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

Fix TF s2s models (#9478)

Browse Source 
* Fix Seq2Seq models for serving

* Apply style

* Fix lonfgormer

* Fix mBart/Pegasus/Blenderbot

* Apply style

* Add a main intermediate layer

* Apply style

* Remove import

* Apply tf.function to Longformer

* Fix utils check_copy

* Update S2S template

* Fix BART + Blenderbot

* Fix BlenderbotSmall

* Fix BlenderbotSmall

* Fix BlenderbotSmall

* Fix MBart

* Fix Marian

* Fix Pegasus + template

* Apply style

* Fix common attributes test

* Forgot to fix the LED test

* Apply Patrick's comment on LED Decoder
This commit is contained in:
Julien Plu
2021-01-21 17:03:29 +01:00
committed by GitHub
parent 23e5a36ee6
commit a7dabfb3d1
20 changed files with 1025 additions and 666 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
tests/test_modeling_tf_common.py
View File

@@ -211,36 +211,35 @@ class TFModelTesterMixin:
def test_saved_model_with_hidden_states_output(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.output_hidden_states = True
config.output_attentions = False
if hasattr(config, "use_cache"):
config.use_cache = False
for model_class in self.all_model_classes:
class_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
# A saved model is always executed in graph mode, since we merged the PR #8777
# the booleans in graph mode are always the ones in the config, then we update
# the use_cache property if it exists in order to have similar booleans with the inputs
if "use_cache" in class_inputs_dict:
config.use_cache = class_inputs_dict.pop("use_cache")
model = model_class(config)
num_out = len(model(class_inputs_dict))
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
saved_model_dir = os.path.join(tmpdirname, "saved_model")
model = tf.keras.models.load_model(saved_model_dir)
model.save_pretrained(tmpdirname, saved_model=True)
model = tf.keras.models.load_model(os.path.join(tmpdirname, "saved_model", "1"))
outputs = model(class_inputs_dict)
if self.is_encoder_decoder:
output = outputs["encoder_hidden_states"] if isinstance(outputs, dict) else outputs[-1]
output = outputs["encoder_hidden_states"]
else:
output = outputs["hidden_states"] if isinstance(outputs, dict) else outputs[-1]
output = outputs["hidden_states"]
hidden_states = [t.numpy() for t in output]
self.assertEqual(len(outputs), num_out)
expected_num_layers = getattr(
self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
)
self.assertEqual(len(hidden_states), expected_num_layers)
self.assertEqual(len(output), expected_num_layers)
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
list(output[0].shape[-2:]),
[self.model_tester.seq_length, self.model_tester.hidden_size],
)
@@ -248,36 +247,33 @@ class TFModelTesterMixin:
def test_saved_model_with_attentions_output(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.output_attentions = True
config.output_hidden_states = False
if hasattr(config, "use_cache"):
config.use_cache = False
encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", self.model_tester.seq_length)
encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
for model_class in self.all_model_classes:
class_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
# A saved model is always executed in graph mode, since we merged the PR #8777
# the booleans in graph mode are always the ones in the config, then we update
# the use_cache property if it exists in order to have similar booleans with the inputs
if "use_cache" in class_inputs_dict:
config.use_cache = class_inputs_dict.pop("use_cache")
model = model_class(config)
num_out = len(model(class_inputs_dict))
with tempfile.TemporaryDirectory() as tmpdirname:
saved_model_dir = os.path.join(tmpdirname, "saved_model")
model.save_pretrained(saved_model_dir)
model = tf.keras.models.load_model(saved_model_dir)
model.save_pretrained(tmpdirname, saved_model=True)
model = tf.keras.models.load_model(os.path.join(tmpdirname, "saved_model", "1"))
outputs = model(class_inputs_dict)
if self.is_encoder_decoder:
output = outputs["encoder_attentions"] if isinstance(outputs, dict) else outputs[-1]
output = outputs["encoder_attentions"]
else:
output = outputs["attentions"] if isinstance(outputs, dict) else outputs[-1]
output = outputs["attentions"]
attentions = [t.numpy() for t in output]
self.assertEqual(len(outputs), num_out)
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
self.assertEqual(len(output), num_out)
self.assertEqual(len(output), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(attentions[0].shape[-3:]),
list(output[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)