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

Fix functional TF Whisper and modernize tests (#24301)

Browse Source 
* Revert whisper change and modify the test_compile_tf_model test

* make fixup

* Tweak test slightly

* Add functional model saving to test

* Ensure TF can infer shapes for data2vec

* Add override for efficientformer

* Mark test as slow
This commit is contained in:
Matt
2023-06-16 14:43:43 +01:00
committed by GitHub
parent ba3fb4b8d7
commit 62d71f4083
13 changed files with 40 additions and 300 deletions
Download Patch File Download Diff File Expand all files Collapse all files

105
tests/test_modeling_tf_common.py
View File

@@ -685,105 +685,30 @@ class TFModelTesterMixin:
if tf_inputs_dict_with_labels:
self.check_pt_tf_models(tf_model, pt_model, tf_inputs_dict_with_labels)
@slow
def test_compile_tf_model(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
max_input = getattr(self.model_tester, "max_position_embeddings", 512)
optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
metric = tf.keras.metrics.SparseCategoricalAccuracy("accuracy")
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
if model_class.__name__ in ["TFSpeech2TextModel", "TFSpeech2TextForConditionalGeneration"]:
inputs = {
"decoder_input_ids": tf.keras.Input(
batch_shape=(2, max_input),
name="decoder_input_ids",
dtype="int32",
),
"input_features": tf.keras.Input(
batch_shape=(
2,
max_input,
self.model_tester.input_feat_per_channel * self.model_tester.input_channels,
),
name="input_features",
dtype="float32",
),
}
elif model_class.__name__ in ["TFWhisperModel", "TFWhisperForConditionalGeneration"]:
inputs = {
"decoder_input_ids": tf.keras.Input(
batch_shape=(2, max_input),
name="decoder_input_ids",
dtype="int32",
),
"input_features": tf.keras.Input(
batch_shape=(
2,
self.model_tester.num_mel_bins,
self.model_tester.seq_length,
),
name="input_features",
dtype="float32",
),
}
elif self.is_encoder_decoder:
inputs = {
"decoder_input_ids": tf.keras.Input(
batch_shape=(2, max_input),
name="decoder_input_ids",
dtype="int32",
),
"input_ids": tf.keras.Input(batch_shape=(2, max_input), name="input_ids", dtype="int32"),
}
# `pixel_values` implies that the input is an image
elif model_class.main_input_name == "pixel_values":
inputs = tf.keras.Input(
batch_shape=(
3,
self.model_tester.num_channels,
self.model_tester.image_size,
self.model_tester.image_size,
),
name="pixel_values",
dtype="float32",
)
elif model_class.__name__ in ["TFCLIPModel", "TFGroupViTModel", "TFBlipModel"]:
inputs = {
"input_ids": tf.keras.Input(batch_shape=(3, max_input), name="input_ids", dtype="int32"),
"pixel_values": tf.keras.Input(
batch_shape=(
3,
self.model_tester.vision_model_tester.num_channels,
self.model_tester.vision_model_tester.image_size,
self.model_tester.vision_model_tester.image_size,
),
name="pixel_values",
dtype="float32",
),
}
elif model_class in get_values(TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING):
inputs = tf.keras.Input(batch_shape=(4, 2, max_input), name="input_ids", dtype="int32")
else:
inputs = tf.keras.Input(batch_shape=(2, max_input), name="input_ids", dtype="int32")
# Prepare our model
model = model_class(config)
model(self._prepare_for_class(inputs_dict, model_class)) # Model must be called before saving.
# Let's load it from the disk to be sure we can use pretrained weights
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname, saved_model=False)
model = model_class.from_pretrained(tmpdirname)
# These are maximally general inputs for the model, with multiple None dimensions
# Hopefully this will catch any conditionals that fail for flexible shapes
functional_inputs = {
key: tf.keras.Input(shape=val.shape[1:], dtype=val.dtype, name=key)
for key, val in model.input_signature.items()
if key in model.dummy_inputs
}
outputs_dict = model(functional_inputs)
outputs_dict = model(inputs)
hidden_states = outputs_dict[0]
# Add a dense layer on top to test integration with other keras modules
outputs = tf.keras.layers.Dense(2, activation="softmax", name="outputs")(hidden_states)
# Compile extended model
extended_model = tf.keras.Model(inputs=[inputs], outputs=[outputs])
extended_model.compile(optimizer=optimizer, loss=loss, metrics=[metric])
functional_model = tf.keras.Model(inputs=functional_inputs, outputs=hidden_states)
model_out = functional_model.predict(model.dummy_inputs) # Check we can pass inputs with the Keras API
self.assertTrue(model_out is not None)
with tempfile.TemporaryDirectory() as tmpdirname:
functional_model.save(tmpdirname) # Ensure we can save/export the whole functional model
def test_keyword_and_dict_args(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()