def check_encoder_decoder_model_from_pretrained(
self,
config,
input_ids,
attention_mask,
encoder_hidden_states,
decoder_config,
decoder_input_ids,
decoder_attention_mask,
return_dict,
**kwargs
):
encoder_model, decoder_model = self.get_encoder_decoder_model(config, decoder_config)
kwargs = {"encoder_model": encoder_model, "decoder_model": decoder_model, "return_dict": return_dict}
enc_dec_model = TFEncoderDecoderModel.from_encoder_decoder_pretrained(**kwargs)
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,
return_dict=True,
kwargs=kwargs,
)
self.assertEqual(
outputs_encoder_decoder["logits"].shape, (decoder_input_ids.shape + (decoder_config.vocab_size,))
)
self.assertEqual(
outputs_encoder_decoder["encoder_last_hidden_state"].shape, (input_ids.shape + (config.hidden_size,))
)
def test_bert2gpt2_summarization(self):
from transformers import EncoderDecoderModel
tokenizer_in = AutoTokenizer.from_pretrained("bert-base-cased")
tokenizer_out = AutoTokenizer.from_pretrained("../gpt2")
"""Not working, because pt checkpoint has `encoder.encoder.layer...` while tf model has `encoder.bert.encoder.layer...`.
(For GPT2 decoder, there is no issue)
model = TFEncoderDecoderModel.from_pretrained("patrickvonplaten/bert2gpt2-cnn_dailymail-fp16", from_pt=True)
"""
# workaround to load from pt
_model = EncoderDecoderModel.from_pretrained("patrickvonplaten/bert2gpt2-cnn_dailymail-fp16")
_model.encoder.save_pretrained("./encoder")
_model.decoder.save_pretrained("./decoder")
model = TFEncoderDecoderModel.from_encoder_decoder_pretrained(
"./encoder", "./decoder", encoder_from_pt=True, decoder_from_pt=True
)
model.config = _model.config
ARTICLE_STUDENTS = """(CNN)Sigma Alpha Epsilon is under fire for a video showing party-bound fraternity members singing a racist chant. SAE's national chapter suspended the students, but University of Oklahoma President David Boren took it a step further, saying the university's affiliation with the fraternity is permanently done. The news is shocking, but it's not the first time SAE has faced controversy. SAE was founded March 9, 1856, at the University of Alabama, five years before the American Civil War, according to the fraternity website. When the war began, the group had fewer than 400 members, of which "369 went to war for the Confederate States and seven for the Union Army," the website says. The fraternity now boasts more than 200,000 living alumni, along with about 15,000 undergraduates populating 219 chapters and 20 "colonies" seeking full membership at universities. SAE has had to work hard to change recently after a string of member deaths, many blamed on the hazing of new recruits, SAE national President Bradley Cohen wrote in a message on the fraternity's website. The fraternity's website lists more than 130 chapters cited or suspended for "health and safety incidents" since 2010. At least 30 of the incidents involved hazing, and dozens more involved alcohol. However, the list is missing numerous incidents from recent months. Among them, according to various media outlets: Yale University banned the SAEs from campus activities last month after members allegedly tried to interfere with a sexual misconduct investigation connected to an initiation rite. Stanford University in December suspended SAE housing privileges after finding sorority members attending a fraternity function were subjected to graphic sexual content. And Johns Hopkins University in November suspended the fraternity for underage drinking. "The media has labeled us as the 'nation's deadliest fraternity,' " Cohen said. In 2011, for example, a student died while being coerced into excessive alcohol consumption, according to a lawsuit. SAE's previous insurer dumped the fraternity. "As a result, we are paying Lloyd's of London the highest insurance rates in the Greek-letter world," Cohen said. Universities have turned down SAE's attempts to open new chapters, and the fraternity had to close 12 in 18 months over hazing incidents."""
EXPECTED_SUMMARY_STUDENTS = """SAS Alpha Epsilon suspended the students, but university president says it's permanent.\nThe fraternity has had to deal with a string of student deaths since 2010.\nSAS has more than 200,000 members, many of whom are students.\nA student died while being forced into excessive alcohol consumption."""
input_dict = tokenizer_in(ARTICLE_STUDENTS, return_tensors="tf")
output_ids = model.generate(input_ids=input_dict["input_ids"], max_length=None).numpy().tolist()
summary = tokenizer_out.batch_decode(output_ids, skip_special_tokens=True)
self.assertEqual(summary, [EXPECTED_SUMMARY_STUDENTS])
def test_encoder_decoder_from_pretrained(self):
load_weight_prefix = TFEncoderDecoderModel.load_weight_prefix
config = self.get_encoder_decoder_config()
encoder_tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
decoder_tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
input_ids = encoder_tokenizer("who sings does he love me with reba", return_tensors="tf").input_ids
decoder_input_ids = decoder_tokenizer("Linda Davis", return_tensors="tf").input_ids
with tempfile.TemporaryDirectory() as tmp_dirname:
# Since most of HF's models don't have pretrained cross-attention layers, they are randomly
# initialized even if we create models using `from_pretrained` method.
# For the tests, the decoder need to be a model with pretrained cross-attention layers.
# So we create pretrained models (without `load_weight_prefix`), save them, and later,
# we load them using `from_pretrained`.
# (we don't need to do this for encoder, but let's make the code more similar between encoder/decoder)
encoder = TFAutoModel.from_pretrained("bert-base-uncased", name="encoder")
# It's necessary to specify `add_cross_attention=True` here.
decoder = TFAutoModelForCausalLM.from_pretrained(
"bert-base-uncased", is_decoder=True, add_cross_attention=True, name="decoder"
)
pretrained_encoder_dir = os.path.join(tmp_dirname, "pretrained_encoder")
pretrained_decoder_dir = os.path.join(tmp_dirname, "pretrained_decoder")
encoder.save_pretrained(pretrained_encoder_dir)
decoder.save_pretrained(pretrained_decoder_dir)
del encoder
del decoder
enc_dec_model = TFEncoderDecoderModel.from_encoder_decoder_pretrained(
pretrained_encoder_dir,
pretrained_decoder_dir,
)
# check that the from pretrained methods work
enc_dec_model.save_pretrained(tmp_dirname)
enc_dec_model = TFEncoderDecoderModel.from_pretrained(tmp_dirname)
output = enc_dec_model(input_ids, decoder_input_ids=decoder_input_ids, labels=decoder_input_ids)
loss_pretrained = output.loss
del enc_dec_model
# Create the model using `__init__` with loaded ``pretrained`` encoder / decoder
encoder = TFAutoModel.from_pretrained(
pretrained_encoder_dir, load_weight_prefix=load_weight_prefix, name="encoder"
)
decoder = TFAutoModelForCausalLM.from_pretrained(
pretrained_decoder_dir, load_weight_prefix=load_weight_prefix, name="decoder"
)
enc_dec_model = TFEncoderDecoderModel(config=config, encoder=encoder, decoder=decoder)
output = enc_dec_model(input_ids, decoder_input_ids=decoder_input_ids, labels=decoder_input_ids)
loss_init = output.loss
max_diff = np.max(np.abs(loss_pretrained - loss_init))
expected_diff = 0.0
self.assertAlmostEqual(max_diff, expected_diff, places=4)
def check_pt_tf_equivalence(self, pt_model, tf_model, inputs_dict):
pt_model.to(torch_device)
pt_model.eval()
# prepare inputs
tf_inputs = inputs_dict
pt_inputs = {k: torch.tensor(v.numpy()) for k, v in tf_inputs.items()}
with torch.no_grad():
pt_outputs = pt_model(**pt_inputs).to_tuple()
tf_outputs = tf_model(**inputs_dict).to_tuple()
self.assertEqual(len(tf_outputs), len(pt_outputs), "Output lengths differ between TF and PyTorch")
for tf_output, pt_output in zip(tf_outputs, pt_outputs):
self.assert_almost_equals(tf_output.numpy(), pt_output.numpy(), 1e-3)
# PT -> TF
with tempfile.TemporaryDirectory() as encoder_tmp_dirname, tempfile.TemporaryDirectory() as decoder_tmp_dirname:
pt_model.encoder.save_pretrained(encoder_tmp_dirname)
pt_model.decoder.save_pretrained(decoder_tmp_dirname)
tf_model_loaded = TFEncoderDecoderModel.from_encoder_decoder_pretrained(
encoder_tmp_dirname, decoder_tmp_dirname, encoder_from_pt=True, decoder_from_pt=True
)
# This is only for copying some specific attributes of this particular model.
tf_model_loaded.config = pt_model.config
tf_outputs_loaded = tf_model_loaded(**inputs_dict).to_tuple()
self.assertEqual(len(tf_outputs_loaded), len(pt_outputs), "Output lengths differ between TF and PyTorch")
for tf_output_loaded, pt_output in zip(tf_outputs_loaded, pt_outputs):
self.assert_almost_equals(tf_output_loaded.numpy(), pt_output.numpy(), 1e-3)
def test_encoder_decoder_save_load_from_encoder_decoder_from_pt(self):
config = self.get_encoder_decoder_config_small()
# create two random BERT models for bert2bert & initialize weights (+cross_attention weights)
encoder_pt = BertModel(config.encoder).to(torch_device).eval()
decoder_pt = BertLMHeadModel(config.decoder).to(torch_device).eval()
encoder_decoder_pt = EncoderDecoderModel(encoder=encoder_pt, decoder=decoder_pt).to(torch_device).eval()
input_ids = ids_tensor([13, 5], encoder_pt.config.vocab_size)
decoder_input_ids = ids_tensor([13, 1], decoder_pt.config.vocab_size)
pt_input_ids = torch.tensor(input_ids.numpy(), device=torch_device, dtype=torch.long)
pt_decoder_input_ids = torch.tensor(decoder_input_ids.numpy(), device=torch_device, dtype=torch.long)
logits_pt = encoder_decoder_pt(input_ids=pt_input_ids, decoder_input_ids=pt_decoder_input_ids).logits
# PyTorch => TensorFlow
with tempfile.TemporaryDirectory() as tmp_dirname_1, tempfile.TemporaryDirectory() as tmp_dirname_2:
encoder_decoder_pt.encoder.save_pretrained(tmp_dirname_1)
encoder_decoder_pt.decoder.save_pretrained(tmp_dirname_2)
encoder_decoder_tf = TFEncoderDecoderModel.from_encoder_decoder_pretrained(
tmp_dirname_1, tmp_dirname_2, encoder_from_pt=True, decoder_from_pt=True
)
logits_tf = encoder_decoder_tf(input_ids=input_ids, decoder_input_ids=decoder_input_ids).logits
max_diff = np.max(np.abs(logits_pt.detach().cpu().numpy() - logits_tf.numpy()))
self.assertAlmostEqual(max_diff, 0.0, places=3)
# Make sure `from_pretrained` following `save_pretrained` work and give the same result
with tempfile.TemporaryDirectory() as tmp_dirname:
encoder_decoder_tf.save_pretrained(tmp_dirname)
encoder_decoder_tf = TFEncoderDecoderModel.from_pretrained(tmp_dirname)
logits_tf_2 = encoder_decoder_tf(input_ids=input_ids, decoder_input_ids=decoder_input_ids).logits
max_diff = np.max(np.abs(logits_tf_2.numpy() - logits_tf.numpy()))
self.assertAlmostEqual(max_diff, 0.0, places=3)
# TensorFlow => PyTorch
with tempfile.TemporaryDirectory() as tmp_dirname:
encoder_decoder_tf.save_pretrained(tmp_dirname)
encoder_decoder_pt = EncoderDecoderModel.from_pretrained(tmp_dirname, from_tf=True)
max_diff = np.max(np.abs(logits_pt.detach().cpu().numpy() - logits_tf.numpy()))
self.assertAlmostEqual(max_diff, 0.0, places=3)
def test_encoder_decoder_save_load_from_encoder_decoder(self):
config = self.get_encoder_decoder_config_small()
# create two random BERT models for bert2bert & initialize weights (+cross_attention weights)
encoder = TFBertModel(config.encoder)
encoder(encoder.dummy_inputs)
decoder = TFBertLMHeadModel(config.decoder)
decoder(decoder.dummy_inputs)
encoder_decoder_orig = TFEncoderDecoderModel(encoder=encoder,
decoder=decoder)
input_ids = ids_tensor([13, 5], encoder.config.vocab_size)
decoder_input_ids = ids_tensor([13, 1], decoder.config.vocab_size)
logits_orig = encoder_decoder_orig(
input_ids=input_ids, decoder_input_ids=decoder_input_ids).logits
with tempfile.TemporaryDirectory() as tmp_dirname:
encoder_path = os.path.join(tmp_dirname, "encoder")
decoder_path = os.path.join(tmp_dirname, "decoder")
encoder.save_pretrained(encoder_path)
decoder.save_pretrained(decoder_path)
encoder_decoder = TFEncoderDecoderModel.from_encoder_decoder_pretrained(
encoder_path, decoder_path)
logits_1 = encoder_decoder(input_ids=input_ids,
decoder_input_ids=decoder_input_ids).logits
self.assertTrue(
logits_orig.numpy().sum() - logits_1.numpy().sum() < 1e-3)
max_diff = np.max(np.abs(logits_1.numpy() - logits_orig.numpy()))
self.assertAlmostEqual(max_diff, 0.0, places=4)
with tempfile.TemporaryDirectory() as tmp_dirname:
encoder_decoder.save_pretrained(tmp_dirname)
encoder_decoder = TFEncoderDecoderModel.from_pretrained(
tmp_dirname)
logits_2 = encoder_decoder(input_ids=input_ids,
decoder_input_ids=decoder_input_ids).logits
max_diff = np.max(np.abs(logits_2.numpy() - logits_orig.numpy()))
self.assertAlmostEqual(max_diff, 0.0, places=4)
def check_equivalence_pt_to_tf(self, config, decoder_config, inputs_dict):
encoder_decoder_config = EncoderDecoderConfig.from_encoder_decoder_configs(config, decoder_config)
pt_model = EncoderDecoderModel(encoder_decoder_config)
with tempfile.TemporaryDirectory() as encoder_tmp_dirname, tempfile.TemporaryDirectory() as decoder_tmp_dirname:
pt_model.encoder.save_pretrained(encoder_tmp_dirname)
pt_model.decoder.save_pretrained(decoder_tmp_dirname)
tf_model = TFEncoderDecoderModel.from_encoder_decoder_pretrained(
encoder_tmp_dirname, decoder_tmp_dirname, encoder_from_pt=True, decoder_from_pt=True
)
# This is only for copying some specific attributes of this particular model.
tf_model.config = pt_model.config
self.check_pt_tf_equivalence(pt_model, tf_model, inputs_dict)
def check_pt_tf_equivalence(self, pt_model, tf_model, inputs_dict):
pt_model.to(torch_device)
pt_model.eval()
# prepare inputs
tf_inputs = inputs_dict
pt_inputs = {k: torch.tensor(v.numpy()) for k, v in tf_inputs.items()}
if "labels" in pt_inputs:
pt_inputs["labels"] = pt_inputs["labels"].type(torch.LongTensor)
# send pytorch inputs to the correct device
pt_inputs = {k: v.to(device=torch_device) if isinstance(v, torch.Tensor) else v for k, v in pt_inputs.items()}
with torch.no_grad():
pt_outputs = pt_model(**pt_inputs).to_tuple()
tf_outputs = tf_model(**inputs_dict)
if "loss" in tf_outputs:
tf_outputs.loss = tf.math.reduce_mean(tf_outputs.loss)
tf_outputs = tf_outputs.to_tuple()
self.assertEqual(len(tf_outputs), len(pt_outputs), "Output lengths differ between TF and PyTorch")
for tf_output, pt_output in zip(tf_outputs, pt_outputs):
self.assert_almost_equals(tf_output.numpy(), pt_output.detach().to("cpu").numpy(), 1e-3)
# PT -> TF
with tempfile.TemporaryDirectory() as encoder_tmp_dirname, tempfile.TemporaryDirectory() as decoder_tmp_dirname:
pt_model.encoder.save_pretrained(encoder_tmp_dirname)
pt_model.decoder.save_pretrained(decoder_tmp_dirname)
tf_model_loaded = TFEncoderDecoderModel.from_encoder_decoder_pretrained(
encoder_tmp_dirname, decoder_tmp_dirname, encoder_from_pt=True, decoder_from_pt=True
)
# This is only for copying some specific attributes of this particular model.
tf_model_loaded.config = pt_model.config
tf_outputs_loaded = tf_model_loaded(**inputs_dict)
if "loss" in tf_outputs_loaded:
tf_outputs_loaded.loss = tf.math.reduce_mean(tf_outputs_loaded.loss)
tf_outputs_loaded = tf_outputs_loaded.to_tuple()
self.assertEqual(len(tf_outputs_loaded), len(pt_outputs), "Output lengths differ between TF and PyTorch")
for tf_output_loaded, pt_output in zip(tf_outputs_loaded, pt_outputs):
self.assert_almost_equals(tf_output_loaded.numpy(), pt_output.detach().to("cpu").numpy(), 1e-3)
def get_from_encoderdecoder_pretrained_model(self):
return TFEncoderDecoderModel.from_encoder_decoder_pretrained("bert-base-cased", "bert-base-cased")
def get_pretrained_model(self):
return TFEncoderDecoderModel.from_encoder_decoder_pretrained(
"hf-internal-testing/tiny-random-rembert",
"hf-internal-testing/tiny-random-rembert",
)
def get_pretrained_model(self):
return TFEncoderDecoderModel.from_encoder_decoder_pretrained("google/rembert", "google/rembert")
def get_pretrained_model(self):
return TFEncoderDecoderModel.from_encoder_decoder_pretrained("roberta-base", "roberta-base")