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

[All Seq2Seq model + CLM models that can be used with EncoderDecoder] Add cross-attention weights to outputs (#8071)

Browse Source 
* Output cross-attention with decoder attention output

* Update src/transformers/modeling_bert.py

* add cross-attention for t5 and bart as well

* fix tests

* correct typo in docs

* add sylvains and sams comments

* correct typo

Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
This commit is contained in:
Yossi Synett
2020-11-06 18:34:48 +00:00
committed by GitHub
parent 30f2507a07
commit bc0d26d1de
16 changed files with 653 additions and 85 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
tests/test_modeling_common.py
View File

@@ -253,9 +253,7 @@ class ModelTesterMixin:
out_len = len(outputs)
if self.is_encoder_decoder:
correct_outlen = (
self.model_tester.base_model_out_len if hasattr(self.model_tester, "base_model_out_len") else 4
)
correct_outlen = 5
# loss is at first position
if "labels" in inputs_dict:
@@ -266,6 +264,7 @@ class ModelTesterMixin:
self.assertEqual(out_len, correct_outlen)
# decoder attentions
decoder_attentions = outputs.decoder_attentions
self.assertIsInstance(decoder_attentions, (list, tuple))
self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
@@ -274,6 +273,19 @@ class ModelTesterMixin:
[self.model_tester.num_attention_heads, decoder_seq_length, decoder_key_length],
)
# cross attentions
cross_attentions = outputs.cross_attentions
self.assertIsInstance(cross_attentions, (list, tuple))
self.assertEqual(len(cross_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(cross_attentions[0].shape[-3:]),
[
self.model_tester.num_attention_heads,
decoder_seq_length,
encoder_key_length,
],
)
# Check attention is always last and order is fine
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = True

60
tests/test_modeling_encoder_decoder.py
View File

@@ -292,6 +292,62 @@ class EncoderDecoderMixin:
outputs_encoder_decoder["encoder_last_hidden_state"].shape, (input_ids.shape + (config.hidden_size,))
)
def check_encoder_decoder_model_output_attentions(
self,
config,
input_ids,
attention_mask,
encoder_hidden_states,
decoder_config,
decoder_input_ids,
decoder_attention_mask,
labels,
**kwargs
):
encoder_model, decoder_model = self.get_encoder_decoder_model(config, decoder_config)
enc_dec_model = EncoderDecoderModel(encoder=encoder_model, decoder=decoder_model)
enc_dec_model.to(torch_device)
outputs_encoder_decoder = enc_dec_model(
input_ids=input_ids,
decoder_input_ids=decoder_input_ids,
attention_mask=attention_mask,
decoder_attention_mask=decoder_attention_mask,
output_attentions=True,
return_dict=True,
)
encoder_attentions = outputs_encoder_decoder["encoder_attentions"]
self.assertEqual(len(encoder_attentions), config.num_hidden_layers)
self.assertListEqual(
list(encoder_attentions[0].shape[-3:]),
[config.num_attention_heads, input_ids.shape[-1], input_ids.shape[-1]],
)
decoder_attentions = outputs_encoder_decoder["decoder_attentions"]
num_decoder_layers = (
decoder_config.num_decoder_layers
if hasattr(decoder_config, "num_decoder_layers")
else decoder_config.num_hidden_layers
)
self.assertEqual(len(decoder_attentions), num_decoder_layers)
self.assertListEqual(
list(decoder_attentions[0].shape[-3:]),
[decoder_config.num_attention_heads, decoder_input_ids.shape[-1], decoder_input_ids.shape[-1]],
)
cross_attentions = outputs_encoder_decoder["cross_attentions"]
self.assertEqual(len(cross_attentions), num_decoder_layers)
cross_attention_input_seq_len = input_ids.shape[-1] * (
1 + (decoder_config.ngram if hasattr(decoder_config, "ngram") else 0)
)
self.assertListEqual(
list(cross_attentions[0].shape[-3:]),
[decoder_config.num_attention_heads, cross_attention_input_seq_len, decoder_input_ids.shape[-1]],
)
def check_encoder_decoder_model_generate(self, input_ids, config, decoder_config, **kwargs):
encoder_model, decoder_model = self.get_encoder_decoder_model(config, decoder_config)
enc_dec_model = EncoderDecoderModel(encoder=encoder_model, decoder=decoder_model)
@@ -413,6 +469,10 @@ class EncoderDecoderMixin:
input_ids_dict = self.prepare_config_and_inputs()
self.check_encoder_decoder_model_labels(**input_ids_dict)
def test_encoder_decoder_model_output_attentions(self):
input_ids_dict = self.prepare_config_and_inputs()
self.check_encoder_decoder_model_output_attentions(**input_ids_dict)
def test_encoder_decoder_model_generate(self):
input_ids_dict = self.prepare_config_and_inputs()
self.check_encoder_decoder_model_generate(**input_ids_dict)

110
tests/test_modeling_prophetnet.py
View File

@@ -916,6 +916,116 @@ class ProphetNetModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.Test
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model_fp16_forward(*config_and_inputs)
# methods overwrite method in `test_modeling_common.py`
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
seq_len = getattr(self.model_tester, "seq_length", None)
decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)
encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
decoder_key_length = getattr(self.model_tester, "decoder_key_length", decoder_seq_length)
encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
chunk_length = getattr(self.model_tester, "chunk_length", None)
if chunk_length is not None and hasattr(self.model_tester, "num_hashes"):
encoder_seq_length = encoder_seq_length * self.model_tester.num_hashes
for model_class in self.all_model_classes:
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = False
config.return_dict = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
# check that output_attentions also work using config
del inputs_dict["output_attentions"]
config.output_attentions = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
if chunk_length is not None:
self.assertListEqual(
list(attentions[0].shape[-4:]),
[self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
)
else:
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)
out_len = len(outputs)
correct_outlen = 7
# loss is at first position
if "labels" in inputs_dict:
correct_outlen += 1 # loss is added to beginning
self.assertEqual(out_len, correct_outlen)
# decoder attentions
decoder_attentions = outputs.decoder_attentions
self.assertIsInstance(decoder_attentions, (list, tuple))
self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(decoder_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, decoder_seq_length, decoder_key_length],
)
# cross attentions
cross_attentions = outputs.cross_attentions
self.assertIsInstance(cross_attentions, (list, tuple))
self.assertEqual(len(cross_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(cross_attentions[0].shape[-3:]),
[
self.model_tester.num_attention_heads,
(self.model_tester.ngram + 1) * decoder_seq_length,
encoder_key_length,
],
)
# Check attention is always last and order is fine
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
if hasattr(self.model_tester, "num_hidden_states_types"):
added_hidden_states = self.model_tester.num_hidden_states_types
elif self.is_encoder_decoder:
added_hidden_states = 2
else:
added_hidden_states = 1
self.assertEqual(out_len + added_hidden_states, len(outputs))
self_attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
if chunk_length is not None:
self.assertListEqual(
list(self_attentions[0].shape[-4:]),
[self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
)
else:
self.assertListEqual(
list(self_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)
@require_torch
class ProphetNetStandaloneDecoderModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):