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

Add FlaxBartForCausalLM (#15995)

Browse Source 
* add causal lm

* add CausalLM tests

* Add FlaxBartForCausalLM

* Add EncoderDecoder model tests

* change docstring

* make repo-consistency

* suggested changes

* remove jax ops

* correction

* rename pre-trained decoder model
This commit is contained in:
Sanchit Gandhi
2022-03-09 19:53:01 +01:00
committed by GitHub
parent 50dd314d93
commit b256f3518d
10 changed files with 543 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

98
tests/bart/test_modeling_flax_bart.py
View File

@@ -11,17 +11,16 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import unittest
import numpy as np
import timeout_decorator # noqa
from transformers import BartConfig, is_flax_available
from transformers import BartConfig, BartTokenizer, is_flax_available
from transformers.testing_utils import require_flax, slow
from ..generation.test_generation_flax_utils import FlaxGenerationTesterMixin
from ..test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor
from ..test_modeling_flax_common import FlaxModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
if is_flax_available():
@@ -34,7 +33,6 @@ if is_flax_available():
import jax
import jax.numpy as jnp
from transformers import BartTokenizer
from transformers.models.bart.modeling_flax_bart import (
FlaxBartForConditionalGeneration,
FlaxBartForQuestionAnswering,
@@ -475,3 +473,95 @@ class FlaxBartModelTest(FlaxModelTesterMixin, unittest.TestCase, FlaxGenerationT
hypotheses_batch.tolist(), clean_up_tokenization_spaces=True, skip_special_tokens=True
)
assert generated_summaries == EXPECTED
class FlaxBartStandaloneDecoderModelTester:
def __init__(
self,
parent,
batch_size=13,
seq_length=7,
is_training=True,
use_attention_mask=True,
use_labels=False,
vocab_size=99,
hidden_size=16,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=4,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=32,
eos_token_id=2,
pad_token_id=1,
bos_token_id=0,
initializer_range=0.02,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_attention_mask = use_attention_mask
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.eos_token_id = eos_token_id
self.pad_token_id = pad_token_id
self.bos_token_id = bos_token_id
self.initializer_range = initializer_range
def prepare_config_and_inputs(self):
input_ids = jnp.clip(ids_tensor([self.batch_size, self.seq_length], self.vocab_size), 3, self.vocab_size)
attention_mask = None
if self.use_attention_mask:
attention_mask = random_attention_mask([self.batch_size, self.seq_length])
config = BartConfig(
vocab_size=self.vocab_size,
d_model=self.hidden_size,
encoder_layers=self.num_hidden_layers,
decoder_layers=self.num_hidden_layers,
encoder_attention_heads=self.num_attention_heads,
decoder_attention_heads=self.num_attention_heads,
encoder_ffn_dim=self.intermediate_size,
decoder_ffn_dim=self.intermediate_size,
dropout=self.hidden_dropout_prob,
attention_dropout=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
eos_token_id=self.eos_token_id,
bos_token_id=self.bos_token_id,
pad_token_id=self.pad_token_id,
initializer_range=self.initializer_range,
use_cache=False,
)
return config, input_ids, attention_mask
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, input_ids, attention_mask = config_and_inputs
inputs_dict = {"input_ids": input_ids, "attention_mask": attention_mask}
return config, inputs_dict
def prepare_config_and_inputs_for_decoder(self):
config, input_ids, attention_mask = self.prepare_config_and_inputs()
encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size])
encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
return (
config,
input_ids,
attention_mask,
encoder_hidden_states,
encoder_attention_mask,
)

40
tests/encoder_decoder/test_modeling_flax_encoder_decoder.py
View File

@@ -22,6 +22,7 @@ import numpy as np
from transformers import is_flax_available, is_torch_available
from transformers.testing_utils import is_pt_flax_cross_test, require_flax, slow, torch_device
from ..bart.test_modeling_flax_bart import FlaxBartStandaloneDecoderModelTester
from ..bert.test_modeling_flax_bert import FlaxBertModelTester
from ..gpt2.test_modeling_flax_gpt2 import FlaxGPT2ModelTester
from ..test_modeling_flax_common import ids_tensor
@@ -31,6 +32,7 @@ if is_flax_available():
from transformers import (
AutoTokenizer,
EncoderDecoderConfig,
FlaxBartForCausalLM,
FlaxBertModel,
FlaxEncoderDecoderModel,
FlaxGPT2LMHeadModel,
@@ -360,6 +362,7 @@ class FlaxEncoderDecoderMixin:
self.assertTrue(decoder_config.cross_attention_hidden_size is None)
# check without `enc_to_dec_proj` projection
decoder_config.hidden_size = config.hidden_size
self.assertTrue(config.hidden_size == decoder_config.hidden_size)
self.check_equivalence_pt_to_flax(config, decoder_config, inputs_dict)
self.check_equivalence_flax_to_pt(config, decoder_config, inputs_dict)
@@ -456,6 +459,43 @@ class FlaxGPT2EncoderDecoderModelTest(FlaxEncoderDecoderMixin, unittest.TestCase
self.assertEqual(summary, [EXPECTED_SUMMARY_STUDENTS])
@require_flax
class FlaxBartEncoderDecoderModelTest(FlaxEncoderDecoderMixin, unittest.TestCase):
def get_encoder_decoder_model(self, config, decoder_config):
encoder_model = FlaxBertModel(config)
decoder_model = FlaxBartForCausalLM(decoder_config)
return encoder_model, decoder_model
def prepare_config_and_inputs(self):
model_tester_encoder = FlaxBertModelTester(self, batch_size=13)
model_tester_decoder = FlaxBartStandaloneDecoderModelTester(self, batch_size=13)
encoder_config_and_inputs = model_tester_encoder.prepare_config_and_inputs()
decoder_config_and_inputs = model_tester_decoder.prepare_config_and_inputs_for_decoder()
(config, input_ids, token_type_ids, attention_mask) = encoder_config_and_inputs
(
decoder_config,
decoder_input_ids,
decoder_attention_mask,
encoder_hidden_states,
encoder_attention_mask,
) = decoder_config_and_inputs
# make sure that cross attention layers are added
decoder_config.add_cross_attention = True
return {
"config": config,
"input_ids": input_ids,
"attention_mask": attention_mask,
"decoder_config": decoder_config,
"decoder_input_ids": decoder_input_ids,
"decoder_attention_mask": decoder_attention_mask,
"encoder_hidden_states": encoder_hidden_states,
}
def get_pretrained_model(self):
return FlaxEncoderDecoderModel.from_encoder_decoder_pretrained("bert-base-cased", "facebook/bart-base")
@require_flax
class FlaxEncoderDecoderModelTest(unittest.TestCase):
def get_from_encoderdecoder_pretrained_model(self):

119
tests/speech_encoder_decoder/test_modeling_flax_speech_encoder_decoder.py
View File

@@ -21,6 +21,7 @@ import numpy as np
from transformers import is_flax_available, is_torch_available
from transformers.testing_utils import is_pt_flax_cross_test, require_flax, slow, torch_device
from ..bart.test_modeling_flax_bart import FlaxBartStandaloneDecoderModelTester
from ..gpt2.test_modeling_flax_gpt2 import FlaxGPT2ModelTester
from ..test_modeling_flax_common import floats_tensor, ids_tensor, random_attention_mask
from ..wav2vec2.test_modeling_flax_wav2vec2 import FlaxWav2Vec2ModelTester
@@ -28,6 +29,7 @@ from ..wav2vec2.test_modeling_flax_wav2vec2 import FlaxWav2Vec2ModelTester
if is_flax_available():
from transformers import (
FlaxBartForCausalLM,
FlaxGPT2LMHeadModel,
FlaxSpeechEncoderDecoderModel,
FlaxWav2Vec2Model,
@@ -553,3 +555,120 @@ class FlaxWav2Vec2GPT2ModelTest(FlaxEncoderDecoderMixin, unittest.TestCase):
self.assertEqual(len(fx_outputs), len(pt_outputs_loaded), "Output lengths differ between Flax and PyTorch")
self.assert_almost_equals(fx_logits, pt_logits_loaded.numpy(), 4e-2)
@require_flax
class FlaxWav2Vec2BartModelTest(FlaxEncoderDecoderMixin, unittest.TestCase):
def get_pretrained_model_and_inputs(self):
model = FlaxSpeechEncoderDecoderModel.from_encoder_decoder_pretrained(
"facebook/wav2vec2-large-lv60", "bart-large"
)
batch_size = 13
input_values = floats_tensor([batch_size, 512], model.config.encoder.vocab_size)
attention_mask = random_attention_mask([batch_size, 512])
decoder_input_ids = ids_tensor([batch_size, 4], model.config.decoder.vocab_size)
decoder_attention_mask = random_attention_mask([batch_size, 4])
inputs = {
"inputs": input_values,
"attention_mask": attention_mask,
"decoder_input_ids": decoder_input_ids,
"decoder_attention_mask": decoder_attention_mask,
}
return model, inputs
def get_encoder_decoder_model(self, config, decoder_config):
encoder_model = FlaxWav2Vec2Model(config)
decoder_model = FlaxBartForCausalLM(decoder_config)
return encoder_model, decoder_model
def prepare_config_and_inputs(self):
model_tester_encoder = FlaxWav2Vec2ModelTester(self, batch_size=13)
model_tester_decoder = FlaxBartStandaloneDecoderModelTester(self, batch_size=13)
encoder_config_and_inputs = model_tester_encoder.prepare_config_and_inputs()
decoder_config_and_inputs = model_tester_decoder.prepare_config_and_inputs_for_decoder()
(config, inputs, attention_mask) = encoder_config_and_inputs
(
decoder_config,
decoder_input_ids,
decoder_attention_mask,
encoder_hidden_states,
encoder_attention_mask,
) = decoder_config_and_inputs
# make sure that cross attention layers are added
decoder_config.add_cross_attention = True
return {
"config": config,
"inputs": inputs,
"attention_mask": attention_mask,
"decoder_config": decoder_config,
"decoder_input_ids": decoder_input_ids,
"decoder_attention_mask": decoder_attention_mask,
"encoder_hidden_states": encoder_hidden_states,
}
@slow
def test_flaxwav2vec2bart_pt_flax_equivalence(self):
pt_model = SpeechEncoderDecoderModel.from_pretrained("patrickvonplaten/wav2vec2-2-bart-large")
fx_model = FlaxSpeechEncoderDecoderModel.from_pretrained(
"patrickvonplaten/wav2vec2-2-bart-large", from_pt=True
)
pt_model.to(torch_device)
pt_model.eval()
# prepare inputs
batch_size = 13
input_values = floats_tensor([batch_size, 512], fx_model.config.encoder.vocab_size)
attention_mask = random_attention_mask([batch_size, 512])
decoder_input_ids = ids_tensor([batch_size, 4], fx_model.config.decoder.vocab_size)
decoder_attention_mask = random_attention_mask([batch_size, 4])
inputs_dict = {
"inputs": input_values,
"attention_mask": attention_mask,
"decoder_input_ids": decoder_input_ids,
"decoder_attention_mask": decoder_attention_mask,
}
flax_inputs = inputs_dict
pt_inputs = {k: torch.tensor(v.tolist()) for k, v in flax_inputs.items()}
with torch.no_grad():
pt_outputs = pt_model(**pt_inputs)
pt_logits = pt_outputs.logits
pt_outputs = pt_outputs.to_tuple()
fx_outputs = fx_model(**inputs_dict)
fx_logits = fx_outputs.logits
fx_outputs = fx_outputs.to_tuple()
self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
self.assert_almost_equals(fx_logits, pt_logits.numpy(), 4e-2)
# PT -> Flax
with tempfile.TemporaryDirectory() as tmpdirname:
pt_model.save_pretrained(tmpdirname)
fx_model_loaded = FlaxSpeechEncoderDecoderModel.from_pretrained(tmpdirname, from_pt=True)
fx_outputs_loaded = fx_model_loaded(**inputs_dict)
fx_logits_loaded = fx_outputs_loaded.logits
fx_outputs_loaded = fx_outputs_loaded.to_tuple()
self.assertEqual(len(fx_outputs_loaded), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
self.assert_almost_equals(fx_logits_loaded, pt_logits.numpy(), 4e-2)
# Flax -> PT
with tempfile.TemporaryDirectory() as tmpdirname:
fx_model.save_pretrained(tmpdirname)
pt_model_loaded = SpeechEncoderDecoderModel.from_pretrained(tmpdirname, from_flax=True)
pt_model_loaded.to(torch_device)
pt_model_loaded.eval()
with torch.no_grad():
pt_outputs_loaded = pt_model_loaded(**pt_inputs)
pt_logits_loaded = pt_outputs_loaded.logits
pt_outputs_loaded = pt_outputs_loaded.to_tuple()
self.assertEqual(len(fx_outputs), len(pt_outputs_loaded), "Output lengths differ between Flax and PyTorch")
self.assert_almost_equals(fx_logits, pt_logits_loaded.numpy(), 4e-2)