class FlaxOPTModelTest(FlaxModelTesterMixin, unittest.TestCase,
FlaxGenerationTesterMixin):
all_model_classes = (FlaxOPTModel,
FlaxOPTForCausalLM) if is_flax_available() else ()
all_generative_model_classes = () if is_flax_available() else ()
def setUp(self):
self.model_tester = FlaxOPTModelTester(self)
def test_use_cache_forward(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
for model_class in self.all_model_classes:
self.model_tester.check_use_cache_forward(model_class, config,
inputs_dict)
def test_use_cache_forward_with_attn_mask(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
for model_class in self.all_model_classes:
self.model_tester.check_use_cache_forward_with_attn_mask(
model_class, config, inputs_dict)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained("facebook/opt-125m")
input_ids = np.ones((1, 1)) * model.config.eos_token_id
outputs = model(input_ids)
self.assertIsNotNone(outputs)
class FlaxAlbertModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = (
(
FlaxAlbertModel,
FlaxAlbertForPreTraining,
FlaxAlbertForMaskedLM,
FlaxAlbertForMultipleChoice,
FlaxAlbertForQuestionAnswering,
FlaxAlbertForSequenceClassification,
FlaxAlbertForTokenClassification,
FlaxAlbertForQuestionAnswering,
)
if is_flax_available()
else ()
)
def setUp(self):
self.model_tester = FlaxAlbertModelTester(self)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained("albert-base-v2")
outputs = model(np.ones((1, 1)))
self.assertIsNotNone(outputs)
class FlaxBertModelTest(FlaxModelTesterMixin, unittest.TestCase):
test_head_masking = True
all_model_classes = ((
FlaxBertModel,
FlaxBertForPreTraining,
FlaxBertForMaskedLM,
FlaxBertForMultipleChoice,
FlaxBertForQuestionAnswering,
FlaxBertForNextSentencePrediction,
FlaxBertForSequenceClassification,
FlaxBertForTokenClassification,
FlaxBertForQuestionAnswering,
) if is_flax_available() else ())
def setUp(self):
self.model_tester = FlaxBertModelTester(self)
@slow
def test_model_from_pretrained(self):
# Only check this for base model, not necessary for all model classes.
# This will also help speed-up tests.
model = FlaxBertModel.from_pretrained("bert-base-cased")
outputs = model(np.ones((1, 1)))
self.assertIsNotNone(outputs)
class FlaxBigBirdModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = ((
FlaxBigBirdModel,
FlaxBigBirdForPreTraining,
FlaxBigBirdForMaskedLM,
FlaxBigBirdForMultipleChoice,
FlaxBigBirdForQuestionAnswering,
FlaxBigBirdForSequenceClassification,
FlaxBigBirdForTokenClassification,
) if is_flax_available() else ())
test_attn_probs = False
def setUp(self):
self.model_tester = FlaxBigBirdModelTester(self)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained(
"google/bigbird-roberta-base", from_pt=True)
outputs = model(np.ones((1, 1)))
self.assertIsNotNone(outputs)
def test_attention_outputs(self):
if self.test_attn_probs:
super().test_attention_outputs()
class FlaxCLIPTextModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = (FlaxCLIPTextModel, ) if is_flax_available() else ()
def setUp(self):
self.model_tester = FlaxCLIPTextModelTester(self)
# FlaxCLIPTextModel does not have any base model
def test_save_load_from_base(self):
pass
# FlaxCLIPVisionModel does not have any base model
def test_save_load_to_base(self):
pass
# FlaxCLIPVisionModel does not have any base model
@is_pt_flax_cross_test
def test_save_load_from_base_pt(self):
pass
# FlaxCLIPVisionModel does not have any base model
@is_pt_flax_cross_test
def test_save_load_to_base_pt(self):
pass
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained(
"openai/clip-vit-base-patch32", from_pt=True)
outputs = model(np.ones((1, 1)))
self.assertIsNotNone(outputs)
class FlaxRoFormerModelTest(FlaxModelTesterMixin, unittest.TestCase):
test_head_masking = True
all_model_classes = (
(
FlaxRoFormerModel,
FlaxRoFormerForMaskedLM,
FlaxRoFormerForSequenceClassification,
FlaxRoFormerForTokenClassification,
FlaxRoFormerForMultipleChoice,
FlaxRoFormerForQuestionAnswering,
)
if is_flax_available()
else ()
)
def setUp(self):
self.model_tester = FlaxRoFormerModelTester(self)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained("junnyu/roformer_chinese_small", from_pt=True)
outputs = model(np.ones((1, 1)))
self.assertIsNotNone(outputs)
def get_all_auto_configured_models():
"""Return the list of all models in at least one auto class."""
result = set(
) # To avoid duplicates we concatenate all model classes in a set.
if is_torch_available():
for attr_name in dir(transformers.models.auto.modeling_auto):
if attr_name.startswith("MODEL_") and attr_name.endswith(
"MAPPING_NAMES"):
result = result | set(
get_values(
getattr(transformers.models.auto.modeling_auto,
attr_name)))
if is_tf_available():
for attr_name in dir(transformers.models.auto.modeling_tf_auto):
if attr_name.startswith("TF_MODEL_") and attr_name.endswith(
"MAPPING_NAMES"):
result = result | set(
get_values(
getattr(transformers.models.auto.modeling_tf_auto,
attr_name)))
if is_flax_available():
for attr_name in dir(transformers.models.auto.modeling_flax_auto):
if attr_name.startswith("FLAX_MODEL_") and attr_name.endswith(
"MAPPING_NAMES"):
result = result | set(
get_values(
getattr(transformers.models.auto.modeling_flax_auto,
attr_name)))
return [cls for cls in result]
def check_all_models_are_auto_configured():
"""Check all models are each in an auto class."""
missing_backends = []
if not is_torch_available():
missing_backends.append("PyTorch")
if not is_tf_available():
missing_backends.append("TensorFlow")
if not is_flax_available():
missing_backends.append("Flax")
if len(missing_backends) > 0:
missing = ", ".join(missing_backends)
if os.getenv("TRANSFORMERS_IS_CI", "").upper() in ENV_VARS_TRUE_VALUES:
raise Exception(
"Full quality checks require all backends to be installed (with `pip install -e .[dev]` in the "
f"Transformers repo, the following are missing: {missing}."
)
else:
warnings.warn(
"Full quality checks require all backends to be installed (with `pip install -e .[dev]` in the "
f"Transformers repo, the following are missing: {missing}. While it's probably fine as long as you "
"didn't make any change in one of those backends modeling files, you should probably execute the "
"command above to be on the safe side."
)
modules = get_model_modules()
all_auto_models = get_all_auto_configured_models()
failures = []
for module in modules:
new_failures = check_models_are_auto_configured(module, all_auto_models)
if new_failures is not None:
failures += new_failures
if len(failures) > 0:
raise Exception(f"There were {len(failures)} failures:\n" + "\n".join(failures))
class FlaxElectraModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = ((
FlaxElectraModel,
FlaxElectraForMaskedLM,
FlaxElectraForPreTraining,
FlaxElectraForTokenClassification,
FlaxElectraForQuestionAnswering,
FlaxElectraForMultipleChoice,
FlaxElectraForSequenceClassification,
) if is_flax_available() else ())
def setUp(self):
self.model_tester = FlaxElectraModelTester(self)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
if model_class_name == FlaxElectraForMaskedLM:
model = model_class_name.from_pretrained(
"google/electra-small-generator")
else:
model = model_class_name.from_pretrained(
"google/electra-small-discriminator")
outputs = model(np.ones((1, 1)))
self.assertIsNotNone(outputs)
class FlaxViTModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = (FlaxViTModel, FlaxViTForImageClassification) if is_flax_available() else ()
def setUp(self) -> None:
self.model_tester = FlaxViTModelTester(self)
self.config_tester = ConfigTester(self, config_class=ViTConfig, has_text_modality=False, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
# We neeed to override this test because ViT's forward signature is different than text models.
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.__call__)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["pixel_values"]
self.assertListEqual(arg_names[:1], expected_arg_names)
# We need to override this test because ViT expects pixel_values instead of input_ids
def test_jit_compilation(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
model = model_class(config)
@jax.jit
def model_jitted(pixel_values, **kwargs):
return model(pixel_values=pixel_values, **kwargs)
with self.subTest("JIT Enabled"):
jitted_outputs = model_jitted(**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = model_jitted(**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained("google/vit-base-patch16-224")
outputs = model(np.ones((1, 3, 224, 224)))
self.assertIsNotNone(outputs)
class FlaxCLIPTextModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = (FlaxCLIPTextModel, ) if is_flax_available() else ()
def setUp(self):
self.model_tester = FlaxCLIPTextModelTester(self)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained(
"openai/clip-vit-base-patch32", from_pt=True)
outputs = model(np.ones((1, 1)))
self.assertIsNotNone(outputs)
class FlaxRobertaModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = (FlaxRobertaModel,) if is_flax_available() else ()
def setUp(self):
self.model_tester = FlaxRobertaModelTester(self)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained("roberta-base")
outputs = model(np.ones((1, 1)))
self.assertIsNotNone(outputs)
class FlaxBertModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = (FlaxBertModel,
FlaxBertForMaskedLM) if is_flax_available() else ()
def setUp(self):
self.model_tester = FlaxBertModelTester(self)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained("bert-base-cased",
from_pt=True)
outputs = model(np.ones((1, 1)))
self.assertIsNotNone(outputs)
class FlaxRobertaModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = (FlaxRobertaModel, ) if is_flax_available() else ()
def setUp(self):
self.model_tester = FlaxRobertaModelTester(self)
class FlaxGPT2ModelTest(FlaxModelTesterMixin, FlaxGenerationTesterMixin,
unittest.TestCase):
all_model_classes = (FlaxGPT2Model,
FlaxGPT2LMHeadModel) if is_flax_available() else ()
all_generative_model_classes = (
FlaxGPT2LMHeadModel, ) if is_flax_available() else ()
def setUp(self):
self.model_tester = FlaxGPT2ModelTester(self)
def test_use_cache_forward(self):
for model_class_name in self.all_model_classes:
config, input_ids, attention_mask = self.model_tester.prepare_config_and_inputs(
)
self.model_tester.check_use_cache_forward(model_class_name, config,
input_ids,
attention_mask)
def test_use_cache_forward_with_attn_mask(self):
for model_class_name in self.all_model_classes:
config, input_ids, attention_mask = self.model_tester.prepare_config_and_inputs(
)
self.model_tester.check_use_cache_forward_with_attn_mask(
model_class_name, config, input_ids, attention_mask)
@slow
def test_batch_generation(self):
tokenizer = GPT2Tokenizer.from_pretrained("gpt2",
pad_token="</s>",
padding_side="left")
inputs = tokenizer(["Hello this is a long string", "Hey"],
return_tensors="jax",
padding=True,
truncation=True)
model = FlaxGPT2LMHeadModel.from_pretrained("gpt2")
model.do_sample = False
model.config.pad_token_id = model.config.eos_token_id
jit_generate = jax.jit(model.generate)
output_sequences = jit_generate(
inputs["input_ids"],
attention_mask=inputs["attention_mask"]).sequences
output_string = tokenizer.batch_decode(output_sequences,
skip_special_tokens=True)
expected_string = [
"Hello this is a long string of words. I'm going to try to explain what I mean.",
"Hey, I'm not sure if I'm going to be able to do",
]
self.assertListEqual(output_string, expected_string)
# overwrite from common since `attention_mask` in combination
# with `causal_mask` behaves slighly differently
@is_pt_flax_cross_test
def test_equivalence_pt_to_flax(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
# prepare inputs
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
pt_inputs = {
k: torch.tensor(v.tolist())
for k, v in prepared_inputs_dict.items()
}
# load corresponding PyTorch class
pt_model_class_name = model_class.__name__[
4:] # Skip the "Flax" at the beginning
pt_model_class = getattr(transformers, pt_model_class_name)
batch_size, seq_length = pt_inputs["input_ids"].shape
rnd_start_indices = np.random.randint(0,
seq_length - 1,
size=(batch_size, ))
for batch_idx, start_index in enumerate(rnd_start_indices):
pt_inputs["attention_mask"][batch_idx, :start_index] = 0
pt_inputs["attention_mask"][batch_idx, start_index:] = 1
prepared_inputs_dict["attention_mask"][
batch_idx, :start_index] = 0
prepared_inputs_dict["attention_mask"][batch_idx,
start_index:] = 1
pt_model = pt_model_class(config).eval()
fx_model = model_class(config, dtype=jnp.float32)
fx_state = convert_pytorch_state_dict_to_flax(
pt_model.state_dict(), fx_model)
fx_model.params = fx_state
with torch.no_grad():
pt_outputs = pt_model(**pt_inputs).to_tuple()
fx_outputs = fx_model(**prepared_inputs_dict).to_tuple()
self.assertEqual(
len(fx_outputs), len(pt_outputs),
"Output lengths differ between Flax and PyTorch")
for fx_output, pt_output in zip(fx_outputs, pt_outputs):
self.assert_almost_equals(fx_output[:, -1],
pt_output[:, -1].numpy(), 4e-2)
with tempfile.TemporaryDirectory() as tmpdirname:
pt_model.save_pretrained(tmpdirname)
fx_model_loaded = model_class.from_pretrained(tmpdirname,
from_pt=True)
fx_outputs_loaded = fx_model_loaded(
**prepared_inputs_dict).to_tuple()
self.assertEqual(
len(fx_outputs_loaded), len(pt_outputs),
"Output lengths differ between Flax and PyTorch")
for fx_output_loaded, pt_output in zip(fx_outputs_loaded,
pt_outputs):
self.assert_almost_equals(fx_output_loaded[:, -1],
pt_output[:, -1].numpy(), 4e-2)
# overwrite from common since `attention_mask` in combination
# with `causal_mask` behaves slighly differently
@is_pt_flax_cross_test
def test_equivalence_flax_to_pt(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
# prepare inputs
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
pt_inputs = {
k: torch.tensor(v.tolist())
for k, v in prepared_inputs_dict.items()
}
# load corresponding PyTorch class
pt_model_class_name = model_class.__name__[
4:] # Skip the "Flax" at the beginning
pt_model_class = getattr(transformers, pt_model_class_name)
pt_model = pt_model_class(config).eval()
fx_model = model_class(config, dtype=jnp.float32)
pt_model = load_flax_weights_in_pytorch_model(
pt_model, fx_model.params)
batch_size, seq_length = pt_inputs["input_ids"].shape
rnd_start_indices = np.random.randint(0,
seq_length - 1,
size=(batch_size, ))
for batch_idx, start_index in enumerate(rnd_start_indices):
pt_inputs["attention_mask"][batch_idx, :start_index] = 0
pt_inputs["attention_mask"][batch_idx, start_index:] = 1
prepared_inputs_dict["attention_mask"][
batch_idx, :start_index] = 0
prepared_inputs_dict["attention_mask"][batch_idx,
start_index:] = 1
# make sure weights are tied in PyTorch
pt_model.tie_weights()
with torch.no_grad():
pt_outputs = pt_model(**pt_inputs).to_tuple()
fx_outputs = fx_model(**prepared_inputs_dict).to_tuple()
self.assertEqual(
len(fx_outputs), len(pt_outputs),
"Output lengths differ between Flax and PyTorch")
for fx_output, pt_output in zip(fx_outputs, pt_outputs):
self.assert_almost_equals(fx_output[:, -1],
pt_output[:, -1].numpy(), 4e-2)
with tempfile.TemporaryDirectory() as tmpdirname:
fx_model.save_pretrained(tmpdirname)
pt_model_loaded = pt_model_class.from_pretrained(
tmpdirname, from_flax=True)
with torch.no_grad():
pt_outputs_loaded = pt_model_loaded(**pt_inputs).to_tuple()
self.assertEqual(
len(fx_outputs), len(pt_outputs_loaded),
"Output lengths differ between Flax and PyTorch")
for fx_output, pt_output in zip(fx_outputs, pt_outputs_loaded):
self.assert_almost_equals(fx_output[:, -1],
pt_output[:, -1].numpy(), 4e-2)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained("gpt2", from_pt=True)
outputs = model(np.ones((1, 1)))
self.assertIsNotNone(outputs)
class FlaxBartModelTest(FlaxModelTesterMixin, unittest.TestCase,
FlaxGenerationTesterMixin):
is_encoder_decoder = True
all_model_classes = ((
FlaxBartModel,
FlaxBartForConditionalGeneration,
FlaxBartForSequenceClassification,
FlaxBartForQuestionAnswering,
) if is_flax_available() else ())
all_generative_model_classes = (
FlaxBartForConditionalGeneration, ) if is_flax_available() else ()
def setUp(self):
self.model_tester = FlaxBartModelTester(self)
def test_use_cache_forward(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
for model_class in self.all_model_classes:
self.model_tester.check_use_cache_forward(model_class, config,
inputs_dict)
def test_use_cache_forward_with_attn_mask(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
for model_class in self.all_model_classes:
self.model_tester.check_use_cache_forward_with_attn_mask(
model_class, config, inputs_dict)
def test_encode(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
model = model_class(config)
@jax.jit
def encode_jitted(input_ids, attention_mask=None, **kwargs):
return model.encode(input_ids=input_ids,
attention_mask=attention_mask)
with self.subTest("JIT Enabled"):
jitted_outputs = encode_jitted(
**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = encode_jitted(
**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
def test_decode(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
model = model_class(config)
encoder_outputs = model.encode(inputs_dict["input_ids"],
inputs_dict["attention_mask"])
prepared_inputs_dict = {
"decoder_input_ids": inputs_dict["decoder_input_ids"],
"decoder_attention_mask":
inputs_dict["decoder_attention_mask"],
"encoder_outputs": encoder_outputs,
}
@jax.jit
def decode_jitted(decoder_input_ids, decoder_attention_mask,
encoder_outputs):
return model.decode(
decoder_input_ids=decoder_input_ids,
decoder_attention_mask=decoder_attention_mask,
encoder_outputs=encoder_outputs,
)
with self.subTest("JIT Enabled"):
jitted_outputs = decode_jitted(
**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = decode_jitted(
**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained("facebook/bart-base",
from_pt=True)
# FlaxBartForSequenceClassification expects eos token in input_ids
input_ids = np.ones((1, 1)) * model.config.eos_token_id
outputs = model(input_ids)
self.assertIsNotNone(outputs)
@slow
def test_summarization_fast(self):
model = FlaxBartForConditionalGeneration.from_pretrained(
"sshleifer/distilbart-cnn-6-6")
tokenizer = BartTokenizer.from_pretrained(
"sshleifer/distilbart-cnn-6-6")
input_str = "This sentence is made of three parts. Each part is important on its own. One part is about animals, the other part about planes, and the last part about housing."
input_ids = tokenizer(input_str, return_tensors="np").input_ids
sequences = model.generate(input_ids, num_beams=2,
max_length=20).sequences
output_str = tokenizer.batch_decode(sequences)[0]
assert (
output_str ==
"</s><s>This sentence is made of three parts. One part is about animals, the other part</s>"
)
@slow
def test_cnn_summarization_same_as_fairseq(self):
model = FlaxBartForConditionalGeneration.from_pretrained(
"facebook/bart-large-cnn")
tokenizer = BartTokenizer.from_pretrained("facebook/bart-large-cnn")
FRANCE_ARTICLE = ' Marseille, France (CNN)The French prosecutor leading an investigation into the crash of Germanwings Flight 9525 insisted Wednesday that he was not aware of any video footage from on board the plane. Marseille prosecutor Brice Robin told CNN that "so far no videos were used in the crash investigation." He added, "A person who has such a video needs to immediately give it to the investigators." Robin\'s comments follow claims by two magazines, German daily Bild and French Paris Match, of a cell phone video showing the harrowing final seconds from on board Germanwings Flight 9525 as it crashed into the French Alps. All 150 on board were killed. Paris Match and Bild reported that the video was recovered from a phone at the wreckage site. The two publications described the supposed video, but did not post it on their websites. The publications said that they watched the video, which was found by a source close to the investigation. "One can hear cries of \'My God\' in several languages," Paris Match reported. "Metallic banging can also be heard more than three times, perhaps of the pilot trying to open the cockpit door with a heavy object. Towards the end, after a heavy shake, stronger than the others, the screaming intensifies. Then nothing." "It is a very disturbing scene," said Julian Reichelt, editor-in-chief of Bild online. An official with France\'s accident investigation agency, the BEA, said the agency is not aware of any such video. Lt. Col. Jean-Marc Menichini, a French Gendarmerie spokesman in charge of communications on rescue efforts around the Germanwings crash site, told CNN that the reports were "completely wrong" and "unwarranted." Cell phones have been collected at the site, he said, but that they "hadn\'t been exploited yet." Menichini said he believed the cell phones would need to be sent to the Criminal Research Institute in Rosny sous-Bois, near Paris, in order to be analyzed by specialized technicians working hand-in-hand with investigators. But none of the cell phones found so far have been sent to the institute, Menichini said. Asked whether staff involved in the search could have leaked a memory card to the media, Menichini answered with a categorical "no." Reichelt told "Erin Burnett: Outfront" that he had watched the video and stood by the report, saying Bild and Paris Match are "very confident" that the clip is real. He noted that investigators only revealed they\'d recovered cell phones from the crash site after Bild and Paris Match published their reports. "That is something we did not know before. ... Overall we can say many things of the investigation weren\'t revealed by the investigation at the beginning," he said. What was mental state of Germanwings co-pilot? German airline Lufthansa confirmed Tuesday that co-pilot Andreas Lubitz had battled depression years before he took the controls of Germanwings Flight 9525, which he\'s accused of deliberately crashing last week in the French Alps. Lubitz told his Lufthansa flight training school in 2009 that he had a "previous episode of severe depression," the airline said Tuesday. Email correspondence between Lubitz and the school discovered in an internal investigation, Lufthansa said, included medical documents he submitted in connection with resuming his flight training. The announcement indicates that Lufthansa, the parent company of Germanwings, knew of Lubitz\'s battle with depression, allowed him to continue training and ultimately put him in the cockpit. Lufthansa, whose CEO Carsten Spohr previously said Lubitz was 100% fit to fly, described its statement Tuesday as a "swift and seamless clarification" and said it was sharing the information and documents -- including training and medical records -- with public prosecutors. Spohr traveled to the crash site Wednesday, where recovery teams have been working for the past week to recover human remains and plane debris scattered across a steep mountainside. He saw the crisis center set up in Seyne-les-Alpes, laid a wreath in the village of Le Vernet, closer to the crash site, where grieving families have left flowers at a simple stone memorial. Menichini told CNN late Tuesday that no visible human remains were left at the site but recovery teams would keep searching. French President Francois Hollande, speaking Tuesday, said that it should be possible to identify all the victims using DNA analysis by the end of the week, sooner than authorities had previously suggested. In the meantime, the recovery of the victims\' personal belongings will start Wednesday, Menichini said. Among those personal belongings could be more cell phones belonging to the 144 passengers and six crew on board. Check out the latest from our correspondents . The details about Lubitz\'s correspondence with the flight school during his training were among several developments as investigators continued to delve into what caused the crash and Lubitz\'s possible motive for downing the jet. A Lufthansa spokesperson told CNN on Tuesday that Lubitz had a valid medical certificate, had passed all his examinations and "held all the licenses required." Earlier, a spokesman for the prosecutor\'s office in Dusseldorf, Christoph Kumpa, said medical records reveal Lubitz suffered from suicidal tendencies at some point before his aviation career and underwent psychotherapy before he got his pilot\'s license. Kumpa emphasized there\'s no evidence suggesting Lubitz was suicidal or acting aggressively before the crash. Investigators are looking into whether Lubitz feared his medical condition would cause him to lose his pilot\'s license, a European government official briefed on the investigation told CNN on Tuesday. While flying was "a big part of his life," the source said, it\'s only one theory being considered. Another source, a law enforcement official briefed on the investigation, also told CNN that authorities believe the primary motive for Lubitz to bring down the plane was that he feared he would not be allowed to fly because of his medical problems. Lubitz\'s girlfriend told investigators he had seen an eye doctor and a neuropsychologist, both of whom deemed him unfit to work recently and concluded he had psychological issues, the European government official said. But no matter what details emerge about his previous mental health struggles, there\'s more to the story, said Brian Russell, a forensic psychologist. "Psychology can explain why somebody would turn rage inward on themselves about the fact that maybe they weren\'t going to keep doing their job and they\'re upset about that and so they\'re suicidal," he said. "But there is no mental illness that explains why somebody then feels entitled to also take that rage and turn it outward on 149 other people who had nothing to do with the person\'s problems." Germanwings crash compensation: What we know . Who was the captain of Germanwings Flight 9525? CNN\'s Margot Haddad reported from Marseille and Pamela Brown from Dusseldorf, while Laura Smith-Spark wrote from London. CNN\'s Frederik Pleitgen, Pamela Boykoff, Antonia Mortensen, Sandrine Amiel and Anna-Maja Rappard contributed to this report.' # @noq
SHORTER_ARTICLE = ' (CNN)The Palestinian Authority officially became the 123rd member of the International Criminal Court on Wednesday, a step that gives the court jurisdiction over alleged crimes in Palestinian territories. The formal accession was marked with a ceremony at The Hague, in the Netherlands, where the court is based. The Palestinians signed the ICC\'s founding Rome Statute in January, when they also accepted its jurisdiction over alleged crimes committed "in the occupied Palestinian territory, including East Jerusalem, since June 13, 2014." Later that month, the ICC opened a preliminary examination into the situation in Palestinian territories, paving the way for possible war crimes investigations against Israelis. As members of the court, Palestinians may be subject to counter-charges as well. Israel and the United States, neither of which is an ICC member, opposed the Palestinians\' efforts to join the body. But Palestinian Foreign Minister Riad al-Malki, speaking at Wednesday\'s ceremony, said it was a move toward greater justice. "As Palestine formally becomes a State Party to the Rome Statute today, the world is also a step closer to ending a long era of impunity and injustice," he said, according to an ICC news release. "Indeed, today brings us closer to our shared goals of justice and peace." Judge Kuniko Ozaki, a vice president of the ICC, said acceding to the treaty was just the first step for the Palestinians. "As the Rome Statute today enters into force for the State of Palestine, Palestine acquires all the rights as well as responsibilities that come with being a State Party to the Statute. These are substantive commitments, which cannot be taken lightly," she said. Rights group Human Rights Watch welcomed the development. "Governments seeking to penalize Palestine for joining the ICC should immediately end their pressure, and countries that support universal acceptance of the court\'s treaty should speak out to welcome its membership," said Balkees Jarrah, international justice counsel for the group. "What\'s objectionable is the attempts to undermine international justice, not Palestine\'s decision to join a treaty to which over 100 countries around the world are members." In January, when the preliminary ICC examination was opened, Israeli Prime Minister Benjamin Netanyahu described it as an outrage, saying the court was overstepping its boundaries. The United States also said it "strongly" disagreed with the court\'s decision. "As we have said repeatedly, we do not believe that Palestine is a state and therefore we do not believe that it is eligible to join the ICC," the State Department said in a statement. It urged the warring sides to resolve their differences through direct negotiations. "We will continue to oppose actions against Israel at the ICC as counterproductive to the cause of peace," it said. But the ICC begs to differ with the definition of a state for its purposes and refers to the territories as "Palestine." While a preliminary examination is not a formal investigation, it allows the court to review evidence and determine whether to investigate suspects on both sides. Prosecutor Fatou Bensouda said her office would "conduct its analysis in full independence and impartiality." The war between Israel and Hamas militants in Gaza last summer left more than 2,000 people dead. The inquiry will include alleged war crimes committed since June. The International Criminal Court was set up in 2002 to prosecute genocide, crimes against humanity and war crimes. CNN\'s Vasco Cotovio, Kareem Khadder and Faith Karimi contributed to this report.'
# The below article tests that we don't add any hypotheses outside of the top n_beams
IRAN_ARTICLE = " (CNN)The United States and its negotiating partners reached a very strong framework agreement with Iran in Lausanne, Switzerland, on Thursday that limits Iran's nuclear program in such a way as to effectively block it from building a nuclear weapon. Expect pushback anyway, if the recent past is any harbinger. Just last month, in an attempt to head off such an agreement, House Speaker John Boehner invited Israeli Prime Minister Benjamin Netanyahu to preemptively blast it before Congress, and 47 senators sent a letter to the Iranian leadership warning them away from a deal. The debate that has already begun since the announcement of the new framework will likely result in more heat than light. It will not be helped by the gathering swirl of dubious assumptions and doubtful assertions. Let us address some of these: . The most misleading assertion, despite universal rejection by experts, is that the negotiations' objective at the outset was the total elimination of any nuclear program in Iran. That is the position of Netanyahu and his acolytes in the U.S. Congress. But that is not and never was the objective. If it had been, there would have been no Iranian team at the negotiating table. Rather, the objective has always been to structure an agreement or series of agreements so that Iran could not covertly develop a nuclear arsenal before the United States and its allies could respond. The new framework has exceeded expectations in achieving that goal. It would reduce Iran's low-enriched uranium stockpile, cut by two-thirds its number of installed centrifuges and implement a rigorous inspection regime. Another dubious assumption of opponents is that the Iranian nuclear program is a covert weapons program. Despite sharp accusations by some in the United States and its allies, Iran denies having such a program, and U.S. intelligence contends that Iran has not yet made the decision to build a nuclear weapon. Iran's continued cooperation with International Atomic Energy Agency inspections is further evidence on this point, and we'll know even more about Iran's program in the coming months and years because of the deal. In fact, the inspections provisions that are part of this agreement are designed to protect against any covert action by the Iranians. What's more, the rhetoric of some members of Congress has implied that the negotiations have been between only the United States and Iran (i.e., the 47 senators' letter warning that a deal might be killed by Congress or a future president). This of course is not the case. The talks were between Iran and the five permanent members of the U.N. Security Council (United States, United Kingdom, France, China and Russia) plus Germany, dubbed the P5+1. While the United States has played a leading role in the effort, it negotiated the terms alongside its partners. If the agreement reached by the P5+1 is rejected by Congress, it could result in an unraveling of the sanctions on Iran and threaten NATO cohesion in other areas. Another questionable assertion is that this agreement contains a sunset clause, after which Iran will be free to do as it pleases. Again, this is not the case. Some of the restrictions on Iran's nuclear activities, such as uranium enrichment, will be eased or eliminated over time, as long as 15 years. But most importantly, the framework agreement includes Iran's ratification of the Additional Protocol, which allows IAEA inspectors expanded access to nuclear sites both declared and nondeclared. This provision will be permanent. It does not sunset. Thus, going forward, if Iran decides to enrich uranium to weapons-grade levels, monitors will be able to detect such a move in a matter of days and alert the U.N. Security Council. Many in Congress have said that the agreement should be a formal treaty requiring the Senate to \"advise and consent.\" But the issue is not suited for a treaty. Treaties impose equivalent obligations on all signatories. For example, the New START treaty limits Russia and the United States to 1,550 deployed strategic warheads. But any agreement with Iran will not be so balanced. The restrictions and obligations in the final framework agreement will be imposed almost exclusively on Iran. The P5+1 are obligated only to ease and eventually remove most but not all economic sanctions, which were imposed as leverage to gain this final deal. Finally some insist that any agreement must address Iranian missile programs, human rights violations or support for Hamas or Hezbollah. As important as these issues are, and they must indeed be addressed, they are unrelated to the most important aim of a nuclear deal: preventing a nuclear Iran. To include them in the negotiations would be a poison pill. This agreement should be judged on its merits and on how it affects the security of our negotiating partners and allies, including Israel. Those judgments should be fact-based, not based on questionable assertions or dubious assumptions."
ARTICLE_SUBWAY = ' New York (CNN)When Liana Barrientos was 23 years old, she got married in Westchester County, New York. A year later, she got married again in Westchester County, but to a different man and without divorcing her first husband. Only 18 days after that marriage, she got hitched yet again. Then, Barrientos declared "I do" five more times, sometimes only within two weeks of each other. In 2010, she married once more, this time in the Bronx. In an application for a marriage license, she stated it was her "first and only" marriage. Barrientos, now 39, is facing two criminal counts of "offering a false instrument for filing in the first degree," referring to her false statements on the 2010 marriage license application, according to court documents. Prosecutors said the marriages were part of an immigration scam. On Friday, she pleaded not guilty at State Supreme Court in the Bronx, according to her attorney, Christopher Wright, who declined to comment further. After leaving court, Barrientos was arrested and charged with theft of service and criminal trespass for allegedly sneaking into the New York subway through an emergency exit, said Detective Annette Markowski, a police spokeswoman. In total, Barrientos has been married 10 times, with nine of her marriages occurring between 1999 and 2002. All occurred either in Westchester County, Long Island, New Jersey or the Bronx. She is believed to still be married to four men, and at one time, she was married to eight men at once, prosecutors say. Prosecutors said the immigration scam involved some of her husbands, who filed for permanent residence status shortly after the marriages. Any divorces happened only after such filings were approved. It was unclear whether any of the men will be prosecuted. The case was referred to the Bronx District Attorney\'s Office by Immigration and Customs Enforcement and the Department of Homeland Security\'s Investigation Division. Seven of the men are from so-called "red-flagged" countries, including Egypt, Turkey, Georgia, Pakistan and Mali. Her eighth husband, Rashid Rajput, was deported in 2006 to his native Pakistan after an investigation by the Joint Terrorism Task Force. If convicted, Barrientos faces up to four years in prison. Her next court appearance is scheduled for May 18.'
dct = tokenizer.batch_encode_plus(
[FRANCE_ARTICLE, SHORTER_ARTICLE, IRAN_ARTICLE, ARTICLE_SUBWAY],
max_length=1024,
padding="max_length",
truncation_strategy="only_first",
truncation=True,
return_tensors="jax",
)
self.assertEqual(1024, dct["input_ids"].shape[1])
hypotheses_batch = model.generate(
input_ids=dct["input_ids"],
attention_mask=dct["attention_mask"],
num_beams=2,
).sequences
assert (hypotheses_batch[:, 1] == 0).all().item()
EXPECTED = [
"A French prosecutor says he is not aware of any video footage from on board the plane. Two German magazines claim to have found a cell phone video showing the crash. The publications say they watched the video, which was found by a source close to the investigation. All 150 on board the Germanwings flight were killed.",
"Palestinian Authority becomes 123rd member of the International Criminal Court. The move gives the court jurisdiction over alleged crimes in Palestinian territories. Israel and the United States opposed the Palestinians' efforts to join the body. But Palestinian Foreign Minister Riad al-Malki said it was a move toward greater justice.",
"U.S. and its negotiating partners reached a strong framework agreement with Iran. Peter Bergen: The debate that has already begun will likely result in more heat than light. Bergen: The most misleading assertion is that the negotiations' objective at the outset was the total elimination of any nuclear program.",
"Liana Barrientos, 39, has been married 10 times, sometimes within two weeks of each other. Prosecutors say the marriages were part of an immigration scam. She pleaded not guilty at State Supreme Court in the Bronx on Friday. If convicted, Barrientos faces up to four years in prison.",
]
generated_summaries = tokenizer.batch_decode(
hypotheses_batch.tolist(),
clean_up_tokenization_spaces=True,
skip_special_tokens=True)
assert generated_summaries == EXPECTED
class FlaxGPT2ModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = (FlaxGPT2Model,
FlaxGPT2LMHeadModel) if is_flax_available() else ()
def setUp(self):
self.model_tester = FlaxGPT2ModelTester(self)
def test_use_cache_forward(self):
for model_class_name in self.all_model_classes:
config, input_ids, attention_mask = self.model_tester.prepare_config_and_inputs(
)
self.model_tester.check_use_cache_forward(model_class_name, config,
input_ids,
attention_mask)
def test_use_cache_forward_with_attn_mask(self):
for model_class_name in self.all_model_classes:
config, input_ids, attention_mask = self.model_tester.prepare_config_and_inputs(
)
self.model_tester.check_use_cache_forward_with_attn_mask(
model_class_name, config, input_ids, attention_mask)
def test_use_cache_generation(self):
config, input_ids, _ = self.model_tester.prepare_config_and_inputs()
self.model_tester.check_use_cache_generation(config, input_ids)
# overwrite from common since `attention_mask` in combination
# with `causal_mask` behaves slighly differently
@is_pt_flax_cross_test
def test_equivalence_pt_to_flax(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
# prepare inputs
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
pt_inputs = {
k: torch.tensor(v.tolist())
for k, v in prepared_inputs_dict.items()
}
# load corresponding PyTorch class
pt_model_class_name = model_class.__name__[
4:] # Skip the "Flax" at the beginning
pt_model_class = getattr(transformers, pt_model_class_name)
batch_size, seq_length = pt_inputs["input_ids"].shape
rnd_start_indices = np.random.randint(0,
seq_length - 1,
size=(batch_size, ))
for batch_idx, start_index in enumerate(rnd_start_indices):
pt_inputs["attention_mask"][batch_idx, :start_index] = 0
pt_inputs["attention_mask"][batch_idx, start_index:] = 1
prepared_inputs_dict["attention_mask"][
batch_idx, :start_index] = 0
prepared_inputs_dict["attention_mask"][batch_idx,
start_index:] = 1
pt_model = pt_model_class(config).eval()
fx_model = model_class(config, dtype=jnp.float32)
fx_state = convert_pytorch_state_dict_to_flax(
pt_model.state_dict(), fx_model)
fx_model.params = fx_state
with torch.no_grad():
pt_outputs = pt_model(**pt_inputs).to_tuple()
fx_outputs = fx_model(**prepared_inputs_dict).to_tuple()
self.assertEqual(
len(fx_outputs), len(pt_outputs),
"Output lengths differ between Flax and PyTorch")
for fx_output, pt_output in zip(fx_outputs, pt_outputs):
self.assert_almost_equals(fx_output[:, -1],
pt_output[:, -1].numpy(), 4e-2)
with tempfile.TemporaryDirectory() as tmpdirname:
pt_model.save_pretrained(tmpdirname)
fx_model_loaded = model_class.from_pretrained(tmpdirname,
from_pt=True)
fx_outputs_loaded = fx_model_loaded(
**prepared_inputs_dict).to_tuple()
self.assertEqual(
len(fx_outputs_loaded), len(pt_outputs),
"Output lengths differ between Flax and PyTorch")
for fx_output_loaded, pt_output in zip(fx_outputs_loaded,
pt_outputs):
self.assert_almost_equals(fx_output_loaded[:, -1],
pt_output[:, -1].numpy(), 4e-2)
# overwrite from common since `attention_mask` in combination
# with `causal_mask` behaves slighly differently
@is_pt_flax_cross_test
def test_equivalence_flax_to_pt(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
# prepare inputs
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
pt_inputs = {
k: torch.tensor(v.tolist())
for k, v in prepared_inputs_dict.items()
}
# load corresponding PyTorch class
pt_model_class_name = model_class.__name__[
4:] # Skip the "Flax" at the beginning
pt_model_class = getattr(transformers, pt_model_class_name)
pt_model = pt_model_class(config).eval()
fx_model = model_class(config, dtype=jnp.float32)
pt_model = load_flax_weights_in_pytorch_model(
pt_model, fx_model.params)
batch_size, seq_length = pt_inputs["input_ids"].shape
rnd_start_indices = np.random.randint(0,
seq_length - 1,
size=(batch_size, ))
for batch_idx, start_index in enumerate(rnd_start_indices):
pt_inputs["attention_mask"][batch_idx, :start_index] = 0
pt_inputs["attention_mask"][batch_idx, start_index:] = 1
prepared_inputs_dict["attention_mask"][
batch_idx, :start_index] = 0
prepared_inputs_dict["attention_mask"][batch_idx,
start_index:] = 1
# make sure weights are tied in PyTorch
pt_model.tie_weights()
with torch.no_grad():
pt_outputs = pt_model(**pt_inputs).to_tuple()
fx_outputs = fx_model(**prepared_inputs_dict).to_tuple()
self.assertEqual(
len(fx_outputs), len(pt_outputs),
"Output lengths differ between Flax and PyTorch")
for fx_output, pt_output in zip(fx_outputs, pt_outputs):
self.assert_almost_equals(fx_output[:, -1],
pt_output[:, -1].numpy(), 4e-2)
with tempfile.TemporaryDirectory() as tmpdirname:
fx_model.save_pretrained(tmpdirname)
pt_model_loaded = pt_model_class.from_pretrained(
tmpdirname, from_flax=True)
with torch.no_grad():
pt_outputs_loaded = pt_model_loaded(**pt_inputs).to_tuple()
self.assertEqual(
len(fx_outputs), len(pt_outputs_loaded),
"Output lengths differ between Flax and PyTorch")
for fx_output, pt_output in zip(fx_outputs, pt_outputs_loaded):
self.assert_almost_equals(fx_output[:, -1],
pt_output[:, -1].numpy(), 4e-2)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained("gpt2", from_pt=True)
outputs = model(np.ones((1, 1)))
self.assertIsNotNone(outputs)
class FlaxCLIPVisionModelTest(FlaxModelTesterMixin, unittest.TestCase):
"""
Here we also overwrite some of the tests of test_modeling_common.py, as CLIP does not use input_ids, inputs_embeds,
attention_mask and seq_length.
"""
all_model_classes = (FlaxCLIPVisionModel, ) if is_flax_available() else ()
def setUp(self):
self.model_tester = FlaxCLIPVisionModelTester(self)
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.__call__)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["pixel_values"]
self.assertListEqual(arg_names[:1], expected_arg_names)
def test_jit_compilation(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
model = model_class(config)
@jax.jit
def model_jitted(pixel_values, **kwargs):
return model(pixel_values=pixel_values,
**kwargs).to_tuple()
with self.subTest("JIT Enabled"):
jitted_outputs = model_jitted(**prepared_inputs_dict)
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = model_jitted(**prepared_inputs_dict)
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
def test_hidden_states_output(self):
def check_hidden_states_output(inputs_dict, config, model_class):
model = model_class(config)
outputs = model(
**self._prepare_for_class(inputs_dict, model_class))
hidden_states = outputs.hidden_states
self.assertEqual(len(hidden_states),
self.model_tester.num_hidden_layers + 1)
# CLIP has a different seq_length
image_size = (self.model_tester.image_size,
self.model_tester.image_size)
patch_size = (self.model_tester.patch_size,
self.model_tester.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] //
patch_size[0])
seq_length = num_patches + 1
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
[seq_length, self.model_tester.hidden_size],
)
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
inputs_dict["output_hidden_states"] = True
check_hidden_states_output(inputs_dict, config, model_class)
# check that output_hidden_states also work using config
del inputs_dict["output_hidden_states"]
config.output_hidden_states = True
check_hidden_states_output(inputs_dict, config, model_class)
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
config.return_dict = True
# in CLIP, the seq_len equals the number of patches + 1 (we add 1 for the [CLS] token)
image_size = (self.model_tester.image_size,
self.model_tester.image_size)
patch_size = (self.model_tester.patch_size,
self.model_tester.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] //
patch_size[0])
seq_length = num_patches + 1
for model_class in self.all_model_classes:
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = False
model = model_class(config)
outputs = model(
**self._prepare_for_class(inputs_dict, model_class))
attentions = 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)
outputs = model(
**self._prepare_for_class(inputs_dict, model_class))
attentions = outputs.attentions
self.assertEqual(len(attentions),
self.model_tester.num_hidden_layers)
self.assertListEqual(
list(attentions[0].shape[-3:]),
[
self.model_tester.num_attention_heads, seq_length,
seq_length
],
)
out_len = len(outputs)
# Check attention is always last and order is fine
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = True
model = model_class(config)
outputs = model(
**self._prepare_for_class(inputs_dict, model_class))
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)
self.assertListEqual(
list(self_attentions[0].shape[-3:]),
[
self.model_tester.num_attention_heads, seq_length,
seq_length
],
)
# FlaxCLIPVisionModel does not have any base model
def test_save_load_from_base(self):
pass
def test_save_load_to_base(self):
pass
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained(
"openai/clip-vit-base-patch32", from_pt=True)
outputs = model(np.ones((1, 3, 224, 224)))
self.assertIsNotNone(outputs)
class FlaxCLIPModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = (FlaxCLIPModel, ) if is_flax_available() else ()
test_attention_outputs = False
def setUp(self):
self.model_tester = FlaxCLIPModelTester(self)
# hidden_states are tested in individual model tests
def test_hidden_states_output(self):
pass
@slow
def test_jit_compilation(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
model = model_class(config)
@jax.jit
def model_jitted(input_ids, pixel_values, **kwargs):
return model(input_ids=input_ids,
pixel_values=pixel_values,
**kwargs).to_tuple()
with self.subTest("JIT Enabled"):
jitted_outputs = model_jitted(**prepared_inputs_dict)
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = model_jitted(**prepared_inputs_dict)
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs[:4],
outputs[:4]):
self.assertEqual(jitted_output.shape, output.shape)
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.__call__)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = [
"input_ids", "pixel_values", "attention_mask", "position_ids"
]
self.assertListEqual(arg_names[:4], expected_arg_names)
def test_get_image_features(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
model = FlaxCLIPModel(config)
@jax.jit
def model_jitted(pixel_values):
return model.get_image_features(pixel_values=pixel_values)
with self.subTest("JIT Enabled"):
jitted_output = model_jitted(inputs_dict["pixel_values"])
with self.subTest("JIT Disabled"):
with jax.disable_jit():
output = model_jitted(inputs_dict["pixel_values"])
self.assertEqual(jitted_output.shape, output.shape)
self.assertTrue(np.allclose(jitted_output, output, atol=1e-3))
def test_get_text_features(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
model = FlaxCLIPModel(config)
@jax.jit
def model_jitted(input_ids, attention_mask, **kwargs):
return model.get_text_features(input_ids=input_ids,
attention_mask=attention_mask)
with self.subTest("JIT Enabled"):
jitted_output = model_jitted(**inputs_dict)
with self.subTest("JIT Disabled"):
with jax.disable_jit():
output = model_jitted(**inputs_dict)
self.assertEqual(jitted_output.shape, output.shape)
self.assertTrue(np.allclose(jitted_output, output, atol=1e-3))
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained(
"openai/clip-vit-base-patch32", from_pt=True)
outputs = model(input_ids=np.ones((1, 1)),
pixel_values=np.ones((1, 3, 224, 224)))
self.assertIsNotNone(outputs)
# overwrite from common since FlaxCLIPModel returns nested output
# which is not supported in the common test
@is_pt_flax_cross_test
def test_equivalence_pt_to_flax(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
# prepare inputs
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
pt_inputs = {
k: torch.tensor(v.tolist())
for k, v in prepared_inputs_dict.items()
}
# load corresponding PyTorch class
pt_model_class_name = model_class.__name__[
4:] # Skip the "Flax" at the beginning
pt_model_class = getattr(transformers, pt_model_class_name)
pt_model = pt_model_class(config).eval()
fx_model = model_class(config, dtype=jnp.float32)
fx_state = convert_pytorch_state_dict_to_flax(
pt_model.state_dict(), fx_model)
fx_model.params = fx_state
with torch.no_grad():
pt_outputs = pt_model(**pt_inputs).to_tuple()
# PyTorch CLIPModel returns loss, we skip it here as we don't return loss in JAX/Flax models
pt_outputs = pt_outputs[1:]
fx_outputs = fx_model(**prepared_inputs_dict).to_tuple()
self.assertEqual(
len(fx_outputs), len(pt_outputs),
"Output lengths differ between Flax and PyTorch")
for fx_output, pt_output in zip(fx_outputs[:4],
pt_outputs[:4]):
self.assert_almost_equals(fx_output, pt_output.numpy(),
4e-2)
with tempfile.TemporaryDirectory() as tmpdirname:
pt_model.save_pretrained(tmpdirname)
fx_model_loaded = model_class.from_pretrained(tmpdirname,
from_pt=True)
fx_outputs_loaded = fx_model_loaded(
**prepared_inputs_dict).to_tuple()
self.assertEqual(
len(fx_outputs_loaded), len(pt_outputs),
"Output lengths differ between Flax and PyTorch")
for fx_output_loaded, pt_output in zip(fx_outputs_loaded[:4],
pt_outputs[:4]):
self.assert_almost_equals(fx_output_loaded,
pt_output.numpy(), 4e-2)
# overwrite from common since FlaxCLIPModel returns nested output
# which is not supported in the common test
@is_pt_flax_cross_test
def test_equivalence_flax_to_pt(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
# prepare inputs
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
pt_inputs = {
k: torch.tensor(v.tolist())
for k, v in prepared_inputs_dict.items()
}
# load corresponding PyTorch class
pt_model_class_name = model_class.__name__[
4:] # Skip the "Flax" at the beginning
pt_model_class = getattr(transformers, pt_model_class_name)
pt_model = pt_model_class(config).eval()
fx_model = model_class(config, dtype=jnp.float32)
pt_model = load_flax_weights_in_pytorch_model(
pt_model, fx_model.params)
# make sure weights are tied in PyTorch
pt_model.tie_weights()
with torch.no_grad():
pt_outputs = pt_model(**pt_inputs).to_tuple()
# PyTorch CLIPModel returns loss, we skip it here as we don't return loss in JAX/Flax models
pt_outputs = pt_outputs[1:]
fx_outputs = fx_model(**prepared_inputs_dict).to_tuple()
self.assertEqual(
len(fx_outputs), len(pt_outputs),
"Output lengths differ between Flax and PyTorch")
for fx_output, pt_output in zip(fx_outputs[:4],
pt_outputs[:4]):
self.assert_almost_equals(fx_output, pt_output.numpy(),
4e-2)
with tempfile.TemporaryDirectory() as tmpdirname:
fx_model.save_pretrained(tmpdirname)
pt_model_loaded = pt_model_class.from_pretrained(
tmpdirname, from_flax=True)
with torch.no_grad():
pt_outputs_loaded = pt_model_loaded(**pt_inputs).to_tuple()
pt_outputs_loaded = pt_outputs_loaded[1:]
self.assertEqual(
len(fx_outputs), len(pt_outputs_loaded),
"Output lengths differ between Flax and PyTorch")
for fx_output, pt_output in zip(fx_outputs[:4],
pt_outputs_loaded[:4]):
self.assert_almost_equals(fx_output, pt_output.numpy(),
4e-2)
# overwrite from common since FlaxCLIPModel returns nested output
# which is not supported in the common test
def test_from_pretrained_save_pretrained(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
if model_class.__name__ != "FlaxBertModel":
continue
with self.subTest(model_class.__name__):
model = model_class(config)
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
outputs = model(**prepared_inputs_dict).to_tuple()
# verify that normal save_pretrained works as expected
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
model_loaded = model_class.from_pretrained(tmpdirname)
outputs_loaded = model_loaded(
**prepared_inputs_dict).to_tuple()[:4]
for output_loaded, output in zip(outputs_loaded, outputs):
self.assert_almost_equals(output_loaded, output, 1e-3)
# verify that save_pretrained for distributed training
# with `params=params` works as expected
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname, params=model.params)
model_loaded = model_class.from_pretrained(tmpdirname)
outputs_loaded = model_loaded(
**prepared_inputs_dict).to_tuple()[:4]
for output_loaded, output in zip(outputs_loaded, outputs):
self.assert_almost_equals(output_loaded, output, 1e-3)
class FlaxBlenderbotModelTest(FlaxModelTesterMixin, unittest.TestCase,
FlaxGenerationTesterMixin):
is_encoder_decoder = True
all_model_classes = ((
FlaxBlenderbotModel,
FlaxBlenderbotForConditionalGeneration,
) if is_flax_available() else ())
all_generative_model_classes = (FlaxBlenderbotForConditionalGeneration,
) if is_flax_available() else ()
def setUp(self):
self.model_tester = FlaxBlenderbotModelTester(self)
def test_use_cache_forward(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
for model_class in self.all_model_classes:
self.model_tester.check_use_cache_forward(model_class, config,
inputs_dict)
def test_use_cache_forward_with_attn_mask(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
for model_class in self.all_model_classes:
self.model_tester.check_use_cache_forward_with_attn_mask(
model_class, config, inputs_dict)
def test_encode(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
model = model_class(config)
@jax.jit
def encode_jitted(input_ids, attention_mask=None, **kwargs):
return model.encode(input_ids=input_ids,
attention_mask=attention_mask)
with self.subTest("JIT Enabled"):
jitted_outputs = encode_jitted(
**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = encode_jitted(
**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
def test_decode(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
model = model_class(config)
encoder_outputs = model.encode(inputs_dict["input_ids"],
inputs_dict["attention_mask"])
prepared_inputs_dict = {
"decoder_input_ids": inputs_dict["decoder_input_ids"],
"decoder_attention_mask":
inputs_dict["decoder_attention_mask"],
"encoder_outputs": encoder_outputs,
}
@jax.jit
def decode_jitted(decoder_input_ids, decoder_attention_mask,
encoder_outputs):
return model.decode(
decoder_input_ids=decoder_input_ids,
decoder_attention_mask=decoder_attention_mask,
encoder_outputs=encoder_outputs,
)
with self.subTest("JIT Enabled"):
jitted_outputs = decode_jitted(
**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = decode_jitted(
**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained(
"facebook/blenderbot-400M-distill")
# FlaxBlenderbotForSequenceClassification expects eos token in input_ids
input_ids = np.ones((1, 1)) * model.config.eos_token_id
outputs = model(input_ids)
self.assertIsNotNone(outputs)
@unittest.skipUnless(jax_device != "cpu", "3B test too slow on CPU.")
@slow
def test_generation_from_short_input_same_as_parlai_3B(self):
FASTER_GEN_KWARGS = dict(num_beams=1,
early_stopping=True,
min_length=15,
max_length=25)
TOK_DECODE_KW = dict(skip_special_tokens=True,
clean_up_tokenization_spaces=True)
model = FlaxBlenderbotForConditionalGeneration.from_pretrained(
"facebook/blenderbot-3B", from_pt=True)
tokenizer = BlenderbotTokenizer.from_pretrained(
"facebook/blenderbot-3B")
src_text = ["Sam"]
model_inputs = tokenizer(src_text, return_tensors="jax")
generated_utterances = model.generate(**model_inputs,
**FASTER_GEN_KWARGS)
tgt_text = 'Sam is a great name. It means "sun" in Gaelic.'
generated_txt = tokenizer.batch_decode(generated_utterances,
**TOK_DECODE_KW)
assert generated_txt[0].strip() == tgt_text
class FlaxBigBirdModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = ((
FlaxBigBirdModel,
FlaxBigBirdForPreTraining,
FlaxBigBirdForMaskedLM,
FlaxBigBirdForMultipleChoice,
FlaxBigBirdForQuestionAnswering,
FlaxBigBirdForSequenceClassification,
FlaxBigBirdForTokenClassification,
) if is_flax_available() else ())
test_attn_probs = False
test_mismatched_shapes = False
def setUp(self):
self.model_tester = FlaxBigBirdModelTester(self)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained(
"google/bigbird-roberta-base", from_pt=True)
outputs = model(np.ones((1, 1)))
self.assertIsNotNone(outputs)
def test_attention_outputs(self):
if self.test_attn_probs:
super().test_attention_outputs()
@slow
# copied from `test_modeling_flax_common` because it takes much longer than other models
def test_jit_compilation(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
model = model_class(config)
@jax.jit
def model_jitted(input_ids, attention_mask=None, **kwargs):
return model(input_ids=input_ids,
attention_mask=attention_mask,
**kwargs)
with self.subTest("JIT Enabled"):
jitted_outputs = model_jitted(
**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = model_jitted(
**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# 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
from transformers import BertConfig, is_flax_available
from transformers.testing_utils import require_flax, slow
from .test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor, random_attention_mask
if is_flax_available():
from transformers.models.bert.modeling_flax_bert import (
FlaxBertForMaskedLM,
FlaxBertForMultipleChoice,
FlaxBertForNextSentencePrediction,
FlaxBertForPreTraining,
FlaxBertForQuestionAnswering,
FlaxBertForSequenceClassification,
FlaxBertForTokenClassification,
FlaxBertModel,
)
class FlaxBertModelTester(unittest.TestCase):
def __init__(
self,
class FlaxWav2Vec2ModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = ((FlaxWav2Vec2Model, FlaxWav2Vec2ForCTC,
FlaxWav2Vec2ForPreTraining)
if is_flax_available() else ())
def setUp(self):
self.model_tester = FlaxWav2Vec2ModelTester(self)
def test_train(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
input_values = inputs_dict["input_values"]
attention_mask = inputs_dict["attention_mask"]
model = FlaxWav2Vec2ForPreTraining(config)
features_shape = (
input_values.shape[0],
model._get_feat_extract_output_lengths(
np.array(input_values.shape[1])),
)
batch_size, sequence_length = features_shape[:2]
mask_prob = 0.5
mask_length = 4
mask_time_indices = _compute_mask_indices(
(batch_size, sequence_length), mask_prob, mask_length)
dropout_rng, gumbel_rng = jax.random.split(jax.random.PRNGKey(0))
output = model(
input_values,
attention_mask=attention_mask,
mask_time_indices=mask_time_indices,
train=True,
dropout_rng=dropout_rng,
gumbel_rng=gumbel_rng,
)[0]
self.assertTrue(output.shape == (batch_size, sequence_length,
model.config.proj_codevector_dim))
# overwrite because of `input_values`
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.__call__)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["input_values", "attention_mask"]
self.assertListEqual(arg_names[:2], expected_arg_names)
# overwrite because of `input_values`
def test_jit_compilation(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
model = model_class(config)
@jax.jit
def model_jitted(input_values, attention_mask=None, **kwargs):
return model(input_values=input_values,
attention_mask=attention_mask,
**kwargs)
with self.subTest("JIT Enabled"):
jitted_outputs = model_jitted(
**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = model_jitted(
**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
def test_freeze_feature_encoder(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
input_values = inputs_dict["input_values"]
attention_mask = inputs_dict["attention_mask"]
model = FlaxWav2Vec2ForPreTraining(config)
params = model.params
# dummy loss function
def compute_loss(params,
input_values,
attention_mask,
freeze_feature_encoder: bool = False,
epsilon: float = 1e-8):
outputs = model(
input_values,
attention_mask=attention_mask,
freeze_feature_encoder=freeze_feature_encoder,
params=params,
)
# compute cosine similarity of projected and projected_quantized states
cosine_sim = optax.cosine_similarity(
outputs.projected_states,
outputs.projected_quantized_states,
epsilon=epsilon)
loss = cosine_sim.sum()
return loss, outputs.to_tuple()
# transform the loss function to get the gradients
grad_fn = jax.value_and_grad(compute_loss, has_aux=True)
# compute loss, outputs and gradients for unfrozen model
(loss, outputs), grads = grad_fn(params,
input_values,
attention_mask,
freeze_feature_encoder=False)
# compare to loss, outputs and gradients for frozen model
(loss_frozen,
outputs_frozen), grads_frozen = grad_fn(params,
input_values,
attention_mask,
freeze_feature_encoder=True)
# ensure that the outputs and losses remain precisely equal
for output, output_frozen in zip(outputs, outputs_frozen):
self.assertTrue((output == output_frozen).all())
self.assertEqual(loss, loss_frozen)
grads = flatten_dict(grads)
grads_frozen = flatten_dict(grads_frozen)
# ensure that the dicts of gradients contain the same keys
self.assertEqual(grads.keys(), grads_frozen.keys())
# ensure that the gradients of the feature extractor layers are precisely zero when frozen and contain non-zero entries when unfrozen
feature_extractor_grads = tuple(grads[k] for k in grads
if "feature_extractor" in k)
feature_extractor_grads_frozen = tuple(grads_frozen[k]
for k in grads_frozen
if "feature_extractor" in k)
for feature_extractor_grad, feature_extractor_grad_frozen in zip(
feature_extractor_grads, feature_extractor_grads_frozen):
self.assertTrue((feature_extractor_grad_frozen == 0.0).all())
self.assertTrue((feature_extractor_grad > 0.0).any())
# ensure that the gradients of all unfrozen layers remain equal, i.e. all layers excluding the frozen 'feature_extractor'
grads = tuple(grads[k] for k in grads if "feature_extractor" not in k)
grads_frozen = tuple(grads_frozen[k] for k in grads_frozen
if "feature_extractor" not in k)
for grad, grad_frozen in zip(grads, grads_frozen):
self.assertTrue((grad == grad_frozen).all())
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained(
"facebook/wav2vec2-large-960h-lv60-self", from_pt=True)
outputs = model(np.ones((1, 1024), dtype="f4"))
self.assertIsNotNone(outputs)
class FlaxWav2Vec2ModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = (
(FlaxWav2Vec2Model, FlaxWav2Vec2ForCTC, FlaxWav2Vec2ForPreTraining) if is_flax_available() else ()
)
def setUp(self):
self.model_tester = FlaxWav2Vec2ModelTester(self)
def test_train(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
input_values = inputs_dict["input_values"]
attention_mask = inputs_dict["attention_mask"]
model = FlaxWav2Vec2ForPreTraining(config)
features_shape = (
input_values.shape[0],
model._get_feat_extract_output_lengths(np.array(input_values.shape[1])),
)
batch_size, sequence_length = features_shape[:2]
mask_prob = 0.5
mask_length = 4
mask_time_indices = _compute_mask_indices((batch_size, sequence_length), mask_prob, mask_length)
dropout_rng, gumbel_rng = jax.random.split(jax.random.PRNGKey(0))
output = model(
input_values,
attention_mask=attention_mask,
mask_time_indices=mask_time_indices,
train=True,
dropout_rng=dropout_rng,
gumbel_rng=gumbel_rng,
)[0]
self.assertTrue(output.shape == (batch_size, sequence_length, model.config.proj_codevector_dim))
# overwrite because of `input_values`
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.__call__)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["input_values", "attention_mask"]
self.assertListEqual(arg_names[:2], expected_arg_names)
@slow
# overwrite because of `input_values`
def test_jit_compilation(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
model = model_class(config)
@jax.jit
def model_jitted(input_values, attention_mask=None, **kwargs):
return model(input_values=input_values, attention_mask=attention_mask, **kwargs)
with self.subTest("JIT Enabled"):
jitted_outputs = model_jitted(**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = model_jitted(**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained("facebook/wav2vec2-large-960h-lv60-self", from_pt=True)
outputs = model(np.ones((1, 1024), dtype="f4"))
self.assertIsNotNone(outputs)
class FlaxLongT5ModelTest(FlaxModelTesterMixin, FlaxGenerationTesterMixin,
unittest.TestCase):
all_model_classes = (
FlaxLongT5Model,
FlaxLongT5ForConditionalGeneration) if is_flax_available() else ()
all_generative_model_classes = (
FlaxLongT5ForConditionalGeneration, ) if is_flax_available() else ()
is_encoder_decoder = True
def setUp(self):
self.model_tester = FlaxLongT5ModelTester(self)
self.config_tester = ConfigTester(self,
config_class=LongT5Config,
d_model=37)
def test_config(self):
self.config_tester.run_common_tests()
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_model_v1_1(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
# check that gated gelu feed forward and different word embeddings work
config = config_and_inputs[0]
config.tie_word_embeddings = False
config.feed_forward_proj = "gated-gelu"
self.model_tester.create_and_check_model(config,
*config_and_inputs[1:])
def test_use_cache_forward_with_attn_mask(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
for model_class in self.all_model_classes:
self.model_tester.check_use_cache_forward_with_attn_mask(
model_class, *config_and_inputs)
def test_encode(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
model = model_class(config)
@jax.jit
def encode_jitted(input_ids, attention_mask=None, **kwargs):
return model.encode(input_ids=input_ids,
attention_mask=attention_mask)
with self.subTest("JIT Enabled"):
jitted_outputs = encode_jitted(
**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = encode_jitted(
**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
def test_decode(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
model = model_class(config)
encoder_outputs = model.encode(inputs_dict["input_ids"],
inputs_dict["attention_mask"])
prepared_inputs_dict = {
"decoder_input_ids": inputs_dict["decoder_input_ids"],
"decoder_attention_mask":
inputs_dict["decoder_attention_mask"],
"encoder_outputs": encoder_outputs,
}
@jax.jit
def decode_jitted(decoder_input_ids, decoder_attention_mask,
encoder_outputs):
return model.decode(
decoder_input_ids=decoder_input_ids,
decoder_attention_mask=decoder_attention_mask,
encoder_outputs=encoder_outputs,
)
with self.subTest("JIT Enabled"):
jitted_outputs = decode_jitted(
**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = decode_jitted(
**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
def test_shift_right(self):
decoder_start_token_id = 0
pad_token_id = 1
labels = np.arange(2, 102).reshape(5, 20)
labels[:2, 15:] = -100
decoder_input_ids = shift_tokens_right(labels, pad_token_id,
decoder_start_token_id)
np_decoder_input_ids = np.array(decoder_input_ids)
padded_slice = np_decoder_input_ids[:2, (15 + 1):]
self.assertTrue((padded_slice == 1).all())
not_padded_slice = np_decoder_input_ids[2:, 1:]
rolled_labels = np.roll(labels[2:], 1)[:, 1:]
self.assertTrue((not_padded_slice == rolled_labels).all())
self.assertTrue((np_decoder_input_ids[:, 0] == 0).all())
# overwrite since special base model prefix is used
def test_save_load_from_base(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
base_class = FLAX_MODEL_MAPPING[config.__class__]
for model_class in self.all_model_classes:
if model_class == base_class:
continue
model = base_class(config)
base_params = flatten_dict(unfreeze(model.params))
# check that all base model weights are loaded correctly
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
head_model = model_class.from_pretrained(tmpdirname)
base_param_from_head = flatten_dict(unfreeze(
head_model.params))
for key in base_param_from_head.keys():
max_diff = (base_params[key] -
base_param_from_head[key]).sum().item()
self.assertLessEqual(max_diff,
1e-3,
msg=f"{key} not identical")
# overwrite since special base model prefix is used
def test_save_load_to_base(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
base_class = FLAX_MODEL_MAPPING[config.__class__]
for model_class in self.all_model_classes:
if model_class == base_class:
continue
model = model_class(config)
base_params_from_head = flatten_dict(unfreeze(model.params))
# check that all base model weights are loaded correctly
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
base_model = base_class.from_pretrained(tmpdirname)
base_params = flatten_dict(unfreeze(base_model.params))
for key in base_params_from_head.keys():
max_diff = (base_params[key] -
base_params_from_head[key]).sum().item()
self.assertLessEqual(max_diff,
1e-3,
msg=f"{key} not identical")
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
config.return_dict = True
seq_length = getattr(self.model_tester, "seq_length", None)
decoder_seq_length = getattr(self.model_tester, "decoder_seq_length",
seq_length)
encoder_seq_length = getattr(self.model_tester, "encoder_seq_length",
seq_length)
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)
block_len = getattr(self.model_tester, "block_len", None)
for model_class in self.all_model_classes:
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = False
model = model_class(config)
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)
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)
self.assertListEqual(
list(attentions[0].shape[-3:]),
[
self.model_tester.num_attention_heads, block_len,
3 * block_len
],
)
out_len = len(outputs)
if self.is_encoder_decoder:
correct_outlen = 5
# Question Answering model returns start_logits and end_logits
if model_class in get_values(
FLAX_MODEL_FOR_QUESTION_ANSWERING_MAPPING):
correct_outlen += 1 # start_logits and end_logits instead of only 1 output
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,
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)
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)
self.assertListEqual(
list(self_attentions[0].shape[-3:]),
[
self.model_tester.num_attention_heads, block_len,
3 * block_len
],
)
# overwrite since special base model prefix is used
@is_pt_flax_cross_test
def test_save_load_from_base_pt(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
base_class = FLAX_MODEL_MAPPING[config.__class__]
for model_class in self.all_model_classes:
if model_class == base_class:
continue
model = base_class(config)
base_params = flatten_dict(unfreeze(model.params))
# convert Flax model to PyTorch model
pt_model_class = getattr(
transformers,
base_class.__name__[4:]) # Skip the "Flax" at the beginning
pt_model = pt_model_class(config).eval()
pt_model = load_flax_weights_in_pytorch_model(
pt_model, model.params)
# check that all base model weights are loaded correctly
with tempfile.TemporaryDirectory() as tmpdirname:
# save pt model
pt_model.save_pretrained(tmpdirname)
head_model = model_class.from_pretrained(tmpdirname,
from_pt=True)
base_param_from_head = flatten_dict(unfreeze(
head_model.params))
for key in base_param_from_head.keys():
max_diff = (base_params[key] -
base_param_from_head[key]).sum().item()
self.assertLessEqual(max_diff,
1e-3,
msg=f"{key} not identical")
# overwrite since special base model prefix is used
@is_pt_flax_cross_test
def test_save_load_to_base_pt(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
base_class = FLAX_MODEL_MAPPING[config.__class__]
for model_class in self.all_model_classes:
if model_class == base_class:
continue
model = model_class(config)
base_params_from_head = flatten_dict(unfreeze(model.params))
# convert Flax model to PyTorch model
pt_model_class = getattr(
transformers,
model_class.__name__[4:]) # Skip the "Flax" at the beginning
pt_model = pt_model_class(config).eval()
pt_model = load_flax_weights_in_pytorch_model(
pt_model, model.params)
# check that all base model weights are loaded correctly
with tempfile.TemporaryDirectory() as tmpdirname:
pt_model.save_pretrained(tmpdirname)
base_model = base_class.from_pretrained(tmpdirname,
from_pt=True)
base_params = flatten_dict(unfreeze(base_model.params))
for key in base_params_from_head.keys():
max_diff = (base_params[key] -
base_params_from_head[key]).sum().item()
self.assertLessEqual(max_diff,
1e-3,
msg=f"{key} not identical")
# overwrite since special base model prefix is used
@is_pt_flax_cross_test
def test_save_load_bf16_to_base_pt(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
base_class = FLAX_MODEL_MAPPING[config.__class__]
for model_class in self.all_model_classes:
if model_class == base_class:
continue
model = model_class(config)
model.params = model.to_bf16(model.params)
base_params_from_head = flatten_dict(unfreeze(model.params))
# convert Flax model to PyTorch model
pt_model_class = getattr(
transformers,
model_class.__name__[4:]) # Skip the "Flax" at the beginning
pt_model = pt_model_class(config).eval()
pt_model = load_flax_weights_in_pytorch_model(
pt_model, model.params)
# check that all base model weights are loaded correctly
with tempfile.TemporaryDirectory() as tmpdirname:
pt_model.save_pretrained(tmpdirname)
base_model = base_class.from_pretrained(tmpdirname,
from_pt=True)
base_params = flatten_dict(unfreeze(base_model.params))
for key in base_params_from_head.keys():
max_diff = (base_params[key] -
base_params_from_head[key]).sum().item()
self.assertLessEqual(max_diff,
1e-3,
msg=f"{key} not identical")
class FlaxPegasusModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = (
(
FlaxPegasusForConditionalGeneration,
FlaxPegasusModel,
)
if is_flax_available()
else ()
)
all_generative_model_classes = (FlaxPegasusForConditionalGeneration,) if is_flax_available() else ()
is_encoder_decoder = True
test_pruning = False
test_head_masking = False
test_onnx = False
def setUp(self):
self.model_tester = FlaxPegasusModelTester(self)
self.config_tester = ConfigTester(self, config_class=PegasusConfig)
def test_config(self):
self.config_tester.run_common_tests()
def test_use_cache_forward(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
self.model_tester.check_use_cache_forward(model_class, config, inputs_dict)
def test_use_cache_forward_with_attn_mask(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
self.model_tester.check_use_cache_forward_with_attn_mask(model_class, config, inputs_dict)
def test_encode(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
model = model_class(config)
@jax.jit
def encode_jitted(input_ids, attention_mask=None, **kwargs):
return model.encode(input_ids=input_ids, attention_mask=attention_mask)
with self.subTest("JIT Enabled"):
jitted_outputs = encode_jitted(**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = encode_jitted(**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
def test_decode(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
model = model_class(config)
encoder_outputs = model.encode(inputs_dict["input_ids"], inputs_dict["attention_mask"])
prepared_inputs_dict = {
"decoder_input_ids": inputs_dict["decoder_input_ids"],
"decoder_attention_mask": inputs_dict["decoder_attention_mask"],
"encoder_outputs": encoder_outputs,
}
@jax.jit
def decode_jitted(decoder_input_ids, decoder_attention_mask, encoder_outputs):
return model.decode(
decoder_input_ids=decoder_input_ids,
decoder_attention_mask=decoder_attention_mask,
encoder_outputs=encoder_outputs,
)
with self.subTest("JIT Enabled"):
jitted_outputs = decode_jitted(**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = decode_jitted(**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained("google/pegasus-large", from_pt=True)
input_ids = np.ones((1, 1))
outputs = model(input_ids)
self.assertIsNotNone(outputs)
@slow
def test_pegasus_xsum_summary(self):
model = FlaxPegasusForConditionalGeneration.from_pretrained("google/pegasus-xsum")
tokenizer = PegasusTokenizer.from_pretrained("google/pegasus-xsum")
src_text = [
""" PG&E stated it scheduled the blackouts in response to forecasts for high winds amid dry conditions. The aim is to reduce the risk of wildfires. Nearly 800 thousand customers were scheduled to be affected by the shutoffs which were expected to last through at least midday tomorrow.""",
""" The London trio are up for best UK act and best album, as well as getting two nominations in the best song category."We got told like this morning 'Oh I think you're nominated'", said Dappy."And I was like 'Oh yeah, which one?' And now we've got nominated for four awards. I mean, wow!"Bandmate Fazer added: "We thought it's best of us to come down and mingle with everyone and say hello to the cameras. And now we find we've got four nominations."The band have two shots at the best song prize, getting the nod for their Tynchy Stryder collaboration Number One, and single Strong Again.Their album Uncle B will also go up against records by the likes of Beyonce and Kanye West.N-Dubz picked up the best newcomer Mobo in 2007, but female member Tulisa said they wouldn't be too disappointed if they didn't win this time around."At the end of the day we're grateful to be where we are in our careers."If it don't happen then it don't happen - live to fight another day and keep on making albums and hits for the fans."Dappy also revealed they could be performing live several times on the night.The group will be doing Number One and also a possible rendition of the War Child single, I Got Soul.The charity song is a re-working of The Killers' All These Things That I've Done and is set to feature artists like Chipmunk, Ironik and Pixie Lott.This year's Mobos will be held outside of London for the first time, in Glasgow on 30 September.N-Dubz said they were looking forward to performing for their Scottish fans and boasted about their recent shows north of the border."We just done Edinburgh the other day," said Dappy."We smashed up an N-Dubz show over there. We done Aberdeen about three or four months ago - we smashed up that show over there! Everywhere we go we smash it up!" """,
]
tgt_text = [
"California's largest electricity provider has turned off power to hundreds of thousands of customers.",
"Pop group N-Dubz have revealed they were surprised to get four nominations for this year's Mobo Awards.",
]
inputs = tokenizer(src_text, return_tensors="np", truncation=True, max_length=512, padding=True)
translated_tokens = model.generate(**inputs, num_beams=2).sequences
decoded = tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)
assert tgt_text == decoded
class FlaxMBartModelTest(FlaxModelTesterMixin, unittest.TestCase,
FlaxGenerationTesterMixin):
is_encoder_decoder = True
all_model_classes = ((
FlaxMBartModel,
FlaxMBartForConditionalGeneration,
FlaxMBartForSequenceClassification,
FlaxMBartForQuestionAnswering,
) if is_flax_available() else ())
all_generative_model_classes = (
FlaxMBartForConditionalGeneration, ) if is_flax_available() else ()
def setUp(self):
self.model_tester = FlaxMBartModelTester(self)
def test_use_cache_forward(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
for model_class in self.all_model_classes:
self.model_tester.check_use_cache_forward(model_class, config,
inputs_dict)
def test_use_cache_forward_with_attn_mask(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
for model_class in self.all_model_classes:
self.model_tester.check_use_cache_forward_with_attn_mask(
model_class, config, inputs_dict)
def test_encode(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
model = model_class(config)
@jax.jit
def encode_jitted(input_ids, attention_mask=None, **kwargs):
return model.encode(input_ids=input_ids,
attention_mask=attention_mask)
with self.subTest("JIT Enabled"):
jitted_outputs = encode_jitted(
**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = encode_jitted(
**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
def test_decode(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
model = model_class(config)
encoder_outputs = model.encode(inputs_dict["input_ids"],
inputs_dict["attention_mask"])
prepared_inputs_dict = {
"decoder_input_ids": inputs_dict["decoder_input_ids"],
"decoder_attention_mask":
inputs_dict["decoder_attention_mask"],
"encoder_outputs": encoder_outputs,
}
@jax.jit
def decode_jitted(decoder_input_ids, decoder_attention_mask,
encoder_outputs):
return model.decode(
decoder_input_ids=decoder_input_ids,
decoder_attention_mask=decoder_attention_mask,
encoder_outputs=encoder_outputs,
)
with self.subTest("JIT Enabled"):
jitted_outputs = decode_jitted(
**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = decode_jitted(
**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained(
"facebook/mbart-large-cc25", from_pt=True)
# FlaxMBartForSequenceClassification expects eos token in input_ids
input_ids = np.ones((1, 1)) * model.config.eos_token_id
outputs = model(input_ids)
self.assertIsNotNone(outputs)
class FlaxT5ModelTest(FlaxModelTesterMixin, FlaxGenerationTesterMixin,
unittest.TestCase):
all_model_classes = (
FlaxT5Model,
FlaxT5ForConditionalGeneration) if is_flax_available() else ()
all_generative_model_classes = (
FlaxT5ForConditionalGeneration, ) if is_flax_available() else ()
is_encoder_decoder = True
def setUp(self):
self.model_tester = FlaxT5ModelTester(self)
self.config_tester = ConfigTester(self,
config_class=T5Config,
d_model=37)
def test_config(self):
self.config_tester.run_common_tests()
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_model_v1_1(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
# check that gated gelu feed forward and different word embeddings work
config = config_and_inputs[0]
config.tie_word_embeddings = False
config.feed_forward_proj = "gated-gelu"
self.model_tester.create_and_check_model(config,
*config_and_inputs[1:])
def test_use_cache_forward_with_attn_mask(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
for model_class in self.all_model_classes:
self.model_tester.check_use_cache_forward_with_attn_mask(
model_class, *config_and_inputs)
def test_encode(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
model = model_class(config)
@jax.jit
def encode_jitted(input_ids, attention_mask=None, **kwargs):
return model.encode(input_ids=input_ids,
attention_mask=attention_mask)
with self.subTest("JIT Enabled"):
jitted_outputs = encode_jitted(
**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = encode_jitted(
**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
def test_decode(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
model = model_class(config)
encoder_outputs = model.encode(inputs_dict["input_ids"],
inputs_dict["attention_mask"])
prepared_inputs_dict = {
"decoder_input_ids": inputs_dict["decoder_input_ids"],
"decoder_attention_mask":
inputs_dict["decoder_attention_mask"],
"encoder_outputs": encoder_outputs,
}
@jax.jit
def decode_jitted(decoder_input_ids, decoder_attention_mask,
encoder_outputs):
return model.decode(
decoder_input_ids=decoder_input_ids,
decoder_attention_mask=decoder_attention_mask,
encoder_outputs=encoder_outputs,
)
with self.subTest("JIT Enabled"):
jitted_outputs = decode_jitted(
**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = decode_jitted(
**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
def test_shift_right(self):
decoder_start_token_id = 0
pad_token_id = 1
labels = np.arange(2, 102).reshape(5, 20)
labels[:2, 15:] = -100
decoder_input_ids = shift_tokens_right(labels, pad_token_id,
decoder_start_token_id)
np_decoder_input_ids = np.array(decoder_input_ids)
padded_slice = np_decoder_input_ids[:2, (15 + 1):]
self.assertTrue((padded_slice == 1).all())
not_padded_slice = np_decoder_input_ids[2:, 1:]
rolled_labels = np.roll(labels[2:], 1)[:, 1:]
self.assertTrue((not_padded_slice == rolled_labels).all())
self.assertTrue((np_decoder_input_ids[:, 0] == 0).all())
# overwrite since special base model prefix is used
def test_save_load_from_base(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
base_class = FLAX_MODEL_MAPPING[config.__class__]
for model_class in self.all_model_classes:
if model_class == base_class:
continue
model = base_class(config)
base_params = flatten_dict(unfreeze(model.params))
# check that all base model weights are loaded correctly
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
head_model = model_class.from_pretrained(tmpdirname)
base_param_from_head = flatten_dict(unfreeze(
head_model.params))
for key in base_param_from_head.keys():
max_diff = (base_params[key] -
base_param_from_head[key]).sum().item()
self.assertLessEqual(max_diff,
1e-3,
msg=f"{key} not identical")
# overwrite since special base model prefix is used
def test_save_load_to_base(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
base_class = FLAX_MODEL_MAPPING[config.__class__]
for model_class in self.all_model_classes:
if model_class == base_class:
continue
model = model_class(config)
base_params_from_head = flatten_dict(unfreeze(model.params))
# check that all base model weights are loaded correctly
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
base_model = base_class.from_pretrained(tmpdirname)
base_params = flatten_dict(unfreeze(base_model.params))
for key in base_params_from_head.keys():
max_diff = (base_params[key] -
base_params_from_head[key]).sum().item()
self.assertLessEqual(max_diff,
1e-3,
msg=f"{key} not identical")
class FlaxViTModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = (FlaxViTModel, FlaxViTForImageClassification) if is_flax_available() else ()
def setUp(self) -> None:
self.model_tester = FlaxViTModelTester(self)
self.config_tester = ConfigTester(self, config_class=ViTConfig, has_text_modality=False, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
# We need to override this test because in ViT, the seq_len equals the number of patches + 1
# we compute that here
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
num_patches = (config.image_size // config.patch_size) ** 2
seq_length = num_patches + 1
for model_class in self.all_model_classes:
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = False
model = model_class(config)
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
attentions = 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)
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
attentions = outputs.attentions
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, seq_length, seq_length],
)
out_len = len(outputs)
# Check attention is always last and order is fine
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = True
model = model_class(config)
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
added_hidden_states = 1
self.assertEqual(out_len + added_hidden_states, len(outputs))
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, seq_length, seq_length],
)
# We neeed to override this test because ViT's forward signature is different than text models.
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.__call__)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["pixel_values"]
self.assertListEqual(arg_names[:1], expected_arg_names)
# We neeed to override this test because ViT expects pixel_values instead of input_ids
def test_jit_compilation(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
model = model_class(config)
@jax.jit
def model_jitted(pixel_values, **kwargs):
return model(pixel_values=pixel_values, **kwargs)
with self.subTest("JIT Enabled"):
jitted_outputs = model_jitted(**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = model_jitted(**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
# We need to override this test because in ViT, the seq_len equals the number of patches + 1
# we compute that here
def test_hidden_states_output(self):
def check_hidden_states_output(inputs_dict, config, model_class):
model = model_class(config)
num_patches = (config.image_size // config.patch_size) ** 2
seq_length = num_patches + 1 # we add 1 for the [CLS] token
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
hidden_states = outputs.hidden_states
self.assertEqual(len(hidden_states), self.model_tester.num_hidden_layers + 1)
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
[seq_length, self.model_tester.hidden_size],
)
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
inputs_dict["output_hidden_states"] = True
check_hidden_states_output(inputs_dict, config, model_class)
# check that output_hidden_states also work using config
del inputs_dict["output_hidden_states"]
config.output_hidden_states = True
check_hidden_states_output(inputs_dict, config, model_class)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained("google/vit-base-patch16-224")
outputs = model(np.ones((1, 3, 224, 224)))
self.assertIsNotNone(outputs)
class FlaxBigBirdModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = ((
FlaxBigBirdForCausalLM,
FlaxBigBirdModel,
FlaxBigBirdForPreTraining,
FlaxBigBirdForMaskedLM,
FlaxBigBirdForMultipleChoice,
FlaxBigBirdForQuestionAnswering,
FlaxBigBirdForSequenceClassification,
FlaxBigBirdForTokenClassification,
) if is_flax_available() else ())
test_attn_probs = False
test_mismatched_shapes = False
def setUp(self):
self.model_tester = FlaxBigBirdModelTester(self)
@slow
# copied from `test_modeling_flax_common` because it takes much longer than other models
def test_from_pretrained_save_pretrained(self):
super().test_from_pretrained_save_pretrained()
@slow
# copied from `test_modeling_flax_common` because it takes much longer than other models
def test_from_pretrained_with_no_automatic_init(self):
super().test_from_pretrained_with_no_automatic_init()
@slow
# copied from `test_modeling_flax_common` because it takes much longer than other models
def test_no_automatic_init(self):
super().test_no_automatic_init()
@slow
# copied from `test_modeling_flax_common` because it takes much longer than other models
def test_hidden_states_output(self):
super().test_hidden_states_output()
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained(
"google/bigbird-roberta-base")
outputs = model(np.ones((1, 1)))
self.assertIsNotNone(outputs)
def test_attention_outputs(self):
if self.test_attn_probs:
super().test_attention_outputs()
@slow
# copied from `test_modeling_flax_common` because it takes much longer than other models
def test_jit_compilation(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
)
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
prepared_inputs_dict = self._prepare_for_class(
inputs_dict, model_class)
model = model_class(config)
@jax.jit
def model_jitted(input_ids, attention_mask=None, **kwargs):
return model(input_ids=input_ids,
attention_mask=attention_mask,
**kwargs)
with self.subTest("JIT Enabled"):
jitted_outputs = model_jitted(
**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = model_jitted(
**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
# overwrite from common in order to skip the check on `attentions`
def check_pt_flax_outputs(self,
fx_outputs,
pt_outputs,
model_class,
tol=1e-5,
name="outputs",
attributes=None):
# `bigbird_block_sparse_attention` in `FlaxBigBird` returns `attention_probs = None`, while in PyTorch version,
# an effort was done to return `attention_probs` (yet to be verified).
if name.startswith("outputs.attentions"):
return
else:
super().check_pt_flax_outputs(fx_outputs, pt_outputs, model_class,
tol, name, attributes)
