Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -1,3 +1,222 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import os
|
2 |
import argparse
|
3 |
import numpy as np
|
@@ -23,12 +242,9 @@ CHECKPOINTS = {
|
|
23 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
24 |
current_vis = []
|
25 |
current_bpe = []
|
26 |
-
current_index = 0
|
27 |
-
|
28 |
|
29 |
def load_model(check_type):
|
30 |
-
|
31 |
-
device = torch.device("cuda")
|
32 |
if check_type == 'R50':
|
33 |
tokenizer = load_tokenizer('tokenizer_path')
|
34 |
model = build_model(argparse.Namespace()).eval()
|
@@ -55,7 +271,7 @@ def load_model(check_type):
|
|
55 |
return model.to(device), tokenizer, transform, device
|
56 |
|
57 |
def process_image(model, tokenizer, transform, device, check_type, image, text):
|
58 |
-
global current_vis, current_bpe
|
59 |
src_size = image.size
|
60 |
if 'TokenOCR' in check_type:
|
61 |
images, target_ratio = dynamic_preprocess(image, min_num=1, max_num=12,
|
@@ -80,10 +296,6 @@ def process_image(model, tokenizer, transform, device, check_type, image, text):
|
|
80 |
text_embeds = model.tok_embeddings(input_ids)
|
81 |
|
82 |
vit_embeds, size1 = model.forward_tokenocr(pixel_values.to(torch.bfloat16).to(device))
|
83 |
-
print("vit_embeds",vit_embeds)
|
84 |
-
print("vit_embeds,shape",vit_embeds.shape)
|
85 |
-
print("target_ratio",target_ratio)
|
86 |
-
print("check_type",check_type)
|
87 |
vit_embeds, size2 = post_process(vit_embeds, target_ratio, check_type)
|
88 |
|
89 |
# 计算相似度
|
@@ -92,43 +304,20 @@ def process_image(model, tokenizer, transform, device, check_type, image, text):
|
|
92 |
similarity = text_embeds @ vit_embeds.T
|
93 |
resized_size = size1 if size1 is not None else size2
|
94 |
|
95 |
-
# print(f"text_embeds shape: {text_embeds.shape}, numel: {text_embeds.numel()}") # text_embeds shape: torch.Size([4, 2048]), numel: 8192
|
96 |
-
# print(f"vit_embeds shape: {vit_embeds.shape}, numel: {vit_embeds.numel()}") # vit_embeds shape: torch.Size([9728, 2048]), numel: 19922944
|
97 |
-
# print(f"similarity shape: {similarity.shape}, numel: {similarity.numel()}")# similarity shape: torch.Size([4, 9728]), numel: 38912
|
98 |
-
|
99 |
-
|
100 |
# 生成可视化
|
101 |
attn_map = similarity.reshape(len(text_embeds), resized_size[0], resized_size[1])
|
102 |
-
# attn_map = similarity.reshape(len(text_embeds), *target_ratio)
|
103 |
all_bpe_strings = [tokenizer.decode(input_id) for input_id in input_ids]
|
104 |
current_vis = generate_similiarity_map([image], attn_map,
|
105 |
[tokenizer.decode([i]) for i in input_ids],
|
106 |
[], target_ratio, src_size)
|
107 |
|
108 |
current_bpe = [tokenizer.decode([i]) for i in input_ids]
|
109 |
-
# current_bpe[-1] = 'Input text'
|
110 |
current_bpe[-1] = text
|
111 |
-
|
112 |
-
print("current_bpe",len(current_bpe))
|
113 |
-
return image, current_vis[0], current_bpe[0]
|
114 |
-
|
115 |
-
# 事件处理函数
|
116 |
-
def update_index(change):
|
117 |
-
global current_vis, current_bpe, current_index
|
118 |
-
current_index = max(0, min(len(current_vis) - 1, current_index + change))
|
119 |
-
return current_vis[current_index], format_bpe_display(current_bpe[current_index])
|
120 |
|
121 |
def format_bpe_display(bpe):
|
122 |
-
# 使用HTML标签来设置字体大小、颜色,加粗,并居中
|
123 |
return f"<div style='text-align:center; font-size:20px;'><strong>Current BPE: <span style='color:red;'>{bpe}</span></strong></div>"
|
124 |
|
125 |
-
def update_slider_index(x):
|
126 |
-
print(f"x: {x}, current_vis length: {len(current_vis)}, current_bpe length: {len(current_bpe)}")
|
127 |
-
if 0 <= x < len(current_vis) and 0 <= x < len(current_bpe):
|
128 |
-
return current_vis[x], format_bpe_display(current_bpe[x])
|
129 |
-
else:
|
130 |
-
return None, "索引超出范围"
|
131 |
-
|
132 |
# Gradio界面
|
133 |
with gr.Blocks(title="BPE Visualization Demo") as demo:
|
134 |
gr.Markdown("## BPE Visualization Demo - TokenFD基座模型能力可视化")
|
@@ -138,7 +327,7 @@ with gr.Blocks(title="BPE Visualization Demo") as demo:
|
|
138 |
model_type = gr.Dropdown(
|
139 |
choices=["TokenFD_4096_English_seg", "TokenFD_2048_Bilingual_seg", "R50", "R50_siglip"],
|
140 |
label="Select model type",
|
141 |
-
value="TokenOCR_4096_English_seg"
|
142 |
)
|
143 |
image_input = gr.Image(label="Upload images", type="pil")
|
144 |
text_input = gr.Textbox(label="Input text")
|
@@ -162,57 +351,24 @@ with gr.Blocks(title="BPE Visualization Demo") as demo:
|
|
162 |
orig_img = gr.Image(label="Original picture", interactive=False)
|
163 |
heatmap = gr.Image(label="BPE visualization", interactive=False)
|
164 |
|
165 |
-
with gr.Row() as controls:
|
166 |
-
prev_btn = gr.Button("⬅ Last", visible=False)
|
167 |
-
index_slider = gr.Slider(0, 1, value=0, step=1, label="BPE index", visible=False)
|
168 |
-
next_btn = gr.Button("⮕ Next", visible=False)
|
169 |
-
|
170 |
bpe_display = gr.Markdown("Current BPE: ", visible=False)
|
171 |
|
172 |
# 事件处理
|
173 |
@spaces.GPU
|
174 |
def on_run_clicked(model_type, image, text):
|
175 |
-
global current_vis, current_bpe
|
176 |
-
current_index = 0 # Reset index when new image is processed
|
177 |
image, vis, bpe = process_image(*load_model(model_type), model_type, image, text)
|
178 |
-
# Update the slider range and set value to 0
|
179 |
-
slider_max_val = len(current_bpe) - 1
|
180 |
bpe_text = format_bpe_display(bpe)
|
181 |
-
|
182 |
-
print("current_bpe",len(current_bpe))
|
183 |
-
return image, vis, bpe_text, slider_max_val
|
184 |
|
185 |
run_btn.click(
|
186 |
on_run_clicked,
|
187 |
inputs=[model_type, image_input, text_input],
|
188 |
-
outputs=[orig_img, heatmap, bpe_display
|
189 |
).then(
|
190 |
-
lambda
|
191 |
-
|
192 |
-
outputs=[prev_btn, index_slider, next_btn, bpe_display],
|
193 |
)
|
194 |
-
|
195 |
-
prev_btn.click(
|
196 |
-
lambda: (*update_index(-1), current_index),
|
197 |
-
outputs=[heatmap, bpe_display, index_slider]
|
198 |
-
)
|
199 |
-
|
200 |
-
next_btn.click(
|
201 |
-
lambda: (*update_index(1), current_index),
|
202 |
-
outputs=[heatmap, bpe_display, index_slider]
|
203 |
-
)
|
204 |
-
|
205 |
-
# index_slider.change(
|
206 |
-
# lambda x: (current_vis[x], format_bpe_display(current_bpe[x])) if 0<=x<len(current_vis else (None,"Invaild")
|
207 |
-
# inputs=index_slider,
|
208 |
-
# outputs=[heatmap, bpe_display]
|
209 |
-
# )
|
210 |
-
|
211 |
-
index_slider.change(
|
212 |
-
update_slider_index,
|
213 |
-
inputs=index_slider,
|
214 |
-
outputs=[heatmap, bpe_display]
|
215 |
-
)
|
216 |
|
217 |
if __name__ == "__main__":
|
218 |
demo.launch()
|
|
|
1 |
+
# import os
|
2 |
+
# import argparse
|
3 |
+
# import numpy as np
|
4 |
+
# from PIL import Image
|
5 |
+
# import torch
|
6 |
+
# import torchvision.transforms as T
|
7 |
+
# from transformers import AutoTokenizer
|
8 |
+
# import gradio as gr
|
9 |
+
# from resnet50 import build_model
|
10 |
+
# from utils import generate_similiarity_map, post_process, load_tokenizer, build_transform_R50
|
11 |
+
# from utils import IMAGENET_MEAN, IMAGENET_STD
|
12 |
+
# from internvl.train.dataset import dynamic_preprocess
|
13 |
+
# from internvl.model.internvl_chat import InternVLChatModel
|
14 |
+
# import spaces
|
15 |
+
|
16 |
+
# # 模型配置
|
17 |
+
# CHECKPOINTS = {
|
18 |
+
# "TokenFD_4096_English_seg": "TongkunGuan/TokenFD_4096_English_seg",
|
19 |
+
# "TokenFD_2048_Bilingual_seg": "TongkunGuan/TokenFD_2048_Bilingual_seg",
|
20 |
+
# }
|
21 |
+
|
22 |
+
# # 全局变量
|
23 |
+
# HF_TOKEN = os.getenv("HF_TOKEN")
|
24 |
+
# current_vis = []
|
25 |
+
# current_bpe = []
|
26 |
+
# current_index = 0
|
27 |
+
|
28 |
+
|
29 |
+
# def load_model(check_type):
|
30 |
+
# # device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
31 |
+
# device = torch.device("cuda")
|
32 |
+
# if check_type == 'R50':
|
33 |
+
# tokenizer = load_tokenizer('tokenizer_path')
|
34 |
+
# model = build_model(argparse.Namespace()).eval()
|
35 |
+
# model.load_state_dict(torch.load(CHECKPOINTS['R50'], map_location='cpu')['model'])
|
36 |
+
# transform = build_transform_R50(normalize_type='imagenet')
|
37 |
+
|
38 |
+
# elif check_type == 'R50_siglip':
|
39 |
+
# tokenizer = load_tokenizer('tokenizer_path')
|
40 |
+
# model = build_model(argparse.Namespace()).eval()
|
41 |
+
# model.load_state_dict(torch.load(CHECKPOINTS['R50_siglip'], map_location='cpu')['model'])
|
42 |
+
# transform = build_transform_R50(normalize_type='imagenet')
|
43 |
+
|
44 |
+
# elif 'TokenFD' in check_type:
|
45 |
+
# model_path = CHECKPOINTS[check_type]
|
46 |
+
# tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True, use_fast=False, use_auth_token=HF_TOKEN)
|
47 |
+
# model = InternVLChatModel.from_pretrained(model_path, torch_dtype=torch.bfloat16).eval()
|
48 |
+
# transform = T.Compose([
|
49 |
+
# T.Lambda(lambda img: img.convert('RGB')),
|
50 |
+
# T.Resize((224, 224)),
|
51 |
+
# T.ToTensor(),
|
52 |
+
# T.Normalize(IMAGENET_MEAN, IMAGENET_STD)
|
53 |
+
# ])
|
54 |
+
|
55 |
+
# return model.to(device), tokenizer, transform, device
|
56 |
+
|
57 |
+
# def process_image(model, tokenizer, transform, device, check_type, image, text):
|
58 |
+
# global current_vis, current_bpe, current_index
|
59 |
+
# src_size = image.size
|
60 |
+
# if 'TokenOCR' in check_type:
|
61 |
+
# images, target_ratio = dynamic_preprocess(image, min_num=1, max_num=12,
|
62 |
+
# image_size=model.config.force_image_size,
|
63 |
+
# use_thumbnail=model.config.use_thumbnail,
|
64 |
+
# return_ratio=True)
|
65 |
+
# pixel_values = torch.stack([transform(img) for img in images]).to(device)
|
66 |
+
# else:
|
67 |
+
# pixel_values = torch.stack([transform(image)]).to(device)
|
68 |
+
# target_ratio = (1, 1)
|
69 |
+
|
70 |
+
# # 文本处理
|
71 |
+
# text += ' '
|
72 |
+
# input_ids = tokenizer(text)['input_ids'][1:]
|
73 |
+
# input_ids = torch.tensor(input_ids, device=device)
|
74 |
+
|
75 |
+
# # 获取嵌入
|
76 |
+
# with torch.no_grad():
|
77 |
+
# if 'R50' in check_type:
|
78 |
+
# text_embeds = model.language_embedding(input_ids)
|
79 |
+
# else:
|
80 |
+
# text_embeds = model.tok_embeddings(input_ids)
|
81 |
+
|
82 |
+
# vit_embeds, size1 = model.forward_tokenocr(pixel_values.to(torch.bfloat16).to(device))
|
83 |
+
# print("vit_embeds",vit_embeds)
|
84 |
+
# print("vit_embeds,shape",vit_embeds.shape)
|
85 |
+
# print("target_ratio",target_ratio)
|
86 |
+
# print("check_type",check_type)
|
87 |
+
# vit_embeds, size2 = post_process(vit_embeds, target_ratio, check_type)
|
88 |
+
|
89 |
+
# # 计算相似度
|
90 |
+
# text_embeds = text_embeds / text_embeds.norm(dim=-1, keepdim=True)
|
91 |
+
# vit_embeds = vit_embeds / vit_embeds.norm(dim=-1, keepdim=True)
|
92 |
+
# similarity = text_embeds @ vit_embeds.T
|
93 |
+
# resized_size = size1 if size1 is not None else size2
|
94 |
+
|
95 |
+
# # print(f"text_embeds shape: {text_embeds.shape}, numel: {text_embeds.numel()}") # text_embeds shape: torch.Size([4, 2048]), numel: 8192
|
96 |
+
# # print(f"vit_embeds shape: {vit_embeds.shape}, numel: {vit_embeds.numel()}") # vit_embeds shape: torch.Size([9728, 2048]), numel: 19922944
|
97 |
+
# # print(f"similarity shape: {similarity.shape}, numel: {similarity.numel()}")# similarity shape: torch.Size([4, 9728]), numel: 38912
|
98 |
+
|
99 |
+
|
100 |
+
# # 生成可视化
|
101 |
+
# attn_map = similarity.reshape(len(text_embeds), resized_size[0], resized_size[1])
|
102 |
+
# # attn_map = similarity.reshape(len(text_embeds), *target_ratio)
|
103 |
+
# all_bpe_strings = [tokenizer.decode(input_id) for input_id in input_ids]
|
104 |
+
# current_vis = generate_similiarity_map([image], attn_map,
|
105 |
+
# [tokenizer.decode([i]) for i in input_ids],
|
106 |
+
# [], target_ratio, src_size)
|
107 |
+
|
108 |
+
# current_bpe = [tokenizer.decode([i]) for i in input_ids]
|
109 |
+
# # current_bpe[-1] = 'Input text'
|
110 |
+
# current_bpe[-1] = text
|
111 |
+
# print("current_vis",len(current_vis))
|
112 |
+
# print("current_bpe",len(current_bpe))
|
113 |
+
# return image, current_vis[0], current_bpe[0]
|
114 |
+
|
115 |
+
# # 事件处理函数
|
116 |
+
# def update_index(change):
|
117 |
+
# global current_vis, current_bpe, current_index
|
118 |
+
# current_index = max(0, min(len(current_vis) - 1, current_index + change))
|
119 |
+
# return current_vis[current_index], format_bpe_display(current_bpe[current_index])
|
120 |
+
|
121 |
+
# def format_bpe_display(bpe):
|
122 |
+
# # 使用HTML标签来设置字体大小、颜色,加粗,并居中
|
123 |
+
# return f"<div style='text-align:center; font-size:20px;'><strong>Current BPE: <span style='color:red;'>{bpe}</span></strong></div>"
|
124 |
+
|
125 |
+
# def update_slider_index(x):
|
126 |
+
# print(f"x: {x}, current_vis length: {len(current_vis)}, current_bpe length: {len(current_bpe)}")
|
127 |
+
# if 0 <= x < len(current_vis) and 0 <= x < len(current_bpe):
|
128 |
+
# return current_vis[x], format_bpe_display(current_bpe[x])
|
129 |
+
# else:
|
130 |
+
# return None, "索引超出范围"
|
131 |
+
|
132 |
+
# # Gradio界面
|
133 |
+
# with gr.Blocks(title="BPE Visualization Demo") as demo:
|
134 |
+
# gr.Markdown("## BPE Visualization Demo - TokenFD基座模型能力可视化")
|
135 |
+
|
136 |
+
# with gr.Row():
|
137 |
+
# with gr.Column(scale=0.5):
|
138 |
+
# model_type = gr.Dropdown(
|
139 |
+
# choices=["TokenFD_4096_English_seg", "TokenFD_2048_Bilingual_seg", "R50", "R50_siglip"],
|
140 |
+
# label="Select model type",
|
141 |
+
# value="TokenOCR_4096_English_seg" # 设置默认值为第一个选项
|
142 |
+
# )
|
143 |
+
# image_input = gr.Image(label="Upload images", type="pil")
|
144 |
+
# text_input = gr.Textbox(label="Input text")
|
145 |
+
|
146 |
+
# run_btn = gr.Button("RUN")
|
147 |
+
|
148 |
+
# gr.Examples(
|
149 |
+
# examples=[
|
150 |
+
# [os.path.join("examples", "examples0.jpg"), "Veterans and Benefits"],
|
151 |
+
# [os.path.join("examples", "examples1.jpg"), "Refreshers"],
|
152 |
+
# [os.path.join("examples", "examples2.png"), "Vision Transformer"]
|
153 |
+
# ],
|
154 |
+
# inputs=[image_input, text_input],
|
155 |
+
# label="Sample input"
|
156 |
+
# )
|
157 |
+
|
158 |
+
# with gr.Column(scale=2):
|
159 |
+
# gr.Markdown("<p style='font-size:20px;'><span style='color:red;'>If the input text is not included in the image</span>, the attention map will show a lot of noise (the actual response value is very low), since we normalize the attention map according to the relative value.</p>")
|
160 |
+
|
161 |
+
# with gr.Row():
|
162 |
+
# orig_img = gr.Image(label="Original picture", interactive=False)
|
163 |
+
# heatmap = gr.Image(label="BPE visualization", interactive=False)
|
164 |
+
|
165 |
+
# with gr.Row() as controls:
|
166 |
+
# prev_btn = gr.Button("⬅ Last", visible=False)
|
167 |
+
# index_slider = gr.Slider(0, 1, value=0, step=1, label="BPE index", visible=False)
|
168 |
+
# next_btn = gr.Button("⮕ Next", visible=False)
|
169 |
+
|
170 |
+
# bpe_display = gr.Markdown("Current BPE: ", visible=False)
|
171 |
+
|
172 |
+
# # 事件处理
|
173 |
+
# @spaces.GPU
|
174 |
+
# def on_run_clicked(model_type, image, text):
|
175 |
+
# global current_vis, current_bpe, current_index
|
176 |
+
# current_index = 0 # Reset index when new image is processed
|
177 |
+
# image, vis, bpe = process_image(*load_model(model_type), model_type, image, text)
|
178 |
+
# # Update the slider range and set value to 0
|
179 |
+
# slider_max_val = len(current_bpe) - 1
|
180 |
+
# bpe_text = format_bpe_display(bpe)
|
181 |
+
# print("current_vis",len(current_vis))
|
182 |
+
# print("current_bpe",len(current_bpe))
|
183 |
+
# return image, vis, bpe_text, slider_max_val
|
184 |
+
|
185 |
+
# run_btn.click(
|
186 |
+
# on_run_clicked,
|
187 |
+
# inputs=[model_type, image_input, text_input],
|
188 |
+
# outputs=[orig_img, heatmap, bpe_display, index_slider],
|
189 |
+
# ).then(
|
190 |
+
# lambda max_val: (gr.update(visible=True), gr.update(visible=True, maximum=max_val, value=0), gr.update(visible=True), gr.update(visible=True)),
|
191 |
+
# inputs=index_slider,
|
192 |
+
# outputs=[prev_btn, index_slider, next_btn, bpe_display],
|
193 |
+
# )
|
194 |
+
|
195 |
+
# prev_btn.click(
|
196 |
+
# lambda: (*update_index(-1), current_index),
|
197 |
+
# outputs=[heatmap, bpe_display, index_slider]
|
198 |
+
# )
|
199 |
+
|
200 |
+
# next_btn.click(
|
201 |
+
# lambda: (*update_index(1), current_index),
|
202 |
+
# outputs=[heatmap, bpe_display, index_slider]
|
203 |
+
# )
|
204 |
+
|
205 |
+
# # index_slider.change(
|
206 |
+
# # lambda x: (current_vis[x], format_bpe_display(current_bpe[x])) if 0<=x<len(current_vis else (None,"Invaild")
|
207 |
+
# # inputs=index_slider,
|
208 |
+
# # outputs=[heatmap, bpe_display]
|
209 |
+
# # )
|
210 |
+
|
211 |
+
# index_slider.change(
|
212 |
+
# update_slider_index,
|
213 |
+
# inputs=index_slider,
|
214 |
+
# outputs=[heatmap, bpe_display]
|
215 |
+
# )
|
216 |
+
|
217 |
+
# if __name__ == "__main__":
|
218 |
+
# demo.launch()
|
219 |
+
|
220 |
import os
|
221 |
import argparse
|
222 |
import numpy as np
|
|
|
242 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
243 |
current_vis = []
|
244 |
current_bpe = []
|
|
|
|
|
245 |
|
246 |
def load_model(check_type):
|
247 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
248 |
if check_type == 'R50':
|
249 |
tokenizer = load_tokenizer('tokenizer_path')
|
250 |
model = build_model(argparse.Namespace()).eval()
|
|
|
271 |
return model.to(device), tokenizer, transform, device
|
272 |
|
273 |
def process_image(model, tokenizer, transform, device, check_type, image, text):
|
274 |
+
global current_vis, current_bpe
|
275 |
src_size = image.size
|
276 |
if 'TokenOCR' in check_type:
|
277 |
images, target_ratio = dynamic_preprocess(image, min_num=1, max_num=12,
|
|
|
296 |
text_embeds = model.tok_embeddings(input_ids)
|
297 |
|
298 |
vit_embeds, size1 = model.forward_tokenocr(pixel_values.to(torch.bfloat16).to(device))
|
|
|
|
|
|
|
|
|
299 |
vit_embeds, size2 = post_process(vit_embeds, target_ratio, check_type)
|
300 |
|
301 |
# 计算相似度
|
|
|
304 |
similarity = text_embeds @ vit_embeds.T
|
305 |
resized_size = size1 if size1 is not None else size2
|
306 |
|
|
|
|
|
|
|
|
|
|
|
307 |
# 生成可视化
|
308 |
attn_map = similarity.reshape(len(text_embeds), resized_size[0], resized_size[1])
|
|
|
309 |
all_bpe_strings = [tokenizer.decode(input_id) for input_id in input_ids]
|
310 |
current_vis = generate_similiarity_map([image], attn_map,
|
311 |
[tokenizer.decode([i]) for i in input_ids],
|
312 |
[], target_ratio, src_size)
|
313 |
|
314 |
current_bpe = [tokenizer.decode([i]) for i in input_ids]
|
|
|
315 |
current_bpe[-1] = text
|
316 |
+
return image, current_vis, current_bpe
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
317 |
|
318 |
def format_bpe_display(bpe):
|
|
|
319 |
return f"<div style='text-align:center; font-size:20px;'><strong>Current BPE: <span style='color:red;'>{bpe}</span></strong></div>"
|
320 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
321 |
# Gradio界面
|
322 |
with gr.Blocks(title="BPE Visualization Demo") as demo:
|
323 |
gr.Markdown("## BPE Visualization Demo - TokenFD基座模型能力可视化")
|
|
|
327 |
model_type = gr.Dropdown(
|
328 |
choices=["TokenFD_4096_English_seg", "TokenFD_2048_Bilingual_seg", "R50", "R50_siglip"],
|
329 |
label="Select model type",
|
330 |
+
value="TokenOCR_4096_English_seg"
|
331 |
)
|
332 |
image_input = gr.Image(label="Upload images", type="pil")
|
333 |
text_input = gr.Textbox(label="Input text")
|
|
|
351 |
orig_img = gr.Image(label="Original picture", interactive=False)
|
352 |
heatmap = gr.Image(label="BPE visualization", interactive=False)
|
353 |
|
|
|
|
|
|
|
|
|
|
|
354 |
bpe_display = gr.Markdown("Current BPE: ", visible=False)
|
355 |
|
356 |
# 事件处理
|
357 |
@spaces.GPU
|
358 |
def on_run_clicked(model_type, image, text):
|
359 |
+
global current_vis, current_bpe
|
|
|
360 |
image, vis, bpe = process_image(*load_model(model_type), model_type, image, text)
|
|
|
|
|
361 |
bpe_text = format_bpe_display(bpe)
|
362 |
+
return image, vis[0], bpe_text
|
|
|
|
|
363 |
|
364 |
run_btn.click(
|
365 |
on_run_clicked,
|
366 |
inputs=[model_type, image_input, text_input],
|
367 |
+
outputs=[orig_img, heatmap, bpe_display],
|
368 |
).then(
|
369 |
+
lambda: (gr.update(visible=True)),
|
370 |
+
outputs=[bpe_display],
|
|
|
371 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
372 |
|
373 |
if __name__ == "__main__":
|
374 |
demo.launch()
|