def test_inference_masked_image_modeling_head(self):
model = BeitForMaskedImageModeling.from_pretrained(
"microsoft/beit-base-patch16-224-pt22k").to(torch_device)
feature_extractor = self.default_feature_extractor
image = prepare_img()
pixel_values = feature_extractor(
images=image, return_tensors="pt").pixel_values.to(torch_device)
# prepare bool_masked_pos
bool_masked_pos = torch.ones((1, 196),
dtype=torch.bool).to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(pixel_values=pixel_values,
bool_masked_pos=bool_masked_pos)
logits = outputs.logits
# verify the logits
expected_shape = torch.Size((1, 196, 8192))
self.assertEqual(logits.shape, expected_shape)
expected_slice = torch.tensor([[-3.2437, 0.5072, -13.9174],
[-3.2456, 0.4948, -13.9401],
[-3.2033, 0.5121,
-13.8550]]).to(torch_device)
self.assertTrue(
torch.allclose(logits[bool_masked_pos][:3, :3],
expected_slice,
atol=1e-2))
def create_and_check_for_masked_lm(self, config, pixel_values, labels,
pixel_labels):
model = BeitForMaskedImageModeling(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
self.parent.assertEqual(
result.logits.shape,
(self.batch_size, self.seq_length - 1, self.vocab_size))
def create_and_check_for_masked_lm(self, config, pixel_values, labels):
model = BeitForMaskedImageModeling(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
# expected sequence length = num_patches
image_size = to_2tuple(self.image_size)
patch_size = to_2tuple(self.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
self.parent.assertEqual(result.logits.shape, (self.batch_size, num_patches, self.vocab_size))
def convert_beit_checkpoint(checkpoint_url, pytorch_dump_folder_path):
"""
Copy/paste/tweak model's weights to our BEiT structure.
"""
# define default BEiT configuration
config = BeitConfig()
has_lm_head = False
is_semantic = False
repo_id = "datasets/huggingface/label-files"
# set config parameters based on URL
if checkpoint_url[-9:-4] == "pt22k":
# masked image modeling
config.use_shared_relative_position_bias = True
config.use_mask_token = True
has_lm_head = True
elif checkpoint_url[-9:-4] == "ft22k":
# intermediate fine-tuning on ImageNet-22k
config.use_relative_position_bias = True
config.num_labels = 21841
filename = "imagenet-22k-id2label.json"
id2label = json.load(open(hf_hub_download(repo_id, filename), "r"))
id2label = {int(k): v for k, v in id2label.items()}
# this dataset contains 21843 labels but the model only has 21841
# we delete the classes as mentioned in https://github.com/google-research/big_transfer/issues/18
del id2label[9205]
del id2label[15027]
config.id2label = id2label
config.label2id = {v: k for k, v in id2label.items()}
elif checkpoint_url[-8:-4] == "to1k":
# fine-tuning on ImageNet-1k
config.use_relative_position_bias = True
config.num_labels = 1000
filename = "imagenet-1k-id2label.json"
id2label = json.load(open(hf_hub_download(repo_id, filename), "r"))
id2label = {int(k): v for k, v in id2label.items()}
config.id2label = id2label
config.label2id = {v: k for k, v in id2label.items()}
if "384" in checkpoint_url:
config.image_size = 384
if "512" in checkpoint_url:
config.image_size = 512
elif "ade20k" in checkpoint_url:
# fine-tuning
config.use_relative_position_bias = True
config.num_labels = 150
filename = "ade20k-id2label.json"
id2label = json.load(open(hf_hub_download(repo_id, filename), "r"))
id2label = {int(k): v for k, v in id2label.items()}
config.id2label = id2label
config.label2id = {v: k for k, v in id2label.items()}
config.image_size = 640
is_semantic = True
else:
raise ValueError(
"Checkpoint not supported, URL should either end with 'pt22k', 'ft22k', 'to1k' or 'ade20k'"
)
# size of the architecture
if "base" in checkpoint_url:
pass
elif "large" in checkpoint_url:
config.hidden_size = 1024
config.intermediate_size = 4096
config.num_hidden_layers = 24
config.num_attention_heads = 16
if "ade20k" in checkpoint_url:
config.image_size = 640
config.out_indices = [7, 11, 15, 23]
else:
raise ValueError(
"Should either find 'base' or 'large' in checkpoint URL")
# load state_dict of original model, remove and rename some keys
state_dict = torch.hub.load_state_dict_from_url(checkpoint_url,
map_location="cpu",
check_hash=True)
state_dict = state_dict[
"model"] if "ade20k" not in checkpoint_url else state_dict["state_dict"]
rename_keys = create_rename_keys(config,
has_lm_head=has_lm_head,
is_semantic=is_semantic)
for src, dest in rename_keys:
rename_key(state_dict, src, dest)
read_in_q_k_v(state_dict,
config,
has_lm_head=has_lm_head,
is_semantic=is_semantic)
if is_semantic:
# add prefix to decoder keys
for key, val in state_dict.copy().items():
val = state_dict.pop(key)
if key.startswith("backbone.fpn"):
key = key.replace("backbone.fpn", "fpn")
state_dict[key] = val
# load HuggingFace model
if checkpoint_url[-9:-4] == "pt22k":
model = BeitForMaskedImageModeling(config)
elif "ade20k" in checkpoint_url:
model = BeitForSemanticSegmentation(config)
else:
model = BeitForImageClassification(config)
model.eval()
model.load_state_dict(state_dict)
# Check outputs on an image
if is_semantic:
feature_extractor = BeitFeatureExtractor(size=config.image_size,
do_center_crop=False)
ds = load_dataset("hf-internal-testing/fixtures_ade20k", split="test")
image = Image.open(ds[0]["file"])
else:
feature_extractor = BeitFeatureExtractor(size=config.image_size,
resample=Image.BILINEAR,
do_center_crop=False)
image = prepare_img()
encoding = feature_extractor(images=image, return_tensors="pt")
pixel_values = encoding["pixel_values"]
outputs = model(pixel_values)
logits = outputs.logits
# verify logits
expected_shape = torch.Size([1, 1000])
if checkpoint_url[:-4].endswith("beit_base_patch16_224_pt22k"):
expected_shape = torch.Size([1, 196, 8192])
elif checkpoint_url[:-4].endswith("beit_large_patch16_224_pt22k"):
expected_shape = torch.Size([1, 196, 8192])
elif checkpoint_url[:-4].endswith("beit_base_patch16_224_pt22k_ft22k"):
expected_shape = torch.Size([1, 21841])
expected_logits = torch.tensor([2.2288, 2.4671, 0.7395])
expected_class_idx = 2397
elif checkpoint_url[:-4].endswith("beit_large_patch16_224_pt22k_ft22k"):
expected_shape = torch.Size([1, 21841])
expected_logits = torch.tensor([1.6881, -0.2787, 0.5901])
expected_class_idx = 2396
elif checkpoint_url[:-4].endswith("beit_base_patch16_224_pt22k_ft1k"):
expected_logits = torch.tensor([0.1241, 0.0798, -0.6569])
expected_class_idx = 285
elif checkpoint_url[:-4].endswith("beit_base_patch16_224_pt22k_ft22kto1k"):
expected_logits = torch.tensor([-1.2385, -1.0987, -1.0108])
expected_class_idx = 281
elif checkpoint_url[:-4].endswith("beit_base_patch16_384_pt22k_ft22kto1k"):
expected_logits = torch.tensor([-1.5303, -0.9484, -0.3147])
expected_class_idx = 761
elif checkpoint_url[:-4].endswith("beit_large_patch16_224_pt22k_ft1k"):
expected_logits = torch.tensor([0.4610, -0.0928, 0.2086])
expected_class_idx = 761
elif checkpoint_url[:-4].endswith(
"beit_large_patch16_224_pt22k_ft22kto1k"):
expected_logits = torch.tensor([-0.4804, 0.6257, -0.1837])
expected_class_idx = 761
elif checkpoint_url[:-4].endswith(
"beit_large_patch16_384_pt22k_ft22kto1k"):
expected_logits = torch.tensor([[-0.5122, 0.5117, -0.2113]])
expected_class_idx = 761
elif checkpoint_url[:-4].endswith(
"beit_large_patch16_512_pt22k_ft22kto1k"):
expected_logits = torch.tensor([-0.3062, 0.7261, 0.4852])
expected_class_idx = 761
elif checkpoint_url[:-4].endswith(
"beit_base_patch16_640_pt22k_ft22ktoade20k"):
expected_shape = (1, 150, 160, 160)
expected_logits = torch.tensor([
[[-4.9225, -2.3954, -3.0522], [-2.8822, -1.0046, -1.7561],
[-2.9549, -1.3228, -2.1347]],
[[-5.8168, -3.4129, -4.0778], [-3.8651, -2.2214, -3.0277],
[-3.8356, -2.4643, -3.3535]],
[[-0.0078, 3.9952, 4.0754], [2.9856, 4.6944, 5.0035],
[3.2413, 4.7813, 4.9969]],
])
elif checkpoint_url[:-4].endswith(
"beit_large_patch16_640_pt22k_ft22ktoade20k"):
expected_shape = (1, 150, 160, 160)
expected_logits = torch.tensor([
[[-4.3305, -2.3049, -3.0161], [-2.9591, -1.5305, -2.2251],
[-3.4198, -1.8004, -2.9062]],
[[-5.8922, -3.7435, -4.3978], [-4.2063, -2.7872, -3.4755],
[-4.2791, -3.1874, -4.1681]],
[[0.9895, 4.3467, 4.7663], [4.2476, 5.6830, 6.1518],
[4.5550, 6.2495, 6.5154]],
])
else:
raise ValueError("Can't verify logits as model is not supported")
assert logits.shape == expected_shape, "Shape of logits not as expected"
if not has_lm_head:
if is_semantic:
assert torch.allclose(
logits[0, :3, :3, :3], expected_logits,
atol=1e-3), "First elements of logits not as expected"
else:
print("Predicted class idx:", logits.argmax(-1).item())
assert torch.allclose(
logits[0, :3], expected_logits,
atol=1e-3), "First elements of logits not as expected"
assert logits.argmax(-1).item(
) == expected_class_idx, "Predicted class index not as expected"
Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
print(f"Saving model to {pytorch_dump_folder_path}")
model.save_pretrained(pytorch_dump_folder_path)
print(f"Saving feature extractor to {pytorch_dump_folder_path}")
feature_extractor.save_pretrained(pytorch_dump_folder_path)
def convert_dit_checkpoint(checkpoint_url,
pytorch_dump_folder_path,
push_to_hub=False):
"""
Copy/paste/tweak model's weights to our BEiT structure.
"""
# define default BEiT configuration
has_lm_head = False if "rvlcdip" in checkpoint_url else True
config = BeitConfig(use_absolute_position_embeddings=True,
use_mask_token=has_lm_head)
# size of the architecture
if "large" in checkpoint_url or "dit-l" in checkpoint_url:
config.hidden_size = 1024
config.intermediate_size = 4096
config.num_hidden_layers = 24
config.num_attention_heads = 16
# labels
if "rvlcdip" in checkpoint_url:
config.num_labels = 16
repo_id = "datasets/huggingface/label-files"
filename = "rvlcdip-id2label.json"
id2label = json.load(
open(cached_download(hf_hub_url(repo_id, filename)), "r"))
id2label = {int(k): v for k, v in id2label.items()}
config.id2label = id2label
config.label2id = {v: k for k, v in id2label.items()}
# load state_dict of original model, remove and rename some keys
state_dict = torch.hub.load_state_dict_from_url(
checkpoint_url, map_location="cpu")["model"]
rename_keys = create_rename_keys(config, has_lm_head=has_lm_head)
for src, dest in rename_keys:
rename_key(state_dict, src, dest)
read_in_q_k_v(state_dict, config, has_lm_head=has_lm_head)
# load HuggingFace model
model = BeitForMaskedImageModeling(
config) if has_lm_head else BeitForImageClassification(config)
model.eval()
model.load_state_dict(state_dict)
# Check outputs on an image
feature_extractor = BeitFeatureExtractor(size=config.image_size,
resample=Image.BILINEAR,
do_center_crop=False)
image = prepare_img()
encoding = feature_extractor(images=image, return_tensors="pt")
pixel_values = encoding["pixel_values"]
outputs = model(pixel_values)
logits = outputs.logits
# verify logits
expected_shape = [1, 16] if "rvlcdip" in checkpoint_url else [1, 196, 8192]
assert logits.shape == torch.Size(
expected_shape), "Shape of logits not as expected"
Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
print(f"Saving model to {pytorch_dump_folder_path}")
model.save_pretrained(pytorch_dump_folder_path)
print(f"Saving feature extractor to {pytorch_dump_folder_path}")
feature_extractor.save_pretrained(pytorch_dump_folder_path)
if push_to_hub:
if has_lm_head:
model_name = "dit-base" if "base" in checkpoint_url else "dit-large"
else:
model_name = "dit-base-finetuned-rvlcdip" if "dit-b" in checkpoint_url else "dit-large-finetuned-rvlcdip"
feature_extractor.push_to_hub(
repo_path_or_name=Path(pytorch_dump_folder_path, model_name),
organization="nielsr",
commit_message="Add feature extractor",
use_temp_dir=True,
)
model.push_to_hub(
repo_path_or_name=Path(pytorch_dump_folder_path, model_name),
organization="nielsr",
commit_message="Add model",
use_temp_dir=True,
)
def convert_beit_checkpoint(checkpoint_url, pytorch_dump_folder_path):
"""
Copy/paste/tweak model's weights to our BEiT structure.
"""
# define default BEiT configuration
config = BeitConfig()
has_lm_head = False
repo_id = "datasets/huggingface/label-files"
# set config parameters based on URL
if checkpoint_url[-9:-4] == "pt22k":
# masked image modeling
config.use_shared_relative_position_bias = True
config.use_mask_token = True
has_lm_head = True
elif checkpoint_url[-9:-4] == "ft22k":
# intermediate fine-tuning on ImageNet-22k
config.use_relative_position_bias = True
config.num_labels = 21841
filename = "imagenet-22k-id2label.json"
id2label = json.load(
open(cached_download(hf_hub_url(repo_id, filename)), "r"))
id2label = {int(k): v for k, v in id2label.items()}
# this dataset contains 21843 labels but the model only has 21841
# we delete the classes as mentioned in https://github.com/google-research/big_transfer/issues/18
del id2label[9205]
del id2label[15027]
config.id2label = id2label
config.label2id = {v: k for k, v in id2label.items()}
elif checkpoint_url[-8:-4] == "to1k":
# fine-tuning on ImageNet-1k
config.use_relative_position_bias = True
config.num_labels = 1000
filename = "imagenet-1k-id2label.json"
id2label = json.load(
open(cached_download(hf_hub_url(repo_id, filename)), "r"))
id2label = {int(k): v for k, v in id2label.items()}
config.id2label = id2label
config.label2id = {v: k for k, v in id2label.items()}
if "384" in checkpoint_url:
config.image_size = 384
if "512" in checkpoint_url:
config.image_size = 512
else:
raise ValueError(
"Checkpoint not supported, URL should either end with 'pt22k', 'ft22k' or 'to1k'"
)
# size of the architecture
if "base" in checkpoint_url:
pass
elif "large" in checkpoint_url:
config.hidden_size = 1024
config.intermediate_size = 4096
config.num_hidden_layers = 24
config.num_attention_heads = 16
else:
raise ValueError(
"Should either find 'base' or 'large' in checkpoint URL")
# load state_dict of original model, remove and rename some keys
state_dict = torch.hub.load_state_dict_from_url(checkpoint_url,
map_location="cpu",
check_hash=True)["model"]
rename_keys = create_rename_keys(config, has_lm_head=has_lm_head)
for src, dest in rename_keys:
rename_key(state_dict, src, dest)
read_in_q_k_v(state_dict, config, has_lm_head=has_lm_head)
# load HuggingFace model
if checkpoint_url[-9:-4] == "pt22k":
model = BeitForMaskedImageModeling(config)
else:
model = BeitForImageClassification(config)
model.eval()
model.load_state_dict(state_dict)
# Check outputs on an image
feature_extractor = BeitFeatureExtractor(size=config.image_size,
resample=Image.BILINEAR,
do_center_crop=False)
encoding = feature_extractor(images=prepare_img(), return_tensors="pt")
pixel_values = encoding["pixel_values"]
outputs = model(pixel_values)
logits = outputs.logits
# verify logits
expected_shape = torch.Size([1, 1000])
if checkpoint_url[:-4].endswith("beit_base_patch16_224_pt22k"):
expected_shape = torch.Size([1, 196, 8192])
elif checkpoint_url[:-4].endswith("beit_large_patch16_224_pt22k"):
expected_shape = torch.Size([1, 196, 8192])
elif checkpoint_url[:-4].endswith("beit_base_patch16_224_pt22k_ft22k"):
expected_shape = torch.Size([1, 21841])
expected_logits = torch.tensor([2.2288, 2.4671, 0.7395])
expected_class_idx = 2397
elif checkpoint_url[:-4].endswith("beit_large_patch16_224_pt22k_ft22k"):
expected_shape = torch.Size([1, 21841])
expected_logits = torch.tensor([1.6881, -0.2787, 0.5901])
expected_class_idx = 2396
elif checkpoint_url[:-4].endswith("beit_base_patch16_224_pt22k_ft1k"):
expected_logits = torch.tensor([0.1241, 0.0798, -0.6569])
expected_class_idx = 285
elif checkpoint_url[:-4].endswith("beit_base_patch16_224_pt22k_ft22kto1k"):
expected_logits = torch.tensor([-1.2385, -1.0987, -1.0108])
expected_class_idx = 281
elif checkpoint_url[:-4].endswith("beit_base_patch16_384_pt22k_ft22kto1k"):
expected_logits = torch.tensor([-1.5303, -0.9484, -0.3147])
expected_class_idx = 761
elif checkpoint_url[:-4].endswith("beit_large_patch16_224_pt22k_ft1k"):
expected_logits = torch.tensor([0.4610, -0.0928, 0.2086])
expected_class_idx = 761
elif checkpoint_url[:-4].endswith(
"beit_large_patch16_224_pt22k_ft22kto1k"):
expected_logits = torch.tensor([-0.4804, 0.6257, -0.1837])
expected_class_idx = 761
elif checkpoint_url[:-4].endswith(
"beit_large_patch16_384_pt22k_ft22kto1k"):
expected_logits = torch.tensor([[-0.5122, 0.5117, -0.2113]])
expected_class_idx = 761
elif checkpoint_url[:-4].endswith(
"beit_large_patch16_512_pt22k_ft22kto1k"):
expected_logits = torch.tensor([-0.3062, 0.7261, 0.4852])
expected_class_idx = 761
else:
raise ValueError("Can't verify logits as model is not supported")
assert logits.shape == expected_shape, "Shape of logits not as expected"
print("Shape of logits:", logits.shape)
if not has_lm_head:
print("Predicted class idx:", logits.argmax(-1).item())
assert torch.allclose(
logits[0, :3], expected_logits,
atol=1e-3), "First elements of logits not as expected"
assert logits.argmax(-1).item(
) == expected_class_idx, "Predicted class index not as expected"
Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
print(f"Saving model to {pytorch_dump_folder_path}")
model.save_pretrained(pytorch_dump_folder_path)
print(f"Saving feature extractor to {pytorch_dump_folder_path}")
feature_extractor.save_pretrained(pytorch_dump_folder_path)
