def test_save_load_pretrained_default(self):
tokenizer_slow = self.get_tokenizer()
tokenizer_fast = self.get_rust_tokenizer()
feature_extractor = self.get_feature_extractor()
processor_slow = OwlViTProcessor(tokenizer=tokenizer_slow,
feature_extractor=feature_extractor)
processor_slow.save_pretrained(self.tmpdirname)
processor_slow = OwlViTProcessor.from_pretrained(self.tmpdirname,
use_fast=False)
processor_fast = OwlViTProcessor(tokenizer=tokenizer_fast,
feature_extractor=feature_extractor)
processor_fast.save_pretrained(self.tmpdirname)
processor_fast = OwlViTProcessor.from_pretrained(self.tmpdirname)
self.assertEqual(processor_slow.tokenizer.get_vocab(),
tokenizer_slow.get_vocab())
self.assertEqual(processor_fast.tokenizer.get_vocab(),
tokenizer_fast.get_vocab())
self.assertEqual(tokenizer_slow.get_vocab(),
tokenizer_fast.get_vocab())
self.assertIsInstance(processor_slow.tokenizer, CLIPTokenizer)
self.assertIsInstance(processor_fast.tokenizer, CLIPTokenizerFast)
self.assertEqual(processor_slow.feature_extractor.to_json_string(),
feature_extractor.to_json_string())
self.assertEqual(processor_fast.feature_extractor.to_json_string(),
feature_extractor.to_json_string())
self.assertIsInstance(processor_slow.feature_extractor,
OwlViTFeatureExtractor)
self.assertIsInstance(processor_fast.feature_extractor,
OwlViTFeatureExtractor)
def test_tokenizer_decode(self):
feature_extractor = self.get_feature_extractor()
tokenizer = self.get_tokenizer()
processor = OwlViTProcessor(tokenizer=tokenizer,
feature_extractor=feature_extractor)
predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
decoded_processor = processor.batch_decode(predicted_ids)
decoded_tok = tokenizer.batch_decode(predicted_ids)
self.assertListEqual(decoded_tok, decoded_processor)
def test_inference_object_detection(self):
model_name = "google/owlvit-base-patch32"
model = OwlViTForObjectDetection.from_pretrained(model_name).to(
torch_device)
processor = OwlViTProcessor.from_pretrained(model_name)
image = prepare_img()
inputs = processor(
text=[["a photo of a cat", "a photo of a dog"]],
images=image,
max_length=16,
padding="max_length",
return_tensors="pt",
).to(torch_device)
with torch.no_grad():
outputs = model(**inputs)
num_queries = int((model.config.vision_config.image_size /
model.config.vision_config.patch_size)**2)
self.assertEqual(outputs.pred_boxes.shape,
torch.Size((1, num_queries, 4)))
expected_slice_boxes = torch.tensor([[0.0948, 0.0471, 0.1915],
[0.3194, 0.0583, 0.6498],
[0.1441, 0.0452,
0.2197]]).to(torch_device)
self.assertTrue(
torch.allclose(outputs.pred_boxes[0, :3, :3],
expected_slice_boxes,
atol=1e-4))
def test_inference(self):
model_name = "google/owlvit-base-patch32"
model = OwlViTModel.from_pretrained(model_name).to(torch_device)
processor = OwlViTProcessor.from_pretrained(model_name)
image = prepare_img()
inputs = processor(
text=[["a photo of a cat", "a photo of a dog"]],
images=image,
max_length=16,
padding="max_length",
return_tensors="pt",
).to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(**inputs)
# verify the logits
self.assertEqual(
outputs.logits_per_image.shape,
torch.Size(
(inputs.pixel_values.shape[0], inputs.input_ids.shape[0])),
)
self.assertEqual(
outputs.logits_per_text.shape,
torch.Size(
(inputs.input_ids.shape[0], inputs.pixel_values.shape[0])),
)
expected_logits = torch.tensor([[4.4420, 0.6181]], device=torch_device)
self.assertTrue(
torch.allclose(outputs.logits_per_image,
expected_logits,
atol=1e-3))
def convert_owlvit_checkpoint(pt_backbone, flax_params, attn_params, pytorch_dump_folder_path, config_path=None):
"""
Copy/paste/tweak model's weights to transformers design.
"""
repo = Repository(pytorch_dump_folder_path, clone_from=f"google/{pytorch_dump_folder_path}")
repo.git_pull()
if config_path is not None:
config = OwlViTConfig.from_pretrained(config_path)
else:
config = OwlViTConfig()
hf_backbone = OwlViTModel(config).eval()
hf_model = OwlViTForObjectDetection(config).eval()
copy_text_model_and_projection(hf_backbone, pt_backbone)
copy_vision_model_and_projection(hf_backbone, pt_backbone)
hf_backbone.logit_scale = pt_backbone.logit_scale
copy_flax_attn_params(hf_backbone, attn_params)
hf_model.owlvit = hf_backbone
copy_class_merge_token(hf_model, flax_params)
copy_class_box_heads(hf_model, flax_params)
# Save HF model
hf_model.save_pretrained(repo.local_dir)
# Initialize feature extractor
feature_extractor = OwlViTFeatureExtractor(
size=config.vision_config.image_size, crop_size=config.vision_config.image_size
)
# Initialize tokenizer
tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32", pad_token="!", model_max_length=16)
# Initialize processor
processor = OwlViTProcessor(feature_extractor=feature_extractor, tokenizer=tokenizer)
feature_extractor.save_pretrained(repo.local_dir)
processor.save_pretrained(repo.local_dir)
repo.git_add()
repo.git_commit("Upload model and processor")
repo.git_push()
def test_processor_with_text_list(self):
model_name = "google/owlvit-base-patch32"
processor = OwlViTProcessor.from_pretrained(model_name)
input_text = ["cat", "nasa badge"]
inputs = processor(text=input_text)
seq_length = 16
self.assertListEqual(list(inputs.keys()),
["input_ids", "attention_mask"])
self.assertEqual(inputs["input_ids"].shape, (2, seq_length))
# test if it raises when no input is passed
with pytest.raises(ValueError):
processor()
def test_save_load_pretrained_additional_features(self):
processor = OwlViTProcessor(
tokenizer=self.get_tokenizer(),
feature_extractor=self.get_feature_extractor())
processor.save_pretrained(self.tmpdirname)
tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)",
eos_token="(EOS)")
feature_extractor_add_kwargs = self.get_feature_extractor(
do_normalize=False)
processor = OwlViTProcessor.from_pretrained(self.tmpdirname,
bos_token="(BOS)",
eos_token="(EOS)",
do_normalize=False)
self.assertEqual(processor.tokenizer.get_vocab(),
tokenizer_add_kwargs.get_vocab())
self.assertIsInstance(processor.tokenizer, CLIPTokenizerFast)
self.assertEqual(processor.feature_extractor.to_json_string(),
feature_extractor_add_kwargs.to_json_string())
self.assertIsInstance(processor.feature_extractor,
OwlViTFeatureExtractor)
def test_tokenizer(self):
feature_extractor = self.get_feature_extractor()
tokenizer = self.get_tokenizer()
processor = OwlViTProcessor(tokenizer=tokenizer,
feature_extractor=feature_extractor)
input_str = "lower newer"
encoded_processor = processor(text=input_str, return_tensors="np")
encoded_tok = tokenizer(input_str, return_tensors="np")
for key in encoded_tok.keys():
self.assertListEqual(encoded_tok[key][0].tolist(),
encoded_processor[key][0].tolist())
def test_feature_extractor(self):
feature_extractor = self.get_feature_extractor()
tokenizer = self.get_tokenizer()
processor = OwlViTProcessor(tokenizer=tokenizer,
feature_extractor=feature_extractor)
image_input = self.prepare_image_inputs()
input_feat_extract = feature_extractor(image_input,
return_tensors="np")
input_processor = processor(images=image_input, return_tensors="np")
for key in input_feat_extract.keys():
self.assertAlmostEqual(input_feat_extract[key].sum(),
input_processor[key].sum(),
delta=1e-2)
def test_processor(self):
feature_extractor = self.get_feature_extractor()
tokenizer = self.get_tokenizer()
processor = OwlViTProcessor(tokenizer=tokenizer,
feature_extractor=feature_extractor)
input_str = "lower newer"
image_input = self.prepare_image_inputs()
inputs = processor(text=input_str, images=image_input)
self.assertListEqual(list(inputs.keys()),
["input_ids", "attention_mask", "pixel_values"])
# test if it raises when no input is passed
with pytest.raises(ValueError):
processor()
def test_processor_case(self):
model_name = "google/owlvit-base-patch32"
processor = OwlViTProcessor.from_pretrained(model_name)
input_texts = ["cat", "nasa badge"]
inputs = processor(text=input_texts)
seq_length = 16
input_ids = inputs["input_ids"]
predicted_ids = [
[49406, 2368, 49407, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[49406, 6841, 11301, 49407, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
]
self.assertListEqual(list(inputs.keys()),
["input_ids", "attention_mask"])
self.assertEqual(inputs["input_ids"].shape, (2, seq_length))
self.assertListEqual(list(input_ids[0]), predicted_ids[0])
self.assertListEqual(list(input_ids[1]), predicted_ids[1])
def test_processor_with_nested_text_list(self):
model_name = "google/owlvit-base-patch32"
processor = OwlViTProcessor.from_pretrained(model_name)
input_texts = [["cat", "nasa badge"], ["person"]]
inputs = processor(text=input_texts)
seq_length = 16
batch_size = len(input_texts)
num_max_text_queries = max([len(texts) for texts in input_texts])
self.assertListEqual(list(inputs.keys()),
["input_ids", "attention_mask"])
self.assertEqual(inputs["input_ids"].shape,
(batch_size * num_max_text_queries, seq_length))
# test if it raises when no input is passed
with pytest.raises(ValueError):
processor()