fix timing

src/transformers/benchmark/benchmark.py

@@ -157,7 +157,7 @@ class PyTorchBenchmark(Benchmark):
         else:
             train_model = model
 
-        model.eval()
+        model.train()
         model.to(self.args.device)
 
         # encoder-decoder has vocab size saved differently
@@ -175,12 +175,12 @@ class PyTorchBenchmark(Benchmark):
         def compute_loss_and_backprob_encoder():
             loss = train_model(input_ids, labels=input_ids)[0]
             loss.backward()
-            train_model.zero_grad()
+            return loss
 
         def compute_loss_and_backprob_encoder_decoder():
             loss = train_model(input_ids, decoder_input_ids=input_ids, labels=input_ids)[0]
             loss.backward()
-            train_model.zero_grad()
+            return loss
 
         _train = (
             compute_loss_and_backprob_encoder_decoder

src/transformers/benchmark/benchmark_tf.py

@@ -21,10 +21,17 @@
 import logging
 import random
 import timeit
+import time
 from functools import wraps
 from typing import Callable, Optional
 
-from transformers import TF_MODEL_MAPPING, PretrainedConfig, is_py3nvml_available, is_tf_available
+from transformers import (
+    TF_MODEL_MAPPING,
+    TF_MODEL_WITH_LM_HEAD_MAPPING,
+    PretrainedConfig,
+    is_py3nvml_available,
+    is_tf_available,
+)
 
 from .benchmark_utils import (
     Benchmark,
@@ -92,10 +99,11 @@ class TensorFlowBenchmark(Benchmark):
         _inference = self._prepare_inference_func(model_name, batch_size, sequence_length)
         return self._measure_speed(_inference)
 
-    def _train_speed(self, model_name, batch_size, sequence_length):
-        raise NotImplementedError(
-            "Training is currently not really implemented." "Wait for TFTrainer to support CLM and MLM."
-        )
+    def _train_speed(self, model_name: str, batch_size: int, sequence_length: int) -> float:
+        strategy = self.args.strategy
+        assert strategy is not None, "A device strategy has to be initialized before using TensorFlow."
+        _train = self._prepare_train_func(model_name, batch_size, sequence_length)
+        return self._measure_speed(_train)
 
     def _inference_memory(
         self, model_name: str, batch_size: int, sequence_length: int
@@ -108,10 +116,16 @@ class TensorFlowBenchmark(Benchmark):
         _inference = self._prepare_inference_func(model_name, batch_size, sequence_length)
         return self._measure_memory(_inference)
 
-    def _train_memory(self, model_name, batch_size, sequence_length):
-        raise NotImplementedError(
-            "Training is currently not really implemented. Wait for TFTrainer to support CLM and MLM."
-        )
+    def _train_memory(
+        self, model_name: str, batch_size: int, sequence_length: int
+    ) -> [Memory, Optional[MemorySummary]]:
+        if self.args.is_gpu:
+            tf.config.experimental.set_memory_growth(self.args.gpu_list[self.args.device_idx], True)
+        strategy = self.args.strategy
+        assert strategy is not None, "A device strategy has to be initialized before using TensorFlow."
+
+        _train = self._prepare_train_func(model_name, batch_size, sequence_length)
+        return self._measure_memory(_train)
 
     def _prepare_inference_func(self, model_name: str, batch_size: int, sequence_length: int) -> Callable[[], None]:
         config = self.config_dict[model_name]
@@ -149,16 +163,68 @@ class TensorFlowBenchmark(Benchmark):
 
         return _inference
 
+    def _prepare_train_func(self, model_name: str, batch_size: int, sequence_length: int) -> Callable[[], None]:
+        config = self.config_dict[model_name]
+
+        assert (
+            self.args.eager_mode is False
+        ), "Training cannot be done in eager mode. Please make sure that `args.eager_mode = False`."
+
+        if self.args.fp16:
+            raise NotImplementedError("Mixed precision is currently not supported.")
+
+        has_model_class_in_config = hasattr(config, "architecture") and len(config.architectures) > 1
+        if not self.args.only_pretrain_model and has_model_class_in_config:
+            try:
+                model_class = "TF" + config.architectures[0]  # prepend 'TF' for tensorflow model
+                transformers_module = __import__("transformers", fromlist=[model_class])
+                model_cls = getattr(transformers_module, model_class)
+                model = model_cls(config)
+            except ImportError:
+                raise ImportError(
+                    f"{model_class} does not exist. If you just want to test the pretrained model, you might want to set `--only_pretrain_model` or `args.only_pretrain_model=True`."
+                )
+        else:
+            model = TF_MODEL_WITH_LM_HEAD_MAPPING[config.__class__](config)
+
+        # encoder-decoder has vocab size saved differently
+        vocab_size = config.vocab_size if hasattr(config, "vocab_size") else config.encoder.vocab_size
+        input_ids = random_input_ids(batch_size, sequence_length, vocab_size)
+
+        @run_with_tf_optimizations(self.args.eager_mode, self.args.use_xla)
+        def encoder_decoder_train():
+            loss = model(input_ids, decoder_input_ids=input_ids, labels=input_ids, training=True)[0]
+            gradients = tf.gradients(loss, model.trainable_variables)
+            return gradients
+
+        @run_with_tf_optimizations(self.args.eager_mode, self.args.use_xla)
+        def encoder_train():
+            loss = model(input_ids, labels=input_ids, training=True)[0]
+            gradients = tf.gradients(loss, model.trainable_variables)
+            return gradients
+
+        _train = encoder_decoder_train if config.is_encoder_decoder else encoder_train
+
+        return _train
+
     def _measure_speed(self, func) -> float:
         with self.args.strategy.scope():
             try:
                 if self.args.is_tpu or self.args.use_xla:
                     # run additional 10 times to stabilize compilation for tpu
                     logger.info("Do inference on TPU. Running model 5 times to stabilize compilation")
+#                    grads = [func() for i in range(5)]
                     timeit.repeat(func, repeat=1, number=5)
 
                 # as written in https://docs.python.org/2/library/timeit.html#timeit.Timer.repeat, min should be taken rather than the average
                 runtimes = timeit.repeat(func, repeat=self.args.repeat, number=10,)
+#                start_time = time.time()
+#                grads = [func() for i in range(10)]
+#                end_time = time.time() - start_time
+#
+#                print("Time", end_time / 10)
+#                print("Grads", grads[0][0])
+#                return end_time / 10
 
                 return min(runtimes) / 10.0
             except ResourceExhaustedError as e: