def test_real_model_save_load_from_pretrained(self):
model_2 = self.get_pretrained_model()
pixel_values = floats_tensor(
[
13,
model_2.config.encoder.num_channels,
model_2.config.encoder.image_size,
model_2.config.encoder.image_size,
]
)
decoder_input_ids = ids_tensor([13, 1], model_2.config.decoder.vocab_size)
outputs = model_2(
pixel_values=pixel_values,
decoder_input_ids=decoder_input_ids,
)
out_2 = np.array(outputs[0])
out_2[np.isnan(out_2)] = 0
with tempfile.TemporaryDirectory() as tmp_dirname:
model_2.save_pretrained(tmp_dirname)
model_1 = TFVisionEncoderDecoderModel.from_pretrained(tmp_dirname)
after_outputs = model_1(pixel_values=pixel_values, decoder_input_ids=decoder_input_ids)
out_1 = np.array(after_outputs[0])
out_1[np.isnan(out_1)] = 0
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
def check_save_and_load(
self,
config,
pixel_values,
encoder_hidden_states,
decoder_config,
decoder_input_ids,
decoder_attention_mask,
**kwargs
):
encoder_model, decoder_model = self.get_encoder_decoder_model(config, decoder_config)
enc_dec_model = TFVisionEncoderDecoderModel(encoder=encoder_model, decoder=decoder_model)
outputs = enc_dec_model(
pixel_values=pixel_values,
decoder_input_ids=decoder_input_ids,
decoder_attention_mask=decoder_attention_mask,
)
out_2 = np.array(outputs[0])
out_2[np.isnan(out_2)] = 0
with tempfile.TemporaryDirectory() as tmpdirname:
enc_dec_model.save_pretrained(tmpdirname)
enc_dec_model = TFVisionEncoderDecoderModel.from_pretrained(tmpdirname)
after_outputs = enc_dec_model(
pixel_values=pixel_values,
decoder_input_ids=decoder_input_ids,
decoder_attention_mask=decoder_attention_mask,
)
out_1 = np.array(after_outputs[0])
out_1[np.isnan(out_1)] = 0
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
def test_inference_coco_en(self):
loc = "ydshieh/vit-gpt2-coco-en"
feature_extractor = ViTFeatureExtractor.from_pretrained(loc)
tokenizer = AutoTokenizer.from_pretrained(loc)
model = TFVisionEncoderDecoderModel.from_pretrained(loc)
# We will verify our results on an image of cute cats
img = Image.open(
"./tests/fixtures/tests_samples/COCO/000000039769.png")
pixel_values = feature_extractor(images=img,
return_tensors="tf").pixel_values
decoder_input_ids = tf.constant([[model.config.decoder_start_token_id]
])
logits = model(pixel_values, decoder_input_ids)[0].numpy()
# verify the logits
expected_shape = (1, 1, model.config.decoder.vocab_size)
self.assertEqual(logits.shape, expected_shape)
EXPECTED_LOGIT_SLICE = np.array([
-38.705807,
-30.639929,
-31.41903,
-39.012012,
-38.38696,
-34.887207,
-33.290855,
-35.68447,
-38.508484,
-36.124645,
])
max_diff = np.amax(np.abs(logits[0, 0, :10] - EXPECTED_LOGIT_SLICE))
self.assertLessEqual(max_diff, 1e-4)
def generate_step(pixel_values):
outputs = model.generate(pixel_values,
max_length=16,
num_beams=4,
return_dict_in_generate=True,
output_scores=True)
output_ids = outputs.sequences
preds = tokenizer.batch_decode(output_ids,
skip_special_tokens=True)
preds = [pred.strip() for pred in preds]
return preds, outputs.scores.numpy()
preds, scores = generate_step(pixel_values)
# should produce
# ["a cat laying on top of a couch next to another cat"]
self.assertEqual(
preds, ["a cat laying on top of a couch next to another cat"])
def test_encoder_decoder_save_load_from_encoder_decoder(self):
config = self.get_encoder_decoder_config_small()
# create two random ViT/GPT2 models for vit-gpt2 & initialize weights (+cross_attention weights)
encoder = TFViTModel(config.encoder)
encoder(encoder.dummy_inputs)
decoder = TFGPT2LMHeadModel(config.decoder)
decoder(decoder.dummy_inputs)
encoder_decoder_orig = TFVisionEncoderDecoderModel(encoder=encoder,
decoder=decoder)
pixel_values = floats_tensor([
13,
encoder.config.num_channels,
encoder.config.image_size,
encoder.config.image_size,
])
decoder_input_ids = ids_tensor([13, 1], decoder.config.vocab_size)
logits_orig = encoder_decoder_orig(
pixel_values=pixel_values,
decoder_input_ids=decoder_input_ids).logits
with tempfile.TemporaryDirectory() as tmp_dirname:
encoder_path = os.path.join(tmp_dirname, "encoder")
decoder_path = os.path.join(tmp_dirname, "decoder")
encoder.save_pretrained(encoder_path)
decoder.save_pretrained(decoder_path)
encoder_decoder = TFVisionEncoderDecoderModel.from_encoder_decoder_pretrained(
encoder_path, decoder_path)
logits_1 = encoder_decoder(pixel_values=pixel_values,
decoder_input_ids=decoder_input_ids).logits
self.assertTrue(
logits_orig.numpy().sum() - logits_1.numpy().sum() < 1e-3)
max_diff = np.max(np.abs(logits_1.numpy() - logits_orig.numpy()))
self.assertAlmostEqual(max_diff, 0.0, places=4)
with tempfile.TemporaryDirectory() as tmp_dirname:
encoder_decoder.save_pretrained(tmp_dirname)
encoder_decoder = TFVisionEncoderDecoderModel.from_pretrained(
tmp_dirname)
logits_2 = encoder_decoder(pixel_values=pixel_values,
decoder_input_ids=decoder_input_ids).logits
max_diff = np.max(np.abs(logits_2.numpy() - logits_orig.numpy()))
self.assertAlmostEqual(max_diff, 0.0, places=4)
def test_encoder_decoder_save_load_from_encoder_decoder_from_pt(self):
config = self.get_encoder_decoder_config_small()
# create two random ViT/GPT2 models for vit-gpt2 & initialize weights (+cross_attention weights)
encoder_pt = ViTModel(config.encoder).to(torch_device).eval()
decoder_pt = GPT2LMHeadModel(config.decoder).to(torch_device).eval()
encoder_decoder_pt = VisionEncoderDecoderModel(encoder=encoder_pt, decoder=decoder_pt).to(torch_device).eval()
pixel_values = floats_tensor(
[
13,
encoder_pt.config.num_channels,
encoder_pt.config.image_size,
encoder_pt.config.image_size,
]
)
decoder_input_ids = ids_tensor([13, 1], decoder_pt.config.vocab_size)
pt_pixel_values = torch.tensor(pixel_values.numpy(), device=torch_device, dtype=torch.float)
pt_decoder_input_ids = torch.tensor(decoder_input_ids.numpy(), device=torch_device, dtype=torch.long)
logits_pt = encoder_decoder_pt(pixel_values=pt_pixel_values, decoder_input_ids=pt_decoder_input_ids).logits
# PyTorch => TensorFlow
with tempfile.TemporaryDirectory() as tmp_dirname_1, tempfile.TemporaryDirectory() as tmp_dirname_2:
encoder_decoder_pt.encoder.save_pretrained(tmp_dirname_1)
encoder_decoder_pt.decoder.save_pretrained(tmp_dirname_2)
encoder_decoder_tf = TFVisionEncoderDecoderModel.from_encoder_decoder_pretrained(
tmp_dirname_1, tmp_dirname_2, encoder_from_pt=True, decoder_from_pt=True
)
logits_tf = encoder_decoder_tf(pixel_values=pixel_values, decoder_input_ids=decoder_input_ids).logits
max_diff = np.max(np.abs(logits_pt.detach().cpu().numpy() - logits_tf.numpy()))
self.assertAlmostEqual(max_diff, 0.0, places=3)
# Make sure `from_pretrained` following `save_pretrained` work and give the same result
# (See https://github.com/huggingface/transformers/pull/14016)
with tempfile.TemporaryDirectory() as tmp_dirname:
encoder_decoder_tf.save_pretrained(tmp_dirname)
encoder_decoder_tf = TFVisionEncoderDecoderModel.from_pretrained(tmp_dirname)
logits_tf_2 = encoder_decoder_tf(pixel_values=pixel_values, decoder_input_ids=decoder_input_ids).logits
max_diff = np.max(np.abs(logits_tf_2.numpy() - logits_tf.numpy()))
self.assertAlmostEqual(max_diff, 0.0, places=3)
def test_encoder_decoder_from_pretrained(self):
load_weight_prefix = TFVisionEncoderDecoderModel.load_weight_prefix
config = self.get_encoder_decoder_config()
feature_extractor = AutoFeatureExtractor.from_pretrained("google/vit-base-patch16-224-in21k")
decoder_tokenizer = AutoTokenizer.from_pretrained("../gpt2")
img = prepare_img()
pixel_values = feature_extractor(images=img, return_tensors="tf").pixel_values
decoder_input_ids = decoder_tokenizer("Linda Davis", return_tensors="tf").input_ids
with tempfile.TemporaryDirectory() as tmp_dirname:
# Since most of HF's models don't have pretrained cross-attention layers, they are randomly
# initialized even if we create models using `from_pretrained` method.
# For the tests, the decoder need to be a model with pretrained cross-attention layers.
# So we create pretrained models (without `load_weight_prefix`), save them, and later,
# we load them using `from_pretrained`.
# (we don't need to do this for encoder, but let's make the code more similar between encoder/decoder)
encoder = TFAutoModel.from_pretrained("google/vit-base-patch16-224-in21k", name="encoder")
# It's necessary to specify `add_cross_attention=True` here.
decoder = TFAutoModelForCausalLM.from_pretrained(
"../gpt2", is_decoder=True, add_cross_attention=True, name="decoder"
)
pretrained_encoder_dir = os.path.join(tmp_dirname, "pretrained_encoder")
pretrained_decoder_dir = os.path.join(tmp_dirname, "pretrained_decoder")
encoder.save_pretrained(pretrained_encoder_dir)
decoder.save_pretrained(pretrained_decoder_dir)
del encoder
del decoder
enc_dec_model = TFVisionEncoderDecoderModel.from_encoder_decoder_pretrained(
pretrained_encoder_dir,
pretrained_decoder_dir,
)
# check that the from pretrained methods work
enc_dec_model.save_pretrained(tmp_dirname)
enc_dec_model = TFVisionEncoderDecoderModel.from_pretrained(tmp_dirname)
output = enc_dec_model(pixel_values, decoder_input_ids=decoder_input_ids, labels=decoder_input_ids)
loss_pretrained = output.loss
del enc_dec_model
# Create the model using `__init__` with loaded ``pretrained`` encoder / decoder
encoder = TFAutoModel.from_pretrained(
pretrained_encoder_dir, load_weight_prefix=load_weight_prefix, name="encoder"
)
decoder = TFAutoModelForCausalLM.from_pretrained(
pretrained_decoder_dir, load_weight_prefix=load_weight_prefix, name="decoder"
)
enc_dec_model = TFVisionEncoderDecoderModel(config=config, encoder=encoder, decoder=decoder)
output = enc_dec_model(pixel_values, decoder_input_ids=decoder_input_ids, labels=decoder_input_ids)
loss_init = output.loss
max_diff = np.max(np.abs(loss_pretrained - loss_init))
expected_diff = 0.0
self.assertAlmostEqual(max_diff, expected_diff, places=4)
def get_encoderdecoder_model(self):
return TFVisionEncoderDecoderModel.from_pretrained("ydshieh/vit-gpt2-coco-en")
