Spaces:
Sleeping
Sleeping
liubangwei
commited on
Commit
·
1855cc2
1
Parent(s):
a72a7d4
init IDMR demo
Browse files- app.py +8 -31
- src/collator.py +11 -139
- src/dataset.py +2 -19
- src/loss.py +5 -11
- src/model.py +1 -15
- src/trainer.py +4 -4
- src/vlm_backbone/intern_vl/modeling_internvl_chat.py +0 -36
- src/vlm_backbone/intern_vl/processing_internvl.py +3 -83
app.py
CHANGED
@@ -8,27 +8,25 @@ from transformers import AutoProcessor
|
|
8 |
from src.model import MMEBModel
|
9 |
from src.arguments import ModelArguments
|
10 |
|
11 |
-
# 假设图片库存储在本地文件夹中
|
12 |
QUERY_DIR = "imgs/queries"
|
13 |
IMAGE_DIR = "imgs/candidates"
|
14 |
-
# IMAGE_DIR = "imgs"
|
15 |
image_paths = [os.path.join(IMAGE_DIR, f) for f in os.listdir(IMAGE_DIR) if f.endswith((".jpg", ".png"))]
|
16 |
global IMAGE_TOKEN, TOP_N
|
17 |
IMAGE_TOKEN = "<|image_1|>"
|
18 |
TOP_N = 5
|
19 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
20 |
print(f"device: {device}")
|
21 |
-
|
|
|
22 |
def load_model():
|
23 |
global IMAGE_TOKEN
|
24 |
-
|
25 |
model_args = ModelArguments(
|
26 |
-
# model_name="/fs-computility/ai-shen/kilab-shared/liubangwei/ckpt/IDMR
|
27 |
-
model_name="/
|
28 |
-
model_backbone="internvl_2_5",
|
29 |
)
|
30 |
|
31 |
-
# 加载处理器
|
32 |
if model_args.model_backbone == "phi35v":
|
33 |
processor = AutoProcessor.from_pretrained(
|
34 |
model_args.model_name,
|
@@ -54,14 +52,12 @@ def load_model():
|
|
54 |
)
|
55 |
IMAGE_TOKEN = "<image>"
|
56 |
|
57 |
-
# 加载模型
|
58 |
model = MMEBModel.load(model_args)
|
59 |
model = model.to(device, dtype=torch.bfloat16)
|
60 |
model.eval()
|
61 |
|
62 |
return model, processor
|
63 |
|
64 |
-
# 加载模型和处理器
|
65 |
model, processor = load_model()
|
66 |
|
67 |
def get_inputs(processor, text, image_path=None, image=None):
|
@@ -84,8 +80,6 @@ def get_inputs(processor, text, image_path=None, image=None):
|
|
84 |
del inputs['pixel_values']
|
85 |
return inputs
|
86 |
|
87 |
-
|
88 |
-
# 将图片库中的图像编码为 embedding
|
89 |
def encode_image_library(image_paths):
|
90 |
embeddings = []
|
91 |
for img_path in image_paths:
|
@@ -97,22 +91,18 @@ def encode_image_library(image_paths):
|
|
97 |
embeddings.append(output["tgt_reps"].float().cpu().numpy())
|
98 |
return np.stack(embeddings)
|
99 |
|
100 |
-
# 保存 embedding 到文件
|
101 |
def save_embeddings(embeddings, file_path="image_embeddings.pkl"):
|
102 |
with open(file_path, "wb") as f:
|
103 |
pickle.dump(embeddings, f)
|
104 |
|
105 |
-
# 加载 embedding 从文件
|
106 |
def load_embeddings(file_path="image_embeddings.pkl"):
|
107 |
with open(file_path, "rb") as f:
|
108 |
return pickle.load(f)
|
109 |
|
110 |
-
# 计算相似度(余弦相似度)
|
111 |
def cosine_similarity(query_embedding, embeddings):
|
112 |
similarity = np.sum(query_embedding * embeddings, axis=-1)
|
113 |
return similarity
|
114 |
|
115 |
-
# 检索逻辑
|
116 |
def retrieve_images(query_text, query_image, top_n=TOP_N):
|
117 |
if query_text:
|
118 |
query_text = f"{IMAGE_TOKEN}\n {query_text}"
|
@@ -129,11 +119,8 @@ def retrieve_images(query_text, query_image, top_n=TOP_N):
|
|
129 |
with torch.no_grad(), torch.autocast(device_type=device, dtype=torch.bfloat16):
|
130 |
query_embedding = model(qry=inputs)["qry_reps"].float().cpu().numpy()
|
131 |
|
132 |
-
|
133 |
-
# 加载图片库的 embedding
|
134 |
embeddings = load_embeddings()
|
135 |
|
136 |
-
# 计算相似度
|
137 |
similarity = cosine_similarity(query_embedding, embeddings)
|
138 |
similarity = similarity.T
|
139 |
print(f"cosine_similarity: {similarity}")
|
@@ -145,29 +132,22 @@ def retrieve_images(query_text, query_image, top_n=TOP_N):
|
|
145 |
|
146 |
return [image_paths[i] for i in top_indices]
|
147 |
|
148 |
-
# 界面逻辑
|
149 |
def demo(query_text, query_image):
|
150 |
-
# 执行检索
|
151 |
# print(f"query_text: {query_text}, query_image: {query_image}, type(query_image): {type(query_image)}, image shape: {query_image.shape if query_image is not None else 'None'}")
|
152 |
|
153 |
retrieved_images = retrieve_images(query_text, query_image)
|
154 |
-
# 返回检索结果(图片列表)
|
155 |
return [Image.open(img) for img in retrieved_images]
|
156 |
|
157 |
-
# 预置示例
|
158 |
def load_examples():
|
159 |
examples = []
|
160 |
-
|
161 |
image_files = [f for f in os.listdir(QUERY_DIR) if f.endswith((".jpg", ".png"))]
|
162 |
|
163 |
for img_file in image_files:
|
164 |
-
# 构建图片完整路径
|
165 |
img_path = os.path.join(QUERY_DIR, img_file)
|
166 |
-
# 获取对应的txt文件名(将图片扩展名替换为.txt)
|
167 |
txt_file = os.path.splitext(img_file)[0] + ".txt"
|
168 |
txt_path = os.path.join(QUERY_DIR, txt_file)
|
169 |
|
170 |
-
# 如果存在对应的txt文件,读取查询文本
|
171 |
if os.path.exists(txt_path):
|
172 |
with open(txt_path, 'r', encoding='utf-8') as f:
|
173 |
query_text = f.read().strip().replace("<|image_1|>\n", "")
|
@@ -175,20 +155,17 @@ def load_examples():
|
|
175 |
|
176 |
return examples
|
177 |
|
178 |
-
# 构建 Gradio 界面
|
179 |
iface = gr.Interface(
|
180 |
fn=demo,
|
181 |
inputs=["text", "image"],
|
182 |
outputs=gr.Gallery(label=f"Retrieved Images (Top {TOP_N})"),
|
183 |
-
examples=load_examples(),
|
184 |
title="Multimodal Retrieval Demo",
|
185 |
description="Enter a query and upload an image to retrieve relevant images from the library. You can click on the example below to use it as a query"
|
186 |
)
|
187 |
|
188 |
-
# 在启动时编码图片库并保存 embedding
|
189 |
if not os.path.exists("image_embeddings.pkl"):
|
190 |
embeddings = encode_image_library(image_paths)
|
191 |
save_embeddings(embeddings)
|
192 |
|
193 |
-
# 启动 Gradio 应用
|
194 |
iface.launch()
|
|
|
8 |
from src.model import MMEBModel
|
9 |
from src.arguments import ModelArguments
|
10 |
|
|
|
11 |
QUERY_DIR = "imgs/queries"
|
12 |
IMAGE_DIR = "imgs/candidates"
|
|
|
13 |
image_paths = [os.path.join(IMAGE_DIR, f) for f in os.listdir(IMAGE_DIR) if f.endswith((".jpg", ".png"))]
|
14 |
global IMAGE_TOKEN, TOP_N
|
15 |
IMAGE_TOKEN = "<|image_1|>"
|
16 |
TOP_N = 5
|
17 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
18 |
print(f"device: {device}")
|
19 |
+
|
20 |
+
|
21 |
def load_model():
|
22 |
global IMAGE_TOKEN
|
23 |
+
|
24 |
model_args = ModelArguments(
|
25 |
+
# model_name="/fs-computility/ai-shen/kilab-shared/liubangwei/ckpt/my_hf/IDMR-2B",
|
26 |
+
model_name="lbw18601752667/IDMR-2B",
|
27 |
+
model_backbone="internvl_2_5",
|
28 |
)
|
29 |
|
|
|
30 |
if model_args.model_backbone == "phi35v":
|
31 |
processor = AutoProcessor.from_pretrained(
|
32 |
model_args.model_name,
|
|
|
52 |
)
|
53 |
IMAGE_TOKEN = "<image>"
|
54 |
|
|
|
55 |
model = MMEBModel.load(model_args)
|
56 |
model = model.to(device, dtype=torch.bfloat16)
|
57 |
model.eval()
|
58 |
|
59 |
return model, processor
|
60 |
|
|
|
61 |
model, processor = load_model()
|
62 |
|
63 |
def get_inputs(processor, text, image_path=None, image=None):
|
|
|
80 |
del inputs['pixel_values']
|
81 |
return inputs
|
82 |
|
|
|
|
|
83 |
def encode_image_library(image_paths):
|
84 |
embeddings = []
|
85 |
for img_path in image_paths:
|
|
|
91 |
embeddings.append(output["tgt_reps"].float().cpu().numpy())
|
92 |
return np.stack(embeddings)
|
93 |
|
|
|
94 |
def save_embeddings(embeddings, file_path="image_embeddings.pkl"):
|
95 |
with open(file_path, "wb") as f:
|
96 |
pickle.dump(embeddings, f)
|
97 |
|
|
|
98 |
def load_embeddings(file_path="image_embeddings.pkl"):
|
99 |
with open(file_path, "rb") as f:
|
100 |
return pickle.load(f)
|
101 |
|
|
|
102 |
def cosine_similarity(query_embedding, embeddings):
|
103 |
similarity = np.sum(query_embedding * embeddings, axis=-1)
|
104 |
return similarity
|
105 |
|
|
|
106 |
def retrieve_images(query_text, query_image, top_n=TOP_N):
|
107 |
if query_text:
|
108 |
query_text = f"{IMAGE_TOKEN}\n {query_text}"
|
|
|
119 |
with torch.no_grad(), torch.autocast(device_type=device, dtype=torch.bfloat16):
|
120 |
query_embedding = model(qry=inputs)["qry_reps"].float().cpu().numpy()
|
121 |
|
|
|
|
|
122 |
embeddings = load_embeddings()
|
123 |
|
|
|
124 |
similarity = cosine_similarity(query_embedding, embeddings)
|
125 |
similarity = similarity.T
|
126 |
print(f"cosine_similarity: {similarity}")
|
|
|
132 |
|
133 |
return [image_paths[i] for i in top_indices]
|
134 |
|
|
|
135 |
def demo(query_text, query_image):
|
|
|
136 |
# print(f"query_text: {query_text}, query_image: {query_image}, type(query_image): {type(query_image)}, image shape: {query_image.shape if query_image is not None else 'None'}")
|
137 |
|
138 |
retrieved_images = retrieve_images(query_text, query_image)
|
|
|
139 |
return [Image.open(img) for img in retrieved_images]
|
140 |
|
|
|
141 |
def load_examples():
|
142 |
examples = []
|
143 |
+
|
144 |
image_files = [f for f in os.listdir(QUERY_DIR) if f.endswith((".jpg", ".png"))]
|
145 |
|
146 |
for img_file in image_files:
|
|
|
147 |
img_path = os.path.join(QUERY_DIR, img_file)
|
|
|
148 |
txt_file = os.path.splitext(img_file)[0] + ".txt"
|
149 |
txt_path = os.path.join(QUERY_DIR, txt_file)
|
150 |
|
|
|
151 |
if os.path.exists(txt_path):
|
152 |
with open(txt_path, 'r', encoding='utf-8') as f:
|
153 |
query_text = f.read().strip().replace("<|image_1|>\n", "")
|
|
|
155 |
|
156 |
return examples
|
157 |
|
|
|
158 |
iface = gr.Interface(
|
159 |
fn=demo,
|
160 |
inputs=["text", "image"],
|
161 |
outputs=gr.Gallery(label=f"Retrieved Images (Top {TOP_N})"),
|
162 |
+
examples=load_examples(),
|
163 |
title="Multimodal Retrieval Demo",
|
164 |
description="Enter a query and upload an image to retrieve relevant images from the library. You can click on the example below to use it as a query"
|
165 |
)
|
166 |
|
|
|
167 |
if not os.path.exists("image_embeddings.pkl"):
|
168 |
embeddings = encode_image_library(image_paths)
|
169 |
save_embeddings(embeddings)
|
170 |
|
|
|
171 |
iface.launch()
|
src/collator.py
CHANGED
@@ -19,8 +19,7 @@ class TrainCollator:
|
|
19 |
"""
|
20 |
:param examples: [{qry:..., qry_image:..., pos_text:..., pos_image:...}] * batch_size
|
21 |
"""
|
22 |
-
|
23 |
-
qry_inputs = self._get_batch_inputs(examples, 0, 1) # qry_inputs: {'input_ids': tensor(batch_size, max_len), 'attention_mask': tensor(batch_size, max_len), 'pixel_values': tensor(batch_size, 4, 224, 224), 'image_sizes': tensor(batch_size, 2)}
|
24 |
pos_inputs = self._get_batch_inputs(examples, 2, 3)
|
25 |
if "hard_neg" in self.data_args.dataset_name:
|
26 |
hard_neg_inputs = self._get_batch_inputs(examples, 4, 5)
|
@@ -45,15 +44,15 @@ class TrainCollator:
|
|
45 |
max_length=self.data_args.max_len,
|
46 |
truncation=True
|
47 |
)
|
48 |
-
elif self.model_args.model_backbone in ["qwen", "qwen2_vl"]:
|
49 |
inputs = self.processor(
|
50 |
-
text=[text],
|
51 |
images=[image] if has_image else None,
|
52 |
return_tensors="pt",
|
53 |
max_length=self.data_args.max_len,
|
54 |
truncation=True
|
55 |
)
|
56 |
-
else:
|
57 |
inputs = self.processor(
|
58 |
text=text,
|
59 |
images=[image] if has_image else None,
|
@@ -62,23 +61,19 @@ class TrainCollator:
|
|
62 |
truncation=True
|
63 |
)
|
64 |
|
65 |
-
# 统一输入格式处理
|
66 |
if has_image:
|
67 |
if self.model_args.model_backbone == "qwen":
|
68 |
pixel_values.append(inputs['pixel_values'].unsqueeze(0))
|
69 |
else:
|
70 |
pixel_values.append(inputs['pixel_values'])
|
71 |
|
72 |
-
# 保持维度对齐原始逻辑
|
73 |
input_ids.append(inputs["input_ids"].squeeze(0).unsqueeze(1))
|
74 |
|
75 |
-
# 处理多模态元数据
|
76 |
if "image_sizes" in inputs:
|
77 |
image_sizes.append(inputs['image_sizes'])
|
78 |
if "image_grid_thw" in inputs:
|
79 |
image_grid_thw.append(inputs['image_grid_thw'])
|
80 |
|
81 |
-
# 保持原始填充逻辑
|
82 |
input_ids = torch._C._nn.pad_sequence(
|
83 |
input_ids,
|
84 |
batch_first=True,
|
@@ -87,89 +82,24 @@ class TrainCollator:
|
|
87 |
|
88 |
attention_mask = input_ids.ne(self.processor.tokenizer.pad_token_id)
|
89 |
|
90 |
-
# 构建返回字典
|
91 |
inputs = {
|
92 |
'input_ids': input_ids,
|
93 |
'attention_mask': attention_mask,
|
94 |
-
'image_mask': torch.tensor(image_mask, dtype=torch.float)
|
95 |
}
|
96 |
|
97 |
-
# 处理图像数据
|
98 |
if any(image_mask):
|
99 |
if pixel_values:
|
100 |
inputs['pixel_values'] = torch.cat(pixel_values, dim=0)
|
101 |
-
if image_sizes:
|
102 |
inputs['image_sizes'] = torch.cat(image_sizes, dim=0)
|
103 |
-
if image_grid_thw:
|
104 |
inputs['image_grid_thw'] = torch.cat(image_grid_thw, dim=0)
|
105 |
|
106 |
-
# InternVL专用字段适配
|
107 |
if self.model_args.model_backbone == "internvl_2_5":
|
108 |
-
inputs['image_flags'] = inputs['image_mask'].to(torch.long)
|
109 |
-
# del inputs['image_mask'] # 根据模型接口调整字段名
|
110 |
|
111 |
return inputs
|
112 |
-
"""
|
113 |
-
def _get_batch_inputs(self, examples, text_idx, image_idx):
|
114 |
-
input_ids, pixel_values, image_sizes, image_grid_thw = [], [], [], []
|
115 |
-
image_mask = []
|
116 |
-
image_exist = False
|
117 |
-
for example in examples:
|
118 |
-
text, image = example[text_idx], example[image_idx] # text: str, image: PIL.Image.Image(765*512)
|
119 |
-
if image is None:
|
120 |
-
image_mask.append(0)
|
121 |
-
if self.model_args.model_backbone == "llava_next":
|
122 |
-
inputs = self.processor(images=None, text=text, return_tensors="pt")
|
123 |
-
elif self.model_args.model_backbone == "qwen":
|
124 |
-
inputs = self.processor(text=[text], images=None, return_tensors="pt",
|
125 |
-
max_length=self.data_args.max_len, truncation=True)
|
126 |
-
else: # 'phi', 'internvl'
|
127 |
-
inputs = self.processor(text=text, images=None, return_tensors="pt",
|
128 |
-
max_length=self.data_args.max_len, truncation=True)
|
129 |
-
input_ids.append(inputs["input_ids"].squeeze(0).unsqueeze(1))
|
130 |
-
else:
|
131 |
-
image_mask.append(1)
|
132 |
-
image_exist = True
|
133 |
-
if self.model_args.model_backbone == "llava_next":
|
134 |
-
inputs = self.processor(images=image, text=text, return_tensors="pt")
|
135 |
-
pixel_values.append(inputs['pixel_values'])
|
136 |
-
elif self.model_args.model_backbone == "qwen":
|
137 |
-
inputs = self.processor(text=[text], images=[image], return_tensors="pt",
|
138 |
-
max_length=self.data_args.max_len, truncation=True)
|
139 |
-
pixel_values.append(inputs['pixel_values'].unsqueeze(0))
|
140 |
-
else:
|
141 |
-
inputs = self.processor(text=text, images=[image], return_tensors="pt",
|
142 |
-
max_length=self.data_args.max_len, truncation=True)
|
143 |
-
pixel_values.append(inputs['pixel_values'])
|
144 |
-
input_ids.append(inputs["input_ids"].squeeze(0).unsqueeze(1))
|
145 |
-
if "image_sizes" in inputs:
|
146 |
-
image_sizes.append(inputs['image_sizes'])
|
147 |
-
if "image_grid_thw" in inputs:
|
148 |
-
image_grid_thw.append(inputs['image_grid_thw'])
|
149 |
-
|
150 |
-
input_ids = torch._C._nn.pad_sequence(
|
151 |
-
input_ids, batch_first=True, padding_value=self.processor.tokenizer.pad_token_id
|
152 |
-
).squeeze(2)
|
153 |
-
attention_mask = input_ids.ne(self.processor.tokenizer.pad_token_id)
|
154 |
-
|
155 |
-
inputs = {
|
156 |
-
'input_ids': input_ids,
|
157 |
-
'attention_mask': attention_mask,
|
158 |
-
}
|
159 |
-
if image_exist:
|
160 |
-
inputs['image_mask'] = torch.Tensor(image_mask)
|
161 |
-
pixel_values = torch.cat(pixel_values, dim=0)
|
162 |
-
inputs['pixel_values'] = pixel_values
|
163 |
-
if image_sizes:
|
164 |
-
image_sizes = torch.cat(image_sizes, dim=0)
|
165 |
-
inputs['image_sizes'] = image_sizes
|
166 |
-
elif image_grid_thw:
|
167 |
-
image_grid_thw = torch.cat(image_grid_thw, dim=0)
|
168 |
-
inputs['image_grid_thw'] = image_grid_thw
|
169 |
-
|
170 |
-
return inputs
|
171 |
-
"""
|
172 |
-
|
173 |
|
174 |
@dataclass
|
175 |
class EvalCollator:
|
@@ -183,72 +113,17 @@ class EvalCollator:
|
|
183 |
"""
|
184 |
inputs = self._get_batch_inputs(examples)
|
185 |
return inputs
|
186 |
-
"""
|
187 |
-
def _get_batch_inputs(self, examples):
|
188 |
-
input_ids, pixel_values, image_sizes = [], [], []
|
189 |
-
image_exist = False
|
190 |
-
for example in examples:
|
191 |
-
text, image = example
|
192 |
-
if image is None:
|
193 |
-
if self.model_args.model_backbone == "llava_next":
|
194 |
-
inputs = self.processor(images=None, text=text, return_tensors="pt")
|
195 |
-
else:
|
196 |
-
inputs = self.processor(text, None, return_tensors="pt", max_length=self.data_args.max_len,
|
197 |
-
truncation=True)
|
198 |
-
input_ids.append(inputs["input_ids"].squeeze(0).unsqueeze(1))
|
199 |
-
pixel_values.append(None)
|
200 |
-
image_sizes.append(None)
|
201 |
-
else:
|
202 |
-
image_exist = True
|
203 |
-
if self.model_args.model_backbone == "llava_next":
|
204 |
-
inputs = self.processor(images=image, text=text, return_tensors="pt")
|
205 |
-
else:
|
206 |
-
inputs = self.processor(text, [image], return_tensors="pt", max_length=self.data_args.max_len, truncation=True)
|
207 |
-
input_ids.append(inputs["input_ids"].squeeze(0).unsqueeze(1))
|
208 |
-
pixel_values.append(inputs['pixel_values'])
|
209 |
-
image_sizes.append(inputs['image_sizes'])
|
210 |
|
211 |
-
input_ids = torch._C._nn.pad_sequence(
|
212 |
-
input_ids, batch_first=True, padding_value=self.processor.tokenizer.pad_token_id
|
213 |
-
).squeeze(2)
|
214 |
-
attention_mask = input_ids.ne(self.processor.tokenizer.pad_token_id)
|
215 |
-
|
216 |
-
if not image_exist:
|
217 |
-
dummy_pixel_values = torch.zeros(input_ids.shape[0], 1)
|
218 |
-
dummy_image_sizes = torch.ones(input_ids.shape[0], 1)
|
219 |
-
inputs = {
|
220 |
-
'input_ids': input_ids,
|
221 |
-
'attention_mask': attention_mask,
|
222 |
-
'pixel_values': dummy_pixel_values,
|
223 |
-
'image_sizes': dummy_image_sizes,
|
224 |
-
}
|
225 |
-
else:
|
226 |
-
pixel_values_shape = list(set(v.shape for v in pixel_values if v is not None))[0]
|
227 |
-
pixel_values = [v if v is not None else torch.zeros(pixel_values_shape) for v in pixel_values]
|
228 |
-
pixel_values = torch.cat(pixel_values, dim=0)
|
229 |
-
image_sizes_shape = list(set(v.shape for v in image_sizes if v is not None))[0]
|
230 |
-
image_sizes = [v if v is not None else torch.ones(image_sizes_shape) for v in image_sizes]
|
231 |
-
image_sizes = torch.cat(image_sizes, dim=0)
|
232 |
-
inputs = {
|
233 |
-
'input_ids': input_ids,
|
234 |
-
'attention_mask': attention_mask,
|
235 |
-
'pixel_values': pixel_values,
|
236 |
-
'image_sizes': image_sizes,
|
237 |
-
}
|
238 |
-
|
239 |
-
return inputs
|
240 |
-
"""
|
241 |
def _get_batch_inputs(self, examples):
|
242 |
input_ids, pixel_values, image_sizes = [], [], []
|
243 |
-
image_mask = []
|
244 |
image_exist = False
|
245 |
for example in examples:
|
246 |
text, image = example
|
247 |
-
# print(text, image)
|
248 |
has_image = image is not None
|
249 |
image_mask.append(1 if has_image else 0)
|
250 |
|
251 |
-
if self.model_args.model_backbone == "internvl_2_5":
|
252 |
inputs = self.processor(
|
253 |
text=text,
|
254 |
images=[image] if has_image else None,
|
@@ -289,22 +164,19 @@ class EvalCollator:
|
|
289 |
attention_mask = input_ids.ne(self.processor.tokenizer.pad_token_id)
|
290 |
|
291 |
if self.model_args.model_backbone == "internvl_2_5":
|
292 |
-
# 构建返回字典
|
293 |
inputs = {
|
294 |
'input_ids': input_ids,
|
295 |
'attention_mask': attention_mask,
|
296 |
'image_mask': torch.tensor(image_mask, dtype=torch.float)
|
297 |
}
|
298 |
|
299 |
-
# 处理图像数据
|
300 |
if any(image_mask):
|
301 |
if pixel_values:
|
302 |
inputs['pixel_values'] = torch.cat(pixel_values, dim=0)
|
303 |
if image_sizes:
|
304 |
inputs['image_sizes'] = torch.cat(image_sizes, dim=0)
|
305 |
-
# InternVL专用字段适配
|
306 |
inputs['image_flags'] = inputs['image_mask'].to(torch.long)
|
307 |
-
del inputs['image_mask']
|
308 |
else:
|
309 |
if not image_exist:
|
310 |
dummy_pixel_values = torch.zeros(input_ids.shape[0], 1)
|
|
|
19 |
"""
|
20 |
:param examples: [{qry:..., qry_image:..., pos_text:..., pos_image:...}] * batch_size
|
21 |
"""
|
22 |
+
qry_inputs = self._get_batch_inputs(examples, 0, 1)
|
|
|
23 |
pos_inputs = self._get_batch_inputs(examples, 2, 3)
|
24 |
if "hard_neg" in self.data_args.dataset_name:
|
25 |
hard_neg_inputs = self._get_batch_inputs(examples, 4, 5)
|
|
|
44 |
max_length=self.data_args.max_len,
|
45 |
truncation=True
|
46 |
)
|
47 |
+
elif self.model_args.model_backbone in ["qwen", "qwen2_vl"]:
|
48 |
inputs = self.processor(
|
49 |
+
text=[text],
|
50 |
images=[image] if has_image else None,
|
51 |
return_tensors="pt",
|
52 |
max_length=self.data_args.max_len,
|
53 |
truncation=True
|
54 |
)
|
55 |
+
else:
|
56 |
inputs = self.processor(
|
57 |
text=text,
|
58 |
images=[image] if has_image else None,
|
|
|
61 |
truncation=True
|
62 |
)
|
63 |
|
|
|
64 |
if has_image:
|
65 |
if self.model_args.model_backbone == "qwen":
|
66 |
pixel_values.append(inputs['pixel_values'].unsqueeze(0))
|
67 |
else:
|
68 |
pixel_values.append(inputs['pixel_values'])
|
69 |
|
|
|
70 |
input_ids.append(inputs["input_ids"].squeeze(0).unsqueeze(1))
|
71 |
|
|
|
72 |
if "image_sizes" in inputs:
|
73 |
image_sizes.append(inputs['image_sizes'])
|
74 |
if "image_grid_thw" in inputs:
|
75 |
image_grid_thw.append(inputs['image_grid_thw'])
|
76 |
|
|
|
77 |
input_ids = torch._C._nn.pad_sequence(
|
78 |
input_ids,
|
79 |
batch_first=True,
|
|
|
82 |
|
83 |
attention_mask = input_ids.ne(self.processor.tokenizer.pad_token_id)
|
84 |
|
|
|
85 |
inputs = {
|
86 |
'input_ids': input_ids,
|
87 |
'attention_mask': attention_mask,
|
88 |
+
'image_mask': torch.tensor(image_mask, dtype=torch.float)
|
89 |
}
|
90 |
|
|
|
91 |
if any(image_mask):
|
92 |
if pixel_values:
|
93 |
inputs['pixel_values'] = torch.cat(pixel_values, dim=0)
|
94 |
+
if image_sizes:
|
95 |
inputs['image_sizes'] = torch.cat(image_sizes, dim=0)
|
96 |
+
if image_grid_thw:
|
97 |
inputs['image_grid_thw'] = torch.cat(image_grid_thw, dim=0)
|
98 |
|
|
|
99 |
if self.model_args.model_backbone == "internvl_2_5":
|
100 |
+
inputs['image_flags'] = inputs['image_mask'].to(torch.long)
|
|
|
101 |
|
102 |
return inputs
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
103 |
|
104 |
@dataclass
|
105 |
class EvalCollator:
|
|
|
113 |
"""
|
114 |
inputs = self._get_batch_inputs(examples)
|
115 |
return inputs
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
116 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
117 |
def _get_batch_inputs(self, examples):
|
118 |
input_ids, pixel_values, image_sizes = [], [], []
|
119 |
+
image_mask = []
|
120 |
image_exist = False
|
121 |
for example in examples:
|
122 |
text, image = example
|
|
|
123 |
has_image = image is not None
|
124 |
image_mask.append(1 if has_image else 0)
|
125 |
|
126 |
+
if self.model_args.model_backbone == "internvl_2_5":
|
127 |
inputs = self.processor(
|
128 |
text=text,
|
129 |
images=[image] if has_image else None,
|
|
|
164 |
attention_mask = input_ids.ne(self.processor.tokenizer.pad_token_id)
|
165 |
|
166 |
if self.model_args.model_backbone == "internvl_2_5":
|
|
|
167 |
inputs = {
|
168 |
'input_ids': input_ids,
|
169 |
'attention_mask': attention_mask,
|
170 |
'image_mask': torch.tensor(image_mask, dtype=torch.float)
|
171 |
}
|
172 |
|
|
|
173 |
if any(image_mask):
|
174 |
if pixel_values:
|
175 |
inputs['pixel_values'] = torch.cat(pixel_values, dim=0)
|
176 |
if image_sizes:
|
177 |
inputs['image_sizes'] = torch.cat(image_sizes, dim=0)
|
|
|
178 |
inputs['image_flags'] = inputs['image_mask'].to(torch.long)
|
179 |
+
del inputs['image_mask']
|
180 |
else:
|
181 |
if not image_exist:
|
182 |
dummy_pixel_values = torch.zeros(input_ids.shape[0], 1)
|
src/dataset.py
CHANGED
@@ -8,18 +8,8 @@ from PIL import Image
|
|
8 |
import os
|
9 |
from torchvision.transforms import RandAugment
|
10 |
|
11 |
-
|
12 |
def get_randaugment_transform(n=2, m=9):
|
13 |
-
"""
|
14 |
-
创建 RandAugment 增强器。
|
15 |
-
|
16 |
-
参数:
|
17 |
-
- n: 每次随机选择的增强操作数量。
|
18 |
-
- m: 每种增强操作的强度。
|
19 |
-
|
20 |
-
返回:
|
21 |
-
- RandAugment 对象。
|
22 |
-
"""
|
23 |
return RandAugment(num_ops=n, magnitude=m)
|
24 |
|
25 |
|
@@ -39,7 +29,7 @@ class TrainDataset(Dataset):
|
|
39 |
self.model_args = model_args
|
40 |
self.transform = None
|
41 |
if self.data_args.randaugment:
|
42 |
-
self.transform = get_randaugment_transform()
|
43 |
train_data = []
|
44 |
|
45 |
if data_args.subset_name is not None:
|
@@ -103,13 +93,6 @@ class TrainDataset(Dataset):
|
|
103 |
return image
|
104 |
|
105 |
def __getitem__(self, item) -> Tuple[str, List[str]]:
|
106 |
-
# qry_text, qry_image_path, pos_text, pos_image_path = (
|
107 |
-
# self.train_data[item]["qry"], self.train_data[item]["qry_image_path"],
|
108 |
-
# self.train_data[item]["pos_text"], self.train_data[item]["pos_image_path"],
|
109 |
-
# )
|
110 |
-
|
111 |
-
# return (qry_text, self._get_image(qry_image_path),
|
112 |
-
# pos_text, self._get_image(pos_image_path))
|
113 |
|
114 |
data_item = self.train_data[item]
|
115 |
qry_text, qry_image_path, pos_text, pos_image_path = (
|
|
|
8 |
import os
|
9 |
from torchvision.transforms import RandAugment
|
10 |
|
11 |
+
|
12 |
def get_randaugment_transform(n=2, m=9):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
return RandAugment(num_ops=n, magnitude=m)
|
14 |
|
15 |
|
|
|
29 |
self.model_args = model_args
|
30 |
self.transform = None
|
31 |
if self.data_args.randaugment:
|
32 |
+
self.transform = get_randaugment_transform()
|
33 |
train_data = []
|
34 |
|
35 |
if data_args.subset_name is not None:
|
|
|
93 |
return image
|
94 |
|
95 |
def __getitem__(self, item) -> Tuple[str, List[str]]:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
96 |
|
97 |
data_item = self.train_data[item]
|
98 |
qry_text, qry_image_path, pos_text, pos_image_path = (
|
src/loss.py
CHANGED
@@ -51,7 +51,7 @@ class HardNegativeContrastiveLoss:
|
|
51 |
# y: positive embeddings
|
52 |
# z: negative embeddings (optional)
|
53 |
|
54 |
-
if z is None:
|
55 |
target_per_qry = y.size(0) // x.size(0)
|
56 |
target = torch.arange(
|
57 |
0, x.size(0) * target_per_qry, target_per_qry,
|
@@ -60,18 +60,12 @@ class HardNegativeContrastiveLoss:
|
|
60 |
loss = F.cross_entropy(logits / self.temperature, target, reduction=reduction)
|
61 |
return loss
|
62 |
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
neg_logits = torch.matmul(x, z.transpose(0, 1)) # [batch_size, num_negs]
|
67 |
|
68 |
-
# 将正负样本的相似度拼接在一起
|
69 |
-
logits = torch.cat([pos_logits, neg_logits], dim=1) # [batch_size, batch_size + num_negs]
|
70 |
-
|
71 |
-
# 创建目标标签(正样本的索引)
|
72 |
target = torch.arange(x.size(0), device=x.device)
|
73 |
-
|
74 |
-
# 计算交叉熵损失
|
75 |
loss = F.cross_entropy(logits / self.temperature, target, reduction=reduction)
|
76 |
return loss
|
77 |
|
|
|
51 |
# y: positive embeddings
|
52 |
# z: negative embeddings (optional)
|
53 |
|
54 |
+
if z is None:
|
55 |
target_per_qry = y.size(0) // x.size(0)
|
56 |
target = torch.arange(
|
57 |
0, x.size(0) * target_per_qry, target_per_qry,
|
|
|
60 |
loss = F.cross_entropy(logits / self.temperature, target, reduction=reduction)
|
61 |
return loss
|
62 |
|
63 |
+
pos_logits = torch.matmul(x, y.transpose(0, 1))
|
64 |
+
neg_logits = torch.matmul(x, z.transpose(0, 1))
|
65 |
+
logits = torch.cat([pos_logits, neg_logits], dim=1)
|
|
|
66 |
|
|
|
|
|
|
|
|
|
67 |
target = torch.arange(x.size(0), device=x.device)
|
68 |
+
|
|
|
69 |
loss = F.cross_entropy(logits / self.temperature, target, reduction=reduction)
|
70 |
return loss
|
71 |
|
src/model.py
CHANGED
@@ -118,20 +118,6 @@ class MMEBModel(nn.Module):
|
|
118 |
trust_remote_code=True)
|
119 |
base_model.padding_side = "right"
|
120 |
|
121 |
-
# # Print all model parameters
|
122 |
-
# import json
|
123 |
-
# import os
|
124 |
-
|
125 |
-
# param_info = {}
|
126 |
-
# for name, param in base_model.named_parameters():
|
127 |
-
# param_info[name] = {
|
128 |
-
# "shape": list(param.shape),
|
129 |
-
# "requires_grad": param.requires_grad
|
130 |
-
# }
|
131 |
-
|
132 |
-
# with open('./model_parameters.json', 'w') as f:
|
133 |
-
# json.dump(param_info, f, indent=4)
|
134 |
-
# import pdb; pdb.set_trace()
|
135 |
if model_args.lora:
|
136 |
if lora_target_modules is None:
|
137 |
lora_target_modules = model_args.lora_target_modules.split(',')
|
@@ -192,7 +178,7 @@ class MMEBModel(nn.Module):
|
|
192 |
trust_remote_code=True
|
193 |
)
|
194 |
config = InternVLChatConfig.from_pretrained(model_args.model_name)
|
195 |
-
# config.vision_config.image_size = data_args.force_image_size
|
196 |
config.use_flash_attn = False
|
197 |
base_model = InternVLChatModel.from_pretrained(
|
198 |
model_args.model_name,
|
|
|
118 |
trust_remote_code=True)
|
119 |
base_model.padding_side = "right"
|
120 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
121 |
if model_args.lora:
|
122 |
if lora_target_modules is None:
|
123 |
lora_target_modules = model_args.lora_target_modules.split(',')
|
|
|
178 |
trust_remote_code=True
|
179 |
)
|
180 |
config = InternVLChatConfig.from_pretrained(model_args.model_name)
|
181 |
+
# config.vision_config.image_size = data_args.force_image_size
|
182 |
config.use_flash_attn = False
|
183 |
base_model = InternVLChatModel.from_pretrained(
|
184 |
model_args.model_name,
|
src/trainer.py
CHANGED
@@ -87,11 +87,11 @@ def split_vlm_inputs(model_input: dict, chunk_size: int):
|
|
87 |
if "image_grid_thw" in keys:
|
88 |
image_grid_thw = arg_val["image_grid_thw"]
|
89 |
chunked_tensors.append(torch.split(image_grid_thw, chunk_image_count))
|
90 |
-
|
91 |
if "image_flags" in keys:
|
92 |
image_flags = arg_val["image_flags"]
|
93 |
chunked_tensors.append(torch.split(image_flags, chunk_size))
|
94 |
-
keys.remove("image_flags")
|
95 |
|
96 |
|
97 |
chunked_arg_val = []
|
@@ -148,7 +148,7 @@ class GradCacheTrainer(Trainer):
|
|
148 |
|
149 |
def training_step(self, model, inputs, *args, **kwargs) -> torch.Tensor:
|
150 |
model.train()
|
151 |
-
|
152 |
if self.args.hard_neg:
|
153 |
queries, passages, negatives = inputs
|
154 |
queries, passages, negatives = {'qry': queries}, {'tgt': passages}, {'neg': negatives}
|
@@ -165,7 +165,7 @@ class GradCacheTrainer(Trainer):
|
|
165 |
print(f"neg_img.shape={negatives['neg']['pixel_values'].shape}")
|
166 |
|
167 |
_distributed = self.args.local_rank > -1
|
168 |
-
self.gc.models = [model, model, model]
|
169 |
loss = self.gc(queries, passages, negatives, no_sync_except_last=_distributed)
|
170 |
else:
|
171 |
queries, passages = inputs
|
|
|
87 |
if "image_grid_thw" in keys:
|
88 |
image_grid_thw = arg_val["image_grid_thw"]
|
89 |
chunked_tensors.append(torch.split(image_grid_thw, chunk_image_count))
|
90 |
+
|
91 |
if "image_flags" in keys:
|
92 |
image_flags = arg_val["image_flags"]
|
93 |
chunked_tensors.append(torch.split(image_flags, chunk_size))
|
94 |
+
keys.remove("image_flags")
|
95 |
|
96 |
|
97 |
chunked_arg_val = []
|
|
|
148 |
|
149 |
def training_step(self, model, inputs, *args, **kwargs) -> torch.Tensor:
|
150 |
model.train()
|
151 |
+
|
152 |
if self.args.hard_neg:
|
153 |
queries, passages, negatives = inputs
|
154 |
queries, passages, negatives = {'qry': queries}, {'tgt': passages}, {'neg': negatives}
|
|
|
165 |
print(f"neg_img.shape={negatives['neg']['pixel_values'].shape}")
|
166 |
|
167 |
_distributed = self.args.local_rank > -1
|
168 |
+
self.gc.models = [model, model, model]
|
169 |
loss = self.gc(queries, passages, negatives, no_sync_except_last=_distributed)
|
170 |
else:
|
171 |
queries, passages = inputs
|
src/vlm_backbone/intern_vl/modeling_internvl_chat.py
CHANGED
@@ -172,53 +172,17 @@ class InternVLChatModel(PreTrainedModel):
|
|
172 |
loss_reduction_all_gather: Optional[bool] = False,
|
173 |
) -> Union[Tuple, CausalLMOutputWithPast]:
|
174 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
175 |
-
# import pdb; pdb.set_trace()
|
176 |
-
# 获取原始batch size和每个样本的序列长度
|
177 |
B, N = input_ids.shape
|
178 |
input_embeds = self.language_model.get_input_embeddings()(input_ids).clone() # [B, N, C]
|
179 |
|
180 |
if pixel_values is not None:
|
181 |
vit_embeds = self.extract_feature(pixel_values) # [num_images, num_patches, C]
|
182 |
-
|
183 |
-
# 找到input_ids中需要替换的图片token位置
|
184 |
selected = torch.eq(input_ids, self.img_context_token_id) # [B, N]
|
185 |
|
186 |
-
# 确保image_flags维度正确
|
187 |
image_flags = image_flags.squeeze(-1) # [B]
|
188 |
|
189 |
-
# # 记录两种方法的时间
|
190 |
-
# import time
|
191 |
-
|
192 |
-
# # 方法1: 循环替换
|
193 |
-
# start_time1 = time.time()
|
194 |
-
# input_embeds2 = input_embeds.clone()
|
195 |
-
# vit_idx = 0
|
196 |
-
# for i in range(B):
|
197 |
-
# if image_flags[i] == 1:
|
198 |
-
# sample_selected = selected[i]
|
199 |
-
# input_embeds2[i, sample_selected] = input_embeds2[i, sample_selected] * 0.0 + vit_embeds[vit_idx]
|
200 |
-
# vit_idx += 1
|
201 |
-
# time1 = time.time() - start_time1
|
202 |
-
|
203 |
-
# 方法2: 向量化替换
|
204 |
-
# start_time2 = time.time()
|
205 |
mask = selected & (image_flags.unsqueeze(-1)) == 1
|
206 |
input_embeds[mask] = vit_embeds.reshape(-1, vit_embeds.shape[-1])
|
207 |
-
# time2 = time.time() - start_time2
|
208 |
-
|
209 |
-
# print(f"循环替换用时: {time1:.6f}秒")
|
210 |
-
# print(f"向量化替换用时: {time2:.6f}秒")
|
211 |
-
# print(f"向量化方法比循环方法快 {time1/time2:.2f}倍")
|
212 |
-
|
213 |
-
# print(f"input_ids.shape = {input_ids.shape}") # [B, N]
|
214 |
-
# print(f"input_embeds.shape = {input_embeds.shape}") # [B, N, C]
|
215 |
-
# print(f"pixel_values.shape = {pixel_values.shape}") # [num_images, ...]
|
216 |
-
# print(f"vit_embeds.shape = {vit_embeds.shape}") # [num_images, num_patches, C]
|
217 |
-
# print(f"image_flags.sum() = {image_flags.sum()}") # 应该等于num_images
|
218 |
-
|
219 |
-
# print(torch.allclose(input_embeds2, input_embeds, rtol=1e-7))
|
220 |
-
# assert torch.allclose(input_embeds2, input_embeds, rtol=1e-5), "input_embeds2 and input_embeds should have the same values"
|
221 |
-
|
222 |
|
223 |
outputs = self.language_model(
|
224 |
inputs_embeds=input_embeds,
|
|
|
172 |
loss_reduction_all_gather: Optional[bool] = False,
|
173 |
) -> Union[Tuple, CausalLMOutputWithPast]:
|
174 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
|
|
|
|
175 |
B, N = input_ids.shape
|
176 |
input_embeds = self.language_model.get_input_embeddings()(input_ids).clone() # [B, N, C]
|
177 |
|
178 |
if pixel_values is not None:
|
179 |
vit_embeds = self.extract_feature(pixel_values) # [num_images, num_patches, C]
|
|
|
|
|
180 |
selected = torch.eq(input_ids, self.img_context_token_id) # [B, N]
|
181 |
|
|
|
182 |
image_flags = image_flags.squeeze(-1) # [B]
|
183 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
184 |
mask = selected & (image_flags.unsqueeze(-1)) == 1
|
185 |
input_embeds[mask] = vit_embeds.reshape(-1, vit_embeds.shape[-1])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
186 |
|
187 |
outputs = self.language_model(
|
188 |
inputs_embeds=input_embeds,
|
src/vlm_backbone/intern_vl/processing_internvl.py
CHANGED
@@ -11,70 +11,6 @@ IMG_START_TOKEN = "<img>"
|
|
11 |
IMG_END_TOKEN = "</img>"
|
12 |
IMG_CONTEXT_TOKEN = "<IMG_CONTEXT>"
|
13 |
|
14 |
-
# class InternVLProcessor(ProcessorMixin):
|
15 |
-
# attributes = ["image_processor", "tokenizer"]
|
16 |
-
# image_processor_class = "AutoImageProcessor"
|
17 |
-
# tokenizer_class = "AutoTokenizer"
|
18 |
-
|
19 |
-
# def __init__(self, image_processor, tokenizer, num_img_tokens=256):
|
20 |
-
# super().__init__(image_processor, tokenizer)
|
21 |
-
# self.num_img_tokens = num_img_tokens
|
22 |
-
# self._add_special_tokens()
|
23 |
-
|
24 |
-
# def _add_special_tokens(self):
|
25 |
-
# special_tokens = [IMG_START_TOKEN, IMG_END_TOKEN, IMG_CONTEXT_TOKEN]
|
26 |
-
# self.tokenizer.add_special_tokens({"additional_special_tokens": special_tokens})
|
27 |
-
|
28 |
-
# def __call__(
|
29 |
-
# self,
|
30 |
-
# text: Union[TextInput, List[TextInput]] = None,
|
31 |
-
# images: ImageInput = None,
|
32 |
-
# padding: Union[bool, str, PaddingStrategy] = False,
|
33 |
-
# truncation: Union[bool, str, TruncationStrategy] = None,
|
34 |
-
# max_length: Optional[int] = None,
|
35 |
-
# return_tensors: Optional[str] = "pt",
|
36 |
-
# ) -> BatchFeature:
|
37 |
-
|
38 |
-
# # Process images
|
39 |
-
# pixel_values = []
|
40 |
-
# if images is not None:
|
41 |
-
# image_inputs = self.image_processor(images, return_tensors=return_tensors)
|
42 |
-
# pixel_values = image_inputs.pixel_values
|
43 |
-
|
44 |
-
# # Process text with image tokens
|
45 |
-
# processed_text = self._insert_image_tokens(text, num_images=len(pixel_values))
|
46 |
-
|
47 |
-
# # Tokenize text
|
48 |
-
# text_inputs = self.tokenizer(
|
49 |
-
# processed_text,
|
50 |
-
# padding=padding,
|
51 |
-
# truncation=truncation,
|
52 |
-
# max_length=max_length,
|
53 |
-
# return_tensors=return_tensors,
|
54 |
-
# add_special_tokens=False
|
55 |
-
# )
|
56 |
-
|
57 |
-
# # Build final inputs
|
58 |
-
# inputs = BatchFeature(data={
|
59 |
-
# **text_inputs,
|
60 |
-
# "pixel_values": pixel_values,
|
61 |
-
# })
|
62 |
-
|
63 |
-
# return inputs
|
64 |
-
|
65 |
-
# def _insert_image_tokens(self, text: str, num_images: int) -> str:
|
66 |
-
# """Replace <image> tags with image context tokens"""
|
67 |
-
# image_tokens = []
|
68 |
-
# for _ in range(num_images):
|
69 |
-
# image_tokens.append(
|
70 |
-
# f"{IMG_START_TOKEN}{IMG_CONTEXT_TOKEN * self.num_img_tokens}{IMG_END_TOKEN}"
|
71 |
-
# )
|
72 |
-
|
73 |
-
# # Replace the first N occurrences of <image>
|
74 |
-
# pattern = re.compile(r"<image>")
|
75 |
-
# return pattern.sub(lambda x: image_tokens.pop(0) if image_tokens else "", text, count=num_images)
|
76 |
-
|
77 |
-
|
78 |
class InternVLProcessor(ProcessorMixin):
|
79 |
attributes = ["image_processor", "tokenizer"]
|
80 |
image_processor_class = "AutoImageProcessor"
|
@@ -91,8 +27,7 @@ class InternVLProcessor(ProcessorMixin):
|
|
91 |
num_added = self.tokenizer.add_special_tokens({
|
92 |
"additional_special_tokens": special_tokens
|
93 |
})
|
94 |
-
|
95 |
-
# assert num_added == 1, f"Failed to add IMG_CONTEXT token, added {num_added}"
|
96 |
|
97 |
def __call__(
|
98 |
self,
|
@@ -103,38 +38,25 @@ class InternVLProcessor(ProcessorMixin):
|
|
103 |
max_length: Optional[int] = None,
|
104 |
return_tensors: str = "pt"
|
105 |
) -> BatchFeature:
|
106 |
-
# import pdb; pdb.set_trace()
|
107 |
-
|
108 |
-
# 处理单样本输入
|
109 |
if isinstance(text, str):
|
110 |
text = [text]
|
111 |
|
112 |
if not isinstance(images, list):
|
113 |
images = [images] if images else []
|
114 |
|
115 |
-
# 生成image_flags
|
116 |
image_flags = [1] if len(images) else [0]
|
117 |
|
118 |
-
# 图像预处理
|
119 |
pixel_values = []
|
120 |
if any(image_flags):
|
121 |
pixel_values = self.image_processor(
|
122 |
-
[img for img in images if img],
|
123 |
return_tensors=return_tensors
|
124 |
-
).pixel_values
|
125 |
|
126 |
-
# 文本预处理
|
127 |
processed_texts = [
|
128 |
self._insert_image_tokens(t, count)
|
129 |
for t, count in zip(text, image_flags)
|
130 |
]
|
131 |
-
# print("process text:")
|
132 |
-
# print(processed_texts)
|
133 |
-
# print("text")
|
134 |
-
# print(text)
|
135 |
-
# print(images)
|
136 |
-
# print(image_flags)
|
137 |
-
# Tokenize文本
|
138 |
text_inputs = self.tokenizer(
|
139 |
processed_texts,
|
140 |
padding=padding,
|
@@ -144,7 +66,6 @@ class InternVLProcessor(ProcessorMixin):
|
|
144 |
add_special_tokens=True
|
145 |
)
|
146 |
|
147 |
-
# 构建最终输入
|
148 |
return BatchFeature({
|
149 |
**text_inputs,
|
150 |
"pixel_values": pixel_values,
|
@@ -152,7 +73,6 @@ class InternVLProcessor(ProcessorMixin):
|
|
152 |
}, tensor_type=return_tensors)
|
153 |
|
154 |
def _insert_image_tokens(self, text: str, image_count: int) -> str:
|
155 |
-
"""动态插入图像token"""
|
156 |
if image_count == 0:
|
157 |
return text
|
158 |
|
|
|
11 |
IMG_END_TOKEN = "</img>"
|
12 |
IMG_CONTEXT_TOKEN = "<IMG_CONTEXT>"
|
13 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
class InternVLProcessor(ProcessorMixin):
|
15 |
attributes = ["image_processor", "tokenizer"]
|
16 |
image_processor_class = "AutoImageProcessor"
|
|
|
27 |
num_added = self.tokenizer.add_special_tokens({
|
28 |
"additional_special_tokens": special_tokens
|
29 |
})
|
30 |
+
|
|
|
31 |
|
32 |
def __call__(
|
33 |
self,
|
|
|
38 |
max_length: Optional[int] = None,
|
39 |
return_tensors: str = "pt"
|
40 |
) -> BatchFeature:
|
|
|
|
|
|
|
41 |
if isinstance(text, str):
|
42 |
text = [text]
|
43 |
|
44 |
if not isinstance(images, list):
|
45 |
images = [images] if images else []
|
46 |
|
|
|
47 |
image_flags = [1] if len(images) else [0]
|
48 |
|
|
|
49 |
pixel_values = []
|
50 |
if any(image_flags):
|
51 |
pixel_values = self.image_processor(
|
52 |
+
[img for img in images if img],
|
53 |
return_tensors=return_tensors
|
54 |
+
).pixel_values
|
55 |
|
|
|
56 |
processed_texts = [
|
57 |
self._insert_image_tokens(t, count)
|
58 |
for t, count in zip(text, image_flags)
|
59 |
]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
60 |
text_inputs = self.tokenizer(
|
61 |
processed_texts,
|
62 |
padding=padding,
|
|
|
66 |
add_special_tokens=True
|
67 |
)
|
68 |
|
|
|
69 |
return BatchFeature({
|
70 |
**text_inputs,
|
71 |
"pixel_values": pixel_values,
|
|
|
73 |
}, tensor_type=return_tensors)
|
74 |
|
75 |
def _insert_image_tokens(self, text: str, image_count: int) -> str:
|
|
|
76 |
if image_count == 0:
|
77 |
return text
|
78 |
|