0mmodule_from_spec[0m[0;34m([0m[0mspec[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0;32m----> 6[0;31m [0mspec[0m[0;34m.[0m[0mloader[0m[0;34m.[0m[0mexec_module[0m[0;34m([0m[0mmodule[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m      7[0m [0msys[0m[0;34m.[0m[0mmodules[0m[0;34m[[0m[0;34m'extract_features_tensorflow'[0m[0;34m][0m [0;34m=[0m [0mmodule[0m[0;34m[0m[0;34m[0m[0m
[1;32m      8[0m [0;34m[0m[0m

[0;32m~/miniconda3/envs/bert/lib/python3.6/importlib/_bootstrap_external.py[0m in [0;36mexec_module[0;34m(self, module)[0m

[0;32m~/miniconda3/envs/bert/lib/python3.6/importlib/_bootstrap.py[0m in [0;36m_call_with_frames_removed[0;34m(f, *args, **kwds)[0m

[0;32m~/Documents/Thomas/Code/HF/BERT/pytorch-pretrained-BERT/tensorflow_code/extract_features_tensorflow.py[0m in [0;36m<module>[0;34m[0m
[1;32m     32[0m [0mFLAGS[0m [0;34m=[0m [0mflags[0m[0;34m.[0m[0mFLAGS[0m[0;34m[0m[0;34m[0m[0m
[1;32m     33[0m [0;34m[0m[0m
[0;32m---> 34[0;31m [0mflags[0m[0;34m.[0m[0mDEFINE_string[0m[0;34m([0m[0;34m"input_file"[0m[0;34m,[0m [0;32mNone[0m[0;34m,[0m [0;34m""[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m     35[0m [0;34m[0m[0m
[1;32m     36[0m [0mflags[0m[0;34m.[0m[0mDEFINE_string[0m[0;34m([0m[0;34m"output_file"[0m[0;34m,[0m [0;32mNone[0m[0;34m,[0m [0;34m""[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m

[0;32m~/miniconda3/envs/bert/lib/python3.6/site-packages/tensorflow/python/platform/flags.py[0m in [0;36mwrapper[0;34m(*args, **kwargs)[0m
[1;32m     56[0m           [0;34m'Use of the keyword argument names (flag_name, default_value, '[0m[0;34m[0m[0;34m[0m[0m
[1;32m     57[0m           'docstring) is deprecated, please use (name, default, help) instead.')
[0;32m---> 58[0;31m     [0;32mreturn[0m [0moriginal_function[0m[0;34m([0m[0;34m*[0m[0margs[0m[0;34m,[0m [0;34m**[0m[0mkwargs[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m     59[0m [0;34m[0m[0m
[1;32m     60[0m   [0;32mreturn[0m [0mtf_decorator[0m[0;34m.[0m[0mmake_decorator[0m[0;34m([0m[0moriginal_function[0m[0;34m,[0m [0mwrapper[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m

[0;32m~/miniconda3/envs/bert/lib/python3.6/site-packages/absl/flags/_defines.py[0m in [0;36mDEFINE_string[0;34m(name, default, help, flag_values, **args)[0m
[1;32m    239[0m   [0mparser[0m [0;34m=[0m [0m_argument_parser[0m[0;34m.[0m[0mArgumentParser[0m[0;34m([0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[1;32m    240[0m   [0mserializer[0m [0;34m=[0m [0m_argument_parser[0m[0;34m.[0m[0mArgumentSerializer[0m[0;34m([0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0;32m--> 241[0;31m   [0mDEFINE[0m[0;34m([0m[0mparser[0m[0;34m,[0m [0mname[0m[0;34m,[0m [0mdefault[0m[0;34m,[0m [0mhelp[0m[0;34m,[0m [0mflag_values[0m[0;34m,[0m [0mserializer[0m[0;34m,[0m [0;34m**[0m[0margs[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m    242[0m [0;34m[0m[0m
[1;32m    243[0m [0;34m[0m[0m

[0;32m~/miniconda3/envs/bert/lib/python3.6/site-packages/absl/flags/_defines.py[0m in [0;36mDEFINE[0;34m(parser, name, default, help, flag_values, serializer, module_name, **args)[0m
[1;32m     80[0m   """
[1;32m     81[0m   DEFINE_flag(_flag.Flag(parser, serializer, name, default, help, **args),
[0;32m---> 82[0;31m               flag_values, module_name)
[0m[1;32m     83[0m [0;34m[0m[0m
[1;32m     84[0m [0;34m[0m[0m

[0;32m~/miniconda3/envs/bert/lib/python3.6/site-packages/absl/flags/_defines.py[0m in [0;36mDEFINE_flag[0;34m(flag, flag_values, module_name)[0m
[1;32m    102[0m   [0;31m# Copying the reference to flag_values prevents pychecker warnings.[0m[0;34m[0m[0;34m[0m[0;34m[0m[0m
[1;32m    103[0m   [0mfv[0m [0;34m=[0m [0mflag_values[0m[0;34m[0m[0;34m[0m[0m
[0;32m--> 104[0;31m   [0mfv[0m[0;34m[[0m[0mflag[0m[0;34m.[0m[0mname[0m[0;34m][0m [0;34m=[0m [0mflag[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m    105[0m   [0;31m# Tell flag_values who's defining the flag.[0m[0;34m[0m[0;34m[0m[0;34m[0m[0m
[1;32m    106[0m   [0;32mif[0m [0mmodule_name[0m[0;34m:[0m[0;34m[0m[0;34m[0m[0m

[0;32m~/miniconda3/envs/bert/lib/python3.6/site-packages/absl/flags/_flagvalues.py[0m in [0;36m__setitem__[0;34m(self, name, flag)[0m
[1;32m    427[0m         [0;31m# module is simply being imported a subsequent time.[0m[0;34m[0m[0;34m[0m[0;34m[0m[0m
[1;32m    428[0m         [0;32mreturn[0m[0;34m[0m[0;34m[0m[0m
[0;32m--> 429[0;31m       [0;32mraise[0m [0m_exceptions[0m[0;34m.[0m[0mDuplicateFlagError[0m[0;34m.[0m[0mfrom_flag[0m[0;34m([0m[0mname[0m[0;34m,[0m [0mself[0m[0;34m)[0m[0;34m[0m[0;34m[0m[0m
[0m[1;32m    430[0m     [0mshort_name[0m [0;34m=[0m [0mflag[0m[0;34m.[0m[0mshort_name[0m[0;34m[0m[0;34m[0m[0m
[1;32m    431[0m     [0;31m# If a new flag overrides an old one, we need to cleanup the old flag's[0m[0;34m[0m[0;34m[0m[0;34m[0m[0m

[0;31mDuplicateFlagError[0m: The flag 'input_file' is defined twice. First from *, Second from *.  Description from first occurrence: (no help available)

In [8]:

layer_indexes = list(range(12))
bert_config = modeling.BertConfig.from_json_file(bert_config_file)
tokenizer = tokenization.FullTokenizer(
    vocab_file=vocab_file, do_lower_case=True)
examples = read_examples(input_file)

features = convert_examples_to_features(
    examples=examples, seq_length=max_seq_length, tokenizer=tokenizer)
unique_id_to_feature = {}
for feature in features:
    unique_id_to_feature[feature.unique_id] = feature

INFO:tensorflow:*** Example ***
INFO:tensorflow:unique_id: 0
INFO:tensorflow:tokens: [CLS] who was jim henson ? [SEP] jim henson was a puppet ##eer [SEP]
INFO:tensorflow:input_ids: 101 2040 2001 3958 27227 1029 102 3958 27227 2001 1037 13997 11510 102 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
INFO:tensorflow:input_mask: 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
INFO:tensorflow:input_type_ids: 0 0 0 0 0 0 0 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

In [9]:

is_per_host = tf.contrib.tpu.InputPipelineConfig.PER_HOST_V2
run_config = tf.contrib.tpu.RunConfig(
    master=None,
    tpu_config=tf.contrib.tpu.TPUConfig(
        num_shards=1,
        per_host_input_for_training=is_per_host))

model_fn = model_fn_builder(
    bert_config=bert_config,
    init_checkpoint=init_checkpoint,
    layer_indexes=layer_indexes,
    use_tpu=False,
    use_one_hot_embeddings=False)

# If TPU is not available, this will fall back to normal Estimator on CPU
# or GPU.
estimator = tf.contrib.tpu.TPUEstimator(
    use_tpu=False,
    model_fn=model_fn,
    config=run_config,
    predict_batch_size=1)

input_fn = input_fn_builder(
    features=features, seq_length=max_seq_length)

WARNING:tensorflow:Estimator's model_fn (<function model_fn_builder.<locals>.model_fn at 0x11ea7f1e0>) includes params argument, but params are not passed to Estimator.
WARNING:tensorflow:Using temporary folder as model directory: /var/folders/yx/cw8n_njx3js5jksyw_qlp8p00000gn/T/tmphs4_nsq9
INFO:tensorflow:Using config: {'_model_dir': '/var/folders/yx/cw8n_njx3js5jksyw_qlp8p00000gn/T/tmphs4_nsq9', '_tf_random_seed': None, '_save_summary_steps': 100, '_save_checkpoints_steps': None, '_save_checkpoints_secs': 600, '_session_config': allow_soft_placement: true
graph_options {
  rewrite_options {
    meta_optimizer_iterations: ONE
  }
}
, '_keep_checkpoint_max': 5, '_keep_checkpoint_every_n_hours': 10000, '_log_step_count_steps': None, '_train_distribute': None, '_device_fn': None, '_protocol': None, '_eval_distribute': None, '_experimental_distribute': None, '_service': None, '_cluster_spec': <tensorflow.python.training.server_lib.ClusterSpec object at 0x121b163c8>, '_task_type': 'worker', '_task_id': 0, '_global_id_in_cluster': 0, '_master': '', '_evaluation_master': '', '_is_chief': True, '_num_ps_replicas': 0, '_num_worker_replicas': 1, '_tpu_config': TPUConfig(iterations_per_loop=2, num_shards=1, num_cores_per_replica=None, per_host_input_for_training=3, tpu_job_name=None, initial_infeed_sleep_secs=None, input_partition_dims=None), '_cluster': None}
WARNING:tensorflow:Setting TPUConfig.num_shards==1 is an unsupported behavior. Please fix as soon as possible (leaving num_shards as None.
INFO:tensorflow:_TPUContext: eval_on_tpu True
WARNING:tensorflow:eval_on_tpu ignored because use_tpu is False.

In [10]:

tensorflow_all_out = []
for result in estimator.predict(input_fn, yield_single_examples=True):
    unique_id = int(result["unique_id"])
    feature = unique_id_to_feature[unique_id]
    output_json = collections.OrderedDict()
    output_json["linex_index"] = unique_id
    tensorflow_all_out_features = []
    # for (i, token) in enumerate(feature.tokens):
    all_layers = []
    for (j, layer_index) in enumerate(layer_indexes):
        print("extracting layer {}".format(j))
        layer_output = result["layer_output_%d" % j]
        layers = collections.OrderedDict()
        layers["index"] = layer_index
        layers["values"] = layer_output
        all_layers.append(layers)
    tensorflow_out_features = collections.OrderedDict()
    tensorflow_out_features["layers"] = all_layers
    tensorflow_all_out_features.append(tensorflow_out_features)

    output_json["features"] = tensorflow_all_out_features
    tensorflow_all_out.append(output_json)

INFO:tensorflow:Could not find trained model in model_dir: /var/folders/yx/cw8n_njx3js5jksyw_qlp8p00000gn/T/tmphs4_nsq9, running initialization to predict.
INFO:tensorflow:Calling model_fn.
INFO:tensorflow:Running infer on CPU
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Graph was finalized.
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
extracting layer 0
extracting layer 1
extracting layer 2
extracting layer 3
extracting layer 4
extracting layer 5
extracting layer 6
extracting layer 7
extracting layer 8
extracting layer 9
extracting layer 10
extracting layer 11
INFO:tensorflow:prediction_loop marked as finished
INFO:tensorflow:prediction_loop marked as finished

In [11]:

print(len(tensorflow_all_out))
print(len(tensorflow_all_out[0]))
print(tensorflow_all_out[0].keys())
print("number of tokens", len(tensorflow_all_out[0]['features']))
print("number of layers", len(tensorflow_all_out[0]['features'][0]['layers']))
tensorflow_all_out[0]['features'][0]['layers'][0]['values'].shape

Out [11]:

1
2
odict_keys(['linex_index', 'features'])
number of tokens 1
number of layers 12

(128, 768)

In [12]:

tensorflow_outputs = list(tensorflow_all_out[0]['features'][0]['layers'][t]['values'] for t in layer_indexes)

2/ PyTorch code

In [ ]:

os.chdir('./examples')

In [17]:

import extract_features
import pytorch_pretrained_bert as ppb
from extract_features import *

In [25]:

init_checkpoint_pt = "../../google_models/uncased_L-12_H-768_A-12/"

In [26]:

device = torch.device("cpu")
model = ppb.BertModel.from_pretrained(init_checkpoint_pt)
model.to(device)

Out [26]:

11/15/2018 16:21:18 - INFO - pytorch_pretrained_bert.modeling -   loading archive file ../../google_models/uncased_L-12_H-768_A-12/
11/15/2018 16:21:18 - INFO - pytorch_pretrained_bert.modeling -   Model config {
  "attention_probs_dropout_prob": 0.1,
  "hidden_act": "gelu",
  "hidden_dropout_prob": 0.1,
  "hidden_size": 768,
  "initializer_range": 0.02,
  "intermediate_size": 3072,
  "max_position_embeddings": 512,
  "num_attention_heads": 12,
  "num_hidden_layers": 12,
  "type_vocab_size": 2,
  "vocab_size": 30522
}

BertModel(
  (embeddings): BertEmbeddings(
    (word_embeddings): Embedding(30522, 768)
    (position_embeddings): Embedding(512, 768)
    (token_type_embeddings): Embedding(2, 768)
    (LayerNorm): BertLayerNorm()
    (dropout): Dropout(p=0.1)
  )
  (encoder): BertEncoder(
    (layer): ModuleList(
      (0): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (1): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (2): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (3): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (4): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (5): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (6): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (7): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (8): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (9): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (10): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (11): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
    )
  )
  (pooler): BertPooler(
    (dense): Linear(in_features=768, out_features=768, bias=True)
    (activation): Tanh()
  )
)

In [27]:

all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
all_input_type_ids = torch.tensor([f.input_type_ids for f in features], dtype=torch.long)
all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)

eval_data = TensorDataset(all_input_ids, all_input_mask, all_input_type_ids, all_example_index)
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=1)

model.eval()

Out [27]:

BertModel(
  (embeddings): BertEmbeddings(
    (word_embeddings): Embedding(30522, 768)
    (position_embeddings): Embedding(512, 768)
    (token_type_embeddings): Embedding(2, 768)
    (LayerNorm): BertLayerNorm()
    (dropout): Dropout(p=0.1)
  )
  (encoder): BertEncoder(
    (layer): ModuleList(
      (0): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (1): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (2): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (3): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (4): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (5): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (6): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (7): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (8): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (9): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (10): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
      (11): BertLayer(
        (attention): BertAttention(
          (self): BertSelfAttention(
            (query): Linear(in_features=768, out_features=768, bias=True)
            (key): Linear(in_features=768, out_features=768, bias=True)
            (value): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1)
          )
          (output): BertSelfOutput(
            (dense): Linear(in_features=768, out_features=768, bias=True)
            (LayerNorm): BertLayerNorm()
            (dropout): Dropout(p=0.1)
          )
        )
        (intermediate): BertIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
        )
        (output): BertOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): BertLayerNorm()
          (dropout): Dropout(p=0.1)
        )
      )
    )
  )
  (pooler): BertPooler(
    (dense): Linear(in_features=768, out_features=768, bias=True)
    (activation): Tanh()
  )
)

In [28]:

layer_indexes = list(range(12))

pytorch_all_out = []
for input_ids, input_mask, input_type_ids, example_indices in eval_dataloader:
    print(input_ids)
    print(input_mask)
    print(example_indices)
    input_ids = input_ids.to(device)
    input_mask = input_mask.to(device)

    all_encoder_layers, _ = model(input_ids, token_type_ids=input_type_ids, attention_mask=input_mask)

    for b, example_index in enumerate(example_indices):
        feature = features[example_index.item()]
        unique_id = int(feature.unique_id)
        # feature = unique_id_to_feature[unique_id]
        output_json = collections.OrderedDict()
        output_json["linex_index"] = unique_id
        all_out_features = []
        # for (i, token) in enumerate(feature.tokens):
        all_layers = []
        for (j, layer_index) in enumerate(layer_indexes):
            print("layer", j, layer_index)
            layer_output = all_encoder_layers[int(layer_index)].detach().cpu().numpy()
            layer_output = layer_output[b]
            layers = collections.OrderedDict()
            layers["index"] = layer_index
            layer_output = layer_output
            layers["values"] = layer_output if not isinstance(layer_output, (int, float)) else [layer_output]
            all_layers.append(layers)

            out_features = collections.OrderedDict()
            out_features["layers"] = all_layers
            all_out_features.append(out_features)
        output_json["features"] = all_out_features
        pytorch_all_out.append(output_json)

tensor([[  101,  2040,  2001,  3958, 27227,  1029,   102,  3958, 27227,  2001,
          1037, 13997, 11510,   102,     0,     0,     0,     0,     0,     0,
             0,     0,     0,     0,     0,     0,     0,     0,     0,     0,
             0,     0,     0,     0,     0,     0,     0,     0,     0,     0,
             0,     0,     0,     0,     0,     0,     0,     0,     0,     0,
             0,     0,     0,     0,     0,     0,     0,     0,     0,     0,
             0,     0,     0,     0,     0,     0,     0,     0,     0,     0,
             0,     0,     0,     0,     0,     0,     0,     0,     0,     0,
             0,     0,     0,     0,     0,     0,     0,     0,     0,     0,
             0,     0,     0,     0,     0,     0,     0,     0,     0,     0,
             0,     0,     0,     0,     0,     0,     0,     0,     0,     0,
             0,     0,     0,     0,     0,     0,     0,     0,     0,     0,
             0,     0,     0,     0,     0,     0,     0,     0]])
tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
         0, 0, 0, 0, 0, 0, 0, 0]])
tensor([0])
layer 0 0
layer 1 1
layer 2 2
layer 3 3
layer 4 4
layer 5 5
layer 6 6
layer 7 7
layer 8 8
layer 9 9
layer 10 10
layer 11 11

In [29]:

print(len(pytorch_all_out))
print(len(pytorch_all_out[0]))
print(pytorch_all_out[0].keys())
print("number of tokens", len(pytorch_all_out))
print("number of layers", len(pytorch_all_out[0]['features'][0]['layers']))
print("hidden_size", len(pytorch_all_out[0]['features'][0]['layers'][0]['values']))
pytorch_all_out[0]['features'][0]['layers'][0]['values'].shape

Out [29]:

1
2
odict_keys(['linex_index', 'features'])
number of tokens 1
number of layers 12
hidden_size 128

(128, 768)

In [30]:

pytorch_outputs = list(pytorch_all_out[0]['features'][0]['layers'][t]['values'] for t in layer_indexes)
print(pytorch_outputs[0].shape)
print(pytorch_outputs[1].shape)

(128, 768)
(128, 768)

In [31]:

print(tensorflow_outputs[0].shape)
print(tensorflow_outputs[1].shape)

(128, 768)
(128, 768)

3/ Comparing the standard deviation on the last layer of both models

In [32]:

import numpy as np

In [33]:

print('shape tensorflow layer, shape pytorch layer, standard deviation')
print('\n'.join(list(str((np.array(tensorflow_outputs[i]).shape,
                          np.array(pytorch_outputs[i]).shape, 
                          np.sqrt(np.mean((np.array(tensorflow_outputs[i]) - np.array(pytorch_outputs[i]))**2.0)))) for i in range(12))))

shape tensorflow layer, shape pytorch layer, standard deviation
((128, 768), (128, 768), 1.5258875e-07)
((128, 768), (128, 768), 2.342731e-07)
((128, 768), (128, 768), 2.801949e-07)
((128, 768), (128, 768), 3.5904986e-07)
((128, 768), (128, 768), 4.2842768e-07)
((128, 768), (128, 768), 5.127951e-07)
((128, 768), (128, 768), 6.14668e-07)
((128, 768), (128, 768), 7.063922e-07)
((128, 768), (128, 768), 7.906173e-07)
((128, 768), (128, 768), 8.475192e-07)
((128, 768), (128, 768), 8.975489e-07)
((128, 768), (128, 768), 4.1671223e-07)