Update gme_inference.py
Browse files- gme_inference.py +63 -111
gme_inference.py
CHANGED
@@ -19,11 +19,11 @@ from transformers import (
|
|
19 |
AutoProcessor,
|
20 |
PreTrainedModel,
|
21 |
Qwen2VLConfig,
|
22 |
-
|
23 |
)
|
24 |
import os
|
|
|
25 |
|
26 |
-
# Define a config class for our model.
|
27 |
class GmeQwen2VLConfig(Qwen2VLConfig):
|
28 |
model_type: str = "gme_qwen2_vl"
|
29 |
|
@@ -39,11 +39,8 @@ class GmeQwen2VLConfig(Qwen2VLConfig):
|
|
39 |
self.min_image_tokens = min_image_tokens
|
40 |
self.max_image_tokens = max_image_tokens
|
41 |
self.max_length = max_length
|
42 |
-
self.device = device
|
43 |
-
AutoConfig.register("gme_qwen2_vl", GmeQwen2VLConfig)
|
44 |
|
45 |
|
46 |
-
# Define the model class so that it can be loaded by AutoModel.from_pretrained.
|
47 |
class GmeQwen2VLForVision2Seq(PreTrainedModel):
|
48 |
config_class = GmeQwen2VLConfig
|
49 |
base_model_prefix: str = "base"
|
@@ -51,29 +48,21 @@ class GmeQwen2VLForVision2Seq(PreTrainedModel):
|
|
51 |
def __init__(self, config: GmeQwen2VLConfig, **kwargs: Any) -> None:
|
52 |
super().__init__(config)
|
53 |
model_name: str = getattr(config, "_name_or_path", "Alibaba-NLP/gme-Qwen2-VL-2B-Instruct")
|
54 |
-
|
55 |
-
self.base =
|
56 |
-
model_name, trust_remote_code=True, **kwargs
|
57 |
-
)
|
58 |
self.normalize: bool = True
|
59 |
-
self.device: str = config.device
|
60 |
|
61 |
min_pixels: int = config.min_image_tokens * 28 * 28
|
62 |
max_pixels: int = config.max_image_tokens * 28 * 28
|
|
|
63 |
self.max_length: int = config.max_length
|
64 |
-
|
65 |
self.processor = AutoProcessor.from_pretrained(
|
66 |
model_name, min_pixels=min_pixels, max_pixels=max_pixels, **kwargs
|
67 |
)
|
68 |
-
self.processor.tokenizer.padding_side =
|
69 |
self.defualt_instruction: str = "You are a helpful assistant."
|
70 |
self.sep: str = " "
|
71 |
|
72 |
-
@classmethod
|
73 |
-
def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs: Any) -> GmeQwen2VLForVision2Seq:
|
74 |
-
config = kwargs.pop("config", GmeQwen2VLConfig.from_pretrained(pretrained_model_name_or_path, **kwargs))
|
75 |
-
return cls(config, **kwargs)
|
76 |
-
|
77 |
def forward(
|
78 |
self,
|
79 |
input_ids: Optional[torch.LongTensor] = None,
|
@@ -82,9 +71,11 @@ class GmeQwen2VLForVision2Seq(PreTrainedModel):
|
|
82 |
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
83 |
inputs_embeds: Optional[torch.FloatTensor] = None,
|
84 |
pixel_values: Optional[torch.Tensor] = None,
|
|
|
85 |
image_grid_thw: Optional[torch.LongTensor] = None,
|
|
|
86 |
pooling_mask: Optional[torch.LongTensor] = None,
|
87 |
-
**kwargs
|
88 |
) -> torch.Tensor:
|
89 |
if inputs_embeds is None:
|
90 |
inputs_embeds = self.base.model.embed_tokens(input_ids)
|
@@ -93,6 +84,11 @@ class GmeQwen2VLForVision2Seq(PreTrainedModel):
|
|
93 |
image_embeds = self.base.visual(pixel_values, grid_thw=image_grid_thw).to(inputs_embeds.device)
|
94 |
image_mask = input_ids == self.base.config.image_token_id
|
95 |
inputs_embeds[image_mask] = image_embeds
|
|
|
|
|
|
|
|
|
|
|
96 |
if attention_mask is not None:
|
97 |
attention_mask = attention_mask.to(inputs_embeds.device)
|
98 |
|
@@ -105,48 +101,37 @@ class GmeQwen2VLForVision2Seq(PreTrainedModel):
|
|
105 |
)
|
106 |
|
107 |
pooling_mask = attention_mask if pooling_mask is None else pooling_mask
|
108 |
-
left_padding
|
109 |
if left_padding:
|
110 |
embeddings = outputs.last_hidden_state[:, -1]
|
111 |
else:
|
112 |
sequence_lengths = pooling_mask.sum(dim=1) - 1
|
113 |
batch_size = outputs.last_hidden_state.shape[0]
|
114 |
-
embeddings = outputs.last_hidden_state[
|
115 |
-
|
116 |
-
|
117 |
-
]
|
118 |
if self.normalize:
|
119 |
embeddings = torch.nn.functional.normalize(embeddings, p=2, dim=1)
|
120 |
return embeddings.contiguous()
|
121 |
|
122 |
-
|
123 |
-
|
124 |
-
texts: List[str],
|
125 |
-
images: List[Image.Image],
|
126 |
-
is_query: bool = True,
|
127 |
-
instruction: Optional[str] = None,
|
128 |
-
**kwargs: Any,
|
129 |
-
) -> torch.Tensor:
|
130 |
self.base.to(self.device)
|
131 |
-
|
132 |
-
input_images
|
133 |
for t, i in zip(texts, images):
|
134 |
if not is_query or instruction is None:
|
135 |
instruction = self.defualt_instruction
|
136 |
-
input_str
|
137 |
if i is None:
|
138 |
input_images = None # All examples in the same batch are consistent
|
139 |
else:
|
140 |
-
input_str +=
|
141 |
i = fetch_image(i)
|
142 |
input_images.append(i)
|
143 |
if t is not None:
|
144 |
input_str += t
|
145 |
-
msg
|
146 |
-
f"<|im_start|>system\n{instruction}<|im_end|>\n"
|
147 |
-
f"<|im_start|>user\n{input_str}<|im_end|>\n"
|
148 |
-
f"<|im_start|>assistant\n<|endoftext|>"
|
149 |
-
)
|
150 |
input_texts.append(msg)
|
151 |
|
152 |
inputs = self.processor(
|
@@ -155,22 +140,22 @@ class GmeQwen2VLForVision2Seq(PreTrainedModel):
|
|
155 |
padding=True,
|
156 |
truncation=True,
|
157 |
max_length=self.max_length,
|
158 |
-
return_tensors=
|
159 |
)
|
160 |
-
inputs = {k: v.to(self.device) for k, v in inputs.items()}
|
161 |
with torch.no_grad():
|
162 |
embeddings = self.forward(**inputs)
|
163 |
return embeddings
|
164 |
|
165 |
-
def encode(self, sentences:
|
166 |
-
|
167 |
-
return self.embed(texts=sentences, images=[None] * len(sentences), **kwargs)
|
168 |
|
169 |
-
def encode_queries(self, queries: List[str], **kwargs
|
170 |
-
|
|
|
171 |
|
172 |
-
def encode_corpus(self, corpus:
|
173 |
-
if
|
174 |
sentences = [
|
175 |
(corpus["title"][i] + self.sep + corpus["text"][i]).strip()
|
176 |
if "title" in corpus
|
@@ -182,49 +167,56 @@ class GmeQwen2VLForVision2Seq(PreTrainedModel):
|
|
182 |
(doc["title"] + self.sep + doc["text"]).strip() if "title" in doc else doc["text"].strip()
|
183 |
for doc in corpus
|
184 |
]
|
185 |
-
|
|
|
186 |
|
187 |
-
def get_image_embeddings(self, images:
|
188 |
return self.get_fused_embeddings(images=images, **kwargs)
|
189 |
|
190 |
-
def get_text_embeddings(self, texts:
|
191 |
return self.get_fused_embeddings(texts=texts, **kwargs)
|
192 |
|
193 |
-
|
194 |
-
def get_fused_embeddings(
|
195 |
-
self,
|
196 |
-
texts: Optional[List[str]] = None,
|
197 |
-
images: Optional[Union[List[Image.Image], DataLoader]] = None,
|
198 |
-
**kwargs: Any,
|
199 |
-
) -> torch.Tensor:
|
200 |
if isinstance(images, DataLoader):
|
201 |
image_loader = images
|
202 |
batch_size = image_loader.batch_size
|
203 |
image_loader.dataset.transform = None
|
204 |
else:
|
205 |
-
batch_size = kwargs.pop(
|
206 |
if images is None:
|
207 |
-
|
208 |
-
image_loader = [None] * ((len(texts) + batch_size - 1) // batch_size)
|
209 |
else:
|
210 |
-
image_loader =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
211 |
|
212 |
-
|
213 |
-
all_embeddings: List[torch.Tensor] = []
|
214 |
none_batch = [None] * batch_size
|
215 |
-
show_progress_bar
|
216 |
-
pbar = tqdm(total=n_batch, disable=not show_progress_bar, mininterval=1, miniters=10, desc=
|
217 |
for n, img_batch in zip(range(0, n_batch * batch_size, batch_size), image_loader):
|
218 |
-
text_batch
|
219 |
img_batch = none_batch if img_batch is None else img_batch
|
220 |
embeddings = self.embed(texts=text_batch, images=img_batch, **kwargs)
|
221 |
pbar.update(1)
|
222 |
all_embeddings.append(embeddings.cpu())
|
223 |
pbar.close()
|
224 |
-
|
|
|
225 |
|
226 |
-
|
227 |
-
|
228 |
|
229 |
# Utility functions (copied from your vision processing code)
|
230 |
IMAGE_FACTOR: int = 28
|
@@ -309,43 +301,3 @@ def fetch_image(image: Union[str, Image.Image], size_factor: int = IMAGE_FACTOR)
|
|
309 |
)
|
310 |
image = image.resize((resized_width, resized_height))
|
311 |
return image
|
312 |
-
|
313 |
-
|
314 |
-
# # For backward compatibility, you can add a from_pretrained classmethod.
|
315 |
-
# @classmethod
|
316 |
-
# def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs: Any) -> GmeQwen2VLForVision2Seq:
|
317 |
-
# config = GmeQwen2VLConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
|
318 |
-
# return cls(config, **kwargs)
|
319 |
-
|
320 |
-
|
321 |
-
# # Monkey-patch the from_pretrained method to our class so that
|
322 |
-
# # one can load the model with AutoModel.from_pretrained.
|
323 |
-
# GmeQwen2VLForVision2Seq.from_pretrained = from_pretrained.__get__(GmeQwen2VLForVision2Seq)
|
324 |
-
|
325 |
-
|
326 |
-
if __name__ == "__main__":
|
327 |
-
texts = [
|
328 |
-
"What kind of car is this?",
|
329 |
-
"The Tesla Cybertruck is a battery electric pickup truck built by Tesla, Inc. since 2023.",
|
330 |
-
]
|
331 |
-
images = [
|
332 |
-
"https://en.wikipedia.org/wiki/File:Tesla_Cybertruck_damaged_window.jpg",
|
333 |
-
"https://en.wikipedia.org/wiki/File:2024_Tesla_Cybertruck_Foundation_Series,_front_left_(Greenwich).jpg",
|
334 |
-
]
|
335 |
-
|
336 |
-
# You can now load your model with AutoModel as long as your repository's config JSON has the "architectures" field set.
|
337 |
-
model = AutoModel.from_pretrained("Alibaba-NLP/gme-Qwen2-VL-2B-Instruct")
|
338 |
-
# Alternatively, load it directly via our class:
|
339 |
-
# model = GmeQwen2VLForVision2Seq.from_pretrained("Alibaba-NLP/gme-Qwen2-VL-2B-Instruct")
|
340 |
-
|
341 |
-
# Single-modal embedding examples:
|
342 |
-
e_text = model.get_text_embeddings(texts=texts)
|
343 |
-
e_image = model.get_image_embeddings(images=images)
|
344 |
-
print("Text-Image similarity:", (e_text * e_image).sum(-1))
|
345 |
-
# Example with different instruction:
|
346 |
-
e_query = model.get_text_embeddings(texts=texts, instruction="Find an image that matches the given text.")
|
347 |
-
e_corpus = model.get_image_embeddings(images=images, is_query=False)
|
348 |
-
print("Query-Corpus similarity:", (e_query * e_corpus).sum(-1))
|
349 |
-
# Fused-modal embedding:
|
350 |
-
e_fused = model.get_fused_embeddings(texts=texts, images=images)
|
351 |
-
print("Fused-modal similarity:", (e_fused[0] * e_fused[1]).sum())
|
|
|
19 |
AutoProcessor,
|
20 |
PreTrainedModel,
|
21 |
Qwen2VLConfig,
|
22 |
+
Qwen2VLForConditionalGeneration,
|
23 |
)
|
24 |
import os
|
25 |
+
from collections.abc import Iterable
|
26 |
|
|
|
27 |
class GmeQwen2VLConfig(Qwen2VLConfig):
|
28 |
model_type: str = "gme_qwen2_vl"
|
29 |
|
|
|
39 |
self.min_image_tokens = min_image_tokens
|
40 |
self.max_image_tokens = max_image_tokens
|
41 |
self.max_length = max_length
|
|
|
|
|
42 |
|
43 |
|
|
|
44 |
class GmeQwen2VLForVision2Seq(PreTrainedModel):
|
45 |
config_class = GmeQwen2VLConfig
|
46 |
base_model_prefix: str = "base"
|
|
|
48 |
def __init__(self, config: GmeQwen2VLConfig, **kwargs: Any) -> None:
|
49 |
super().__init__(config)
|
50 |
model_name: str = getattr(config, "_name_or_path", "Alibaba-NLP/gme-Qwen2-VL-2B-Instruct")
|
51 |
+
|
52 |
+
self.base = Qwen2VLForConditionalGeneration(config)
|
|
|
|
|
53 |
self.normalize: bool = True
|
|
|
54 |
|
55 |
min_pixels: int = config.min_image_tokens * 28 * 28
|
56 |
max_pixels: int = config.max_image_tokens * 28 * 28
|
57 |
+
|
58 |
self.max_length: int = config.max_length
|
|
|
59 |
self.processor = AutoProcessor.from_pretrained(
|
60 |
model_name, min_pixels=min_pixels, max_pixels=max_pixels, **kwargs
|
61 |
)
|
62 |
+
self.processor.tokenizer.padding_side = 'right'
|
63 |
self.defualt_instruction: str = "You are a helpful assistant."
|
64 |
self.sep: str = " "
|
65 |
|
|
|
|
|
|
|
|
|
|
|
66 |
def forward(
|
67 |
self,
|
68 |
input_ids: Optional[torch.LongTensor] = None,
|
|
|
71 |
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
72 |
inputs_embeds: Optional[torch.FloatTensor] = None,
|
73 |
pixel_values: Optional[torch.Tensor] = None,
|
74 |
+
# pixel_values_videos: Optional[torch.FloatTensor] = None,
|
75 |
image_grid_thw: Optional[torch.LongTensor] = None,
|
76 |
+
# video_grid_thw: Optional[torch.LongTensor] = None,
|
77 |
pooling_mask: Optional[torch.LongTensor] = None,
|
78 |
+
**kwargs
|
79 |
) -> torch.Tensor:
|
80 |
if inputs_embeds is None:
|
81 |
inputs_embeds = self.base.model.embed_tokens(input_ids)
|
|
|
84 |
image_embeds = self.base.visual(pixel_values, grid_thw=image_grid_thw).to(inputs_embeds.device)
|
85 |
image_mask = input_ids == self.base.config.image_token_id
|
86 |
inputs_embeds[image_mask] = image_embeds
|
87 |
+
# if pixel_values_videos is not None:
|
88 |
+
# pixel_values_videos = pixel_values_videos.type(self.base.visual.get_dtype())
|
89 |
+
# video_embeds = self.base.visual(pixel_values_videos, grid_thw=video_grid_thw).to(inputs_embeds.device)
|
90 |
+
# video_mask = input_ids == self.base.config.video_token_id
|
91 |
+
# inputs_embeds[video_mask] = video_embeds
|
92 |
if attention_mask is not None:
|
93 |
attention_mask = attention_mask.to(inputs_embeds.device)
|
94 |
|
|
|
101 |
)
|
102 |
|
103 |
pooling_mask = attention_mask if pooling_mask is None else pooling_mask
|
104 |
+
left_padding = (pooling_mask[:, -1].sum() == pooling_mask.shape[0]) # TODO
|
105 |
if left_padding:
|
106 |
embeddings = outputs.last_hidden_state[:, -1]
|
107 |
else:
|
108 |
sequence_lengths = pooling_mask.sum(dim=1) - 1
|
109 |
batch_size = outputs.last_hidden_state.shape[0]
|
110 |
+
embeddings = outputs.last_hidden_state[torch.arange(
|
111 |
+
batch_size, device=outputs.last_hidden_state.device
|
112 |
+
), sequence_lengths]
|
|
|
113 |
if self.normalize:
|
114 |
embeddings = torch.nn.functional.normalize(embeddings, p=2, dim=1)
|
115 |
return embeddings.contiguous()
|
116 |
|
117 |
+
|
118 |
+
def embed(self, texts: list[str], images: list[Image.Image], is_query=True, instruction=None, **kwargs):
|
|
|
|
|
|
|
|
|
|
|
|
|
119 |
self.base.to(self.device)
|
120 |
+
# Inputs must be batched
|
121 |
+
input_texts, input_images = list(), list()
|
122 |
for t, i in zip(texts, images):
|
123 |
if not is_query or instruction is None:
|
124 |
instruction = self.defualt_instruction
|
125 |
+
input_str = ''
|
126 |
if i is None:
|
127 |
input_images = None # All examples in the same batch are consistent
|
128 |
else:
|
129 |
+
input_str += '<|vision_start|><|image_pad|><|vision_end|>'
|
130 |
i = fetch_image(i)
|
131 |
input_images.append(i)
|
132 |
if t is not None:
|
133 |
input_str += t
|
134 |
+
msg = f'<|im_start|>system\n{instruction}<|im_end|>\n<|im_start|>user\n{input_str}<|im_end|>\n<|im_start|>assistant\n<|endoftext|>'
|
|
|
|
|
|
|
|
|
135 |
input_texts.append(msg)
|
136 |
|
137 |
inputs = self.processor(
|
|
|
140 |
padding=True,
|
141 |
truncation=True,
|
142 |
max_length=self.max_length,
|
143 |
+
return_tensors='pt'
|
144 |
)
|
145 |
+
inputs = {k: v.to(self.device) for k, v in inputs.items()} # TODO
|
146 |
with torch.no_grad():
|
147 |
embeddings = self.forward(**inputs)
|
148 |
return embeddings
|
149 |
|
150 |
+
def encode(self, sentences: list[str], *, prompt_name=None, **kwargs):
|
151 |
+
return self.get_fused_embeddings(texts=sentences, prompt_name=prompt_name, **kwargs)
|
|
|
152 |
|
153 |
+
def encode_queries(self, queries: List[str], **kwargs):
|
154 |
+
embeddings = self.encode(queries, **kwargs)
|
155 |
+
return embeddings
|
156 |
|
157 |
+
def encode_corpus(self, corpus: List[Dict[str, str]], **kwargs):
|
158 |
+
if type(corpus) is dict:
|
159 |
sentences = [
|
160 |
(corpus["title"][i] + self.sep + corpus["text"][i]).strip()
|
161 |
if "title" in corpus
|
|
|
167 |
(doc["title"] + self.sep + doc["text"]).strip() if "title" in doc else doc["text"].strip()
|
168 |
for doc in corpus
|
169 |
]
|
170 |
+
embeddings = self.encode(sentences, is_query=False, **kwargs)
|
171 |
+
return embeddings
|
172 |
|
173 |
+
def get_image_embeddings(self, images: list[Image.Image] | DataLoader, **kwargs):
|
174 |
return self.get_fused_embeddings(images=images, **kwargs)
|
175 |
|
176 |
+
def get_text_embeddings(self, texts: list[str], **kwargs):
|
177 |
return self.get_fused_embeddings(texts=texts, **kwargs)
|
178 |
|
179 |
+
def get_fused_embeddings(self, texts: list[str] = None, images: list[Image.Image] | DataLoader = None, **kwargs):
|
|
|
|
|
|
|
|
|
|
|
|
|
180 |
if isinstance(images, DataLoader):
|
181 |
image_loader = images
|
182 |
batch_size = image_loader.batch_size
|
183 |
image_loader.dataset.transform = None
|
184 |
else:
|
185 |
+
batch_size = kwargs.pop('batch_size', 32)
|
186 |
if images is None:
|
187 |
+
image_loader = None
|
|
|
188 |
else:
|
189 |
+
image_loader = DataLoader(
|
190 |
+
images,
|
191 |
+
batch_size=batch_size,
|
192 |
+
shuffle=False,
|
193 |
+
collate_fn=custom_collate_fn,
|
194 |
+
num_workers=min(math.floor(os.cpu_count() / 2), 8),
|
195 |
+
)
|
196 |
+
|
197 |
+
if texts is None:
|
198 |
+
assert image_loader is not None
|
199 |
+
n_batch = len(image_loader)
|
200 |
+
else:
|
201 |
+
n_batch = len(texts) // batch_size + int(len(texts) % batch_size > 0)
|
202 |
+
image_loader = image_loader or [None] * n_batch
|
203 |
|
204 |
+
all_embeddings = list()
|
|
|
205 |
none_batch = [None] * batch_size
|
206 |
+
show_progress_bar = kwargs.pop('show_progress_bar', True)
|
207 |
+
pbar = tqdm(total=n_batch, disable=not show_progress_bar, mininterval=1, miniters=10, desc='encode')
|
208 |
for n, img_batch in zip(range(0, n_batch * batch_size, batch_size), image_loader):
|
209 |
+
text_batch = none_batch if texts is None else texts[n: n+batch_size]
|
210 |
img_batch = none_batch if img_batch is None else img_batch
|
211 |
embeddings = self.embed(texts=text_batch, images=img_batch, **kwargs)
|
212 |
pbar.update(1)
|
213 |
all_embeddings.append(embeddings.cpu())
|
214 |
pbar.close()
|
215 |
+
all_embeddings = torch.cat(all_embeddings, dim=0)
|
216 |
+
return all_embeddings
|
217 |
|
218 |
+
def custom_collate_fn(batch):
|
219 |
+
return batch
|
220 |
|
221 |
# Utility functions (copied from your vision processing code)
|
222 |
IMAGE_FACTOR: int = 28
|
|
|
301 |
)
|
302 |
image = image.resize((resized_width, resized_height))
|
303 |
return image
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|