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

16
tests/test_modeling_tf_bart.py
View File

@@ -264,22 +264,8 @@ class TFBartModelTest(TFModelTesterMixin, unittest.TestCase):
models_equal = False
self.assertTrue(models_equal)
@slow
def test_saved_model_with_hidden_states_output(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
@slow
def test_saved_model_with_attentions_output(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
def test_saved_model_creation(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
def test_saved_model_creation_extended(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
# This test is too long (>30sec) and makes fail the CI
pass

16
tests/test_modeling_tf_blenderbot.py
View File

@@ -200,22 +200,8 @@ class TFBlenderbotModelTest(TFModelTesterMixin, unittest.TestCase):
name = model.get_bias()
assert name is None
@slow
def test_saved_model_with_hidden_states_output(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
@slow
def test_saved_model_with_attentions_output(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
def test_saved_model_creation(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
def test_saved_model_creation_extended(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
# This test is too long (>30sec) and makes fail the CI
pass
def test_resize_token_embeddings(self):

37
tests/test_modeling_tf_blenderbot_small.py
View File

@@ -188,28 +188,19 @@ class TFBlenderbotSmallModelTest(TFModelTesterMixin, unittest.TestCase):
for model_class in self.all_model_classes:
model = model_class(config)
assert isinstance(model.get_input_embeddings(), tf.keras.layers.Layer)
x = model.get_output_layer_with_bias()
assert x is None
name = model.get_prefix_bias_name()
assert name is None
@slow
def test_saved_model_with_hidden_states_output(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
@slow
def test_saved_model_with_attentions_output(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
def test_saved_model_creation(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
def test_saved_model_creation_extended(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
if model_class in self.all_generative_model_classes:
x = model.get_output_embeddings()
assert isinstance(x, tf.keras.layers.Layer)
name = model.get_bias()
assert isinstance(name, dict)
for k, v in name.items():
assert isinstance(v, tf.Variable)
else:
x = model.get_output_embeddings()
assert x is None
name = model.get_bias()
assert name is None
def test_resize_token_embeddings(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -274,6 +265,10 @@ class TFBlenderbotSmallModelTest(TFModelTesterMixin, unittest.TestCase):
models_equal = False
self.assertTrue(models_equal)
def test_saved_model_creation(self):
# This test is too long (>30sec) and makes fail the CI
pass
def _assert_tensors_equal(a, b, atol=1e-12, prefix=""):
"""If tensors not close, or a and b arent both tensors, raise a nice Assertion error."""

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],
)

36
tests/test_modeling_tf_led.py
View File

@@ -352,30 +352,20 @@ class TFLEDModelTest(TFModelTesterMixin, unittest.TestCase):
self.assertEqual(model.config.output_hidden_states, True)
check_encoder_attentions_output(outputs)
@slow
def test_saved_model_with_attentions_output(self):
# longformer has special attentions which are not
# compatible in graph mode
pass
@slow
def test_saved_model_with_hidden_states_output(self):
# TODO(JPLU, PVP) this test should pass!!! PVP:
# IMO there is a problem with the signature check.
# Test passes for TFLEDModel, but not for TFLEDForConditionalGeneration
# IMO the reason is that the tensor variable name cannot be changed
# from decoder_input_ids -> input_ids, which poses a BIG restrictions
pass
@slow
def test_saved_model_creation_extended(self):
# All the tests about building a saved model
# fails because the Seq2Seq models uses model in a model
# as a layer.
# TODO(JPLU) WARNING: NEED TO BE FIXED ASAP
pass
def test_saved_model_creation(self):
# This test is too long (>30sec) and makes fail the CI
pass
def test_saved_model_with_attentions_output(self):
# This test don't pass because of the error:
# condition [13,8,4,5], then [13,8,4,5], and else [13,8,4,6] must be broadcastable
# This occurs line 323 in modeling_tf_led.py because the condition line 255
# returns a tensor of shape
# [batch_size, seq_len, self.num_heads, self.one_sided_attn_window_size * 2 + 2]
# if is_global_attn is True and a tensor of shape
# [batch_size, seq_len, self.num_heads, self.one_sided_attn_window_size * 2 + 1]
# This is due to the tf.concat call line 703 that adds one dimension
# Need to check with PVP how to properly fix this
pass

16
tests/test_modeling_tf_marian.py
View File

@@ -233,22 +233,8 @@ class TFMarianModelTest(TFModelTesterMixin, unittest.TestCase):
name = model.get_bias()
assert name is None
@slow
def test_saved_model_with_hidden_states_output(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
@slow
def test_saved_model_with_attentions_output(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
def test_saved_model_creation(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
def test_saved_model_creation_extended(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
# This test is too long (>30sec) and makes fail the CI
pass
def test_resize_token_embeddings(self):

16
tests/test_modeling_tf_mbart.py
View File

@@ -204,22 +204,8 @@ class TFMBartModelTest(TFModelTesterMixin, unittest.TestCase):
name = model.get_bias()
assert name is None
@slow
def test_saved_model_with_hidden_states_output(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
@slow
def test_saved_model_with_attentions_output(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
def test_saved_model_creation(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
def test_saved_model_creation_extended(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
# This test is too long (>30sec) and makes fail the CI
pass
def test_resize_token_embeddings(self):

16
tests/test_modeling_tf_pegasus.py
View File

@@ -231,22 +231,8 @@ class TFPegasusModelTest(TFModelTesterMixin, unittest.TestCase):
name = model.get_bias()
assert name is None
@slow
def test_saved_model_with_hidden_states_output(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
@slow
def test_saved_model_with_attentions_output(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
def test_saved_model_creation(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
pass
def test_saved_model_creation_extended(self):
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
# This test is too long (>30sec) and makes fail the CI
pass
def test_resize_token_embeddings(self):