SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Tensorflow improvements (#4530)

Browse Source 
* Better None gradients handling

* Apply Style

* Apply Style

* Create a loss class per task to compute its respective loss

* Add loss classes to the ALBERT TF models

* Add loss classes to the BERT TF models

* Add question answering and multiple choice to TF Camembert

* Remove prints

* Add multiple choice model to TF DistilBERT + loss computation

* Add question answering model to TF Electra + loss computation

* Add token classification, question answering and multiple choice models to TF Flaubert

* Add multiple choice model to TF Roberta + loss computation

* Add multiple choice model to TF XLM + loss computation

* Add multiple choice and question answering models to TF XLM-Roberta

* Add multiple choice model to TF XLNet + loss computation

* Remove unused parameters

* Add task loss classes

* Reorder TF imports + add new model classes

* Add new model classes

* Bugfix in TF T5 model

* Bugfix for TF T5 tests

* Bugfix in TF T5 model

* Fix TF T5 model tests

* Fix T5 tests + some renaming

* Fix inheritance issue in the AutoX tests

* Add tests for TF Flaubert and TF XLM Roberta

* Add tests for TF Flaubert and TF XLM Roberta

* Remove unused piece of code in the TF trainer

* bugfix and remove unused code

* Bugfix for TF 2.2

* Apply Style

* Divide TFSequenceClassificationAndMultipleChoiceLoss into their two respective name

* Apply style

* Mirror the PT Trainer in the TF one: fp16, optimizers and tb_writer as class parameter and better dataset handling

* Fix TF optimizations tests and apply style

* Remove useless parameter

* Bugfix and apply style

* Fix TF Trainer prediction

* Now the TF models return the loss such as their PyTorch couterparts

* Apply Style

* Ignore some tests output

* Take into account the SQuAD cls_index, p_mask and is_impossible parameters for the QuestionAnswering task models.

* Fix names for SQuAD data

* Apply Style

* Fix conflicts with 2.11 release

* Fix conflicts with 2.11

* Fix wrongname

* Add better documentation on the new create_optimizer function

* Fix isort

* logging_dir: use same default as PyTorch

Co-authored-by: Julien Chaumond <chaumond@gmail.com>
This commit is contained in:
Julien Plu
2020-06-05 01:45:53 +02:00
committed by GitHub
parent ccd26c2862
commit f9414f7553
27 changed files with 2380 additions and 558 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
tests/test_modeling_tf_common.py
View File

@@ -30,7 +30,7 @@ if is_tf_available():
import tensorflow as tf
import numpy as np
from transformers import tf_top_k_top_p_filtering, TFAdaptiveEmbedding
from transformers import tf_top_k_top_p_filtering, TFAdaptiveEmbedding, TFSharedEmbeddings
if _tf_gpu_memory_limit is not None:
gpus = tf.config.list_physical_devices("GPU")
@@ -107,26 +107,45 @@ class TFModelTesterMixin:
and getattr(module_member, "_keras_serializable", False)
)
for main_layer_class in tf_main_layer_classes:
main_layer = main_layer_class(config)
# T5MainLayer needs an embed_tokens parameter when called without the inputs_embeds parameter
if "T5" in main_layer_class.__name__:
# Take the same values than in TFT5ModelTester for this shared layer
shared = TFSharedEmbeddings(99, 32, name="shared")
main_layer = main_layer_class(config, embed_tokens=shared)
else:
main_layer = main_layer_class(config)
symbolic_inputs = {
name: tf.keras.Input(tensor.shape[1:], dtype=tensor.dtype) for name, tensor in inputs_dict.items()
}
model = tf.keras.Model(symbolic_inputs, outputs=main_layer(symbolic_inputs))
outputs = model(inputs_dict)
with tempfile.TemporaryDirectory() as tmpdirname:
filepath = os.path.join(tmpdirname, "keras_model.h5")
model.save(filepath)
model = tf.keras.models.load_model(
filepath, custom_objects={main_layer_class.__name__: main_layer_class}
)
if "T5" in main_layer_class.__name__:
model = tf.keras.models.load_model(
filepath,
custom_objects={
main_layer_class.__name__: main_layer_class,
"TFSharedEmbeddings": TFSharedEmbeddings,
},
)
else:
model = tf.keras.models.load_model(
filepath, custom_objects={main_layer_class.__name__: main_layer_class}
)
assert isinstance(model, tf.keras.Model)
after_outputs = model(inputs_dict)
self.assert_outputs_same(after_outputs, outputs)
def assert_outputs_same(self, after_outputs, outputs):
# Make sure we don't have nans
out_1 = after_outputs[0].numpy()
if isinstance(after_outputs, tf.Tensor):
out_1 = after_outputs.numpy()
else:
out_1 = after_outputs[0].numpy()
out_2 = outputs[0].numpy()
self.assertEqual(out_1.shape, out_2.shape)
out_1 = out_1[~np.isnan(out_1)]
@@ -269,7 +288,6 @@ class TFModelTesterMixin:
inputs_keywords = copy.deepcopy(inputs_dict)
input_ids = inputs_keywords.pop("input_ids" if not self.is_encoder_decoder else "inputs", None,)
outputs_keywords = model(input_ids, **inputs_keywords)
output_dict = outputs_dict[0].numpy()
output_keywords = outputs_keywords[0].numpy()