Spaces:
Sleeping
Sleeping
Commit
·
c16cb4b
1
Parent(s):
d1c139b
:tada: first commit
Browse files- app.py +169 -0
- requirements.txt +6 -0
app.py
ADDED
@@ -0,0 +1,169 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import einops
|
3 |
+
import matplotlib.pyplot as plt
|
4 |
+
from torchvision.transforms import ToPILImage
|
5 |
+
from PIL import Image
|
6 |
+
import os
|
7 |
+
import math
|
8 |
+
|
9 |
+
from transformers import AutoTokenizer, AutoImageProcessor, VisionEncoderDecoderModel
|
10 |
+
import gradio as gr
|
11 |
+
from concurrent.futures import ThreadPoolExecutor
|
12 |
+
|
13 |
+
############################## RATIONAL BEHIND ###############################
|
14 |
+
|
15 |
+
# Load the model, tokenizer, and image processor with error handling
|
16 |
+
def load_model_and_components(model_name):
|
17 |
+
model = VisionEncoderDecoderModel.from_pretrained(model_name)
|
18 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
19 |
+
image_processor = AutoImageProcessor.from_pretrained(model_name)
|
20 |
+
return model, tokenizer, image_processor
|
21 |
+
|
22 |
+
# Preload both models in parallel
|
23 |
+
def preload_models():
|
24 |
+
models = {}
|
25 |
+
model_names = ["laicsiifes/swin-distilbertimbau"] #, "laicsiifes/swin-gportuguese-2"]
|
26 |
+
with ThreadPoolExecutor() as executor:
|
27 |
+
results = executor.map(load_model_and_components, model_names)
|
28 |
+
for name, result in zip(model_names, results):
|
29 |
+
models[name] = result
|
30 |
+
return models
|
31 |
+
|
32 |
+
models = preload_models()
|
33 |
+
|
34 |
+
# Predefined images for selection
|
35 |
+
image_folder = "images"
|
36 |
+
predefined_images = [
|
37 |
+
Image.open(os.path.join(image_folder, fname)).convert("RGB")
|
38 |
+
for fname in os.listdir(image_folder)
|
39 |
+
if fname.lower().endswith(('.png', '.jpg', '.jpeg', '.gif', '.bmp', '.ppm'))
|
40 |
+
]
|
41 |
+
|
42 |
+
# Function to preprocess the image to RGB format
|
43 |
+
def preprocess_image(image):
|
44 |
+
if image is None:
|
45 |
+
return None, None
|
46 |
+
pil_image = image.convert("RGB")
|
47 |
+
return pil_image, None
|
48 |
+
|
49 |
+
# Function to process the image in tokens with its attention maps
|
50 |
+
def get_attn_map(model, image, processor, tokenizer):
|
51 |
+
pixel_values = processor(image, return_tensors="pt").pixel_values
|
52 |
+
model.eval()
|
53 |
+
with torch.no_grad():
|
54 |
+
output = model.generate(
|
55 |
+
pixel_values=pixel_values,
|
56 |
+
return_dict_in_generate=True,
|
57 |
+
output_hidden_states=True,
|
58 |
+
output_attentions=True,
|
59 |
+
max_length=25,
|
60 |
+
num_beams=5
|
61 |
+
)
|
62 |
+
|
63 |
+
last_layers = [tensor_tuple[-1] for tensor_tuple in output.cross_attentions]
|
64 |
+
attention_maps = torch.stack(last_layers, dim=0)
|
65 |
+
attention_maps = einops.reduce(
|
66 |
+
attention_maps,
|
67 |
+
'token batch head sequence (height width) -> token sequence (height width)',
|
68 |
+
height=7, width=7,
|
69 |
+
reduction='mean'
|
70 |
+
)
|
71 |
+
|
72 |
+
tokens = output.sequences[0]
|
73 |
+
token_texts = tokenizer.convert_ids_to_tokens(tokens)
|
74 |
+
valid_token_texts = token_texts[1:]
|
75 |
+
|
76 |
+
return valid_token_texts, attention_maps, output
|
77 |
+
|
78 |
+
# Function to preprocess the captions tokens and attention maps
|
79 |
+
# e.g. tokens `sent` and `##ada` yield the word `sentada`
|
80 |
+
def join_tokens(text_tokens, attention_maps, connect_symbol='##'):
|
81 |
+
tokens = text_tokens.copy()
|
82 |
+
attn_map = attention_maps.detach().clone()
|
83 |
+
|
84 |
+
i = 0
|
85 |
+
while i < len(tokens) and tokens[i] != '[SEP]':
|
86 |
+
if tokens[i].startswith(connect_symbol):
|
87 |
+
tokens[i] = tokens[i - 1] + tokens[i].replace(connect_symbol, '')
|
88 |
+
tokens.pop(i - 1)
|
89 |
+
attn_map[i][0] = attn_map[i - 1][0] + attn_map[i][0]
|
90 |
+
attn_map = torch.cat((attn_map[:i - 1], attn_map[i:]), dim=0)
|
91 |
+
i -= 1
|
92 |
+
i += 1
|
93 |
+
|
94 |
+
tokens = tokens[1:i - 1]
|
95 |
+
attn_map = attn_map[1:i - 1]
|
96 |
+
|
97 |
+
return tokens, attn_map
|
98 |
+
|
99 |
+
# Make the attention maps visually organized and presentable
|
100 |
+
def generate_attention_gallery(image, selected_model):
|
101 |
+
if image is None:
|
102 |
+
return []
|
103 |
+
|
104 |
+
model, tokenizer, processor = models[selected_model]
|
105 |
+
tokens, attention_maps, _ = get_attn_map(model, image, processor, tokenizer)
|
106 |
+
joined_tokens, joined_attn_maps = join_tokens(tokens, attention_maps)
|
107 |
+
|
108 |
+
grid_size = int(joined_attn_maps.size(-1) ** 0.5)
|
109 |
+
gallery_output = []
|
110 |
+
|
111 |
+
for i, token in enumerate(joined_tokens):
|
112 |
+
att_map = joined_attn_maps[i].view(grid_size, grid_size)
|
113 |
+
att_map = (att_map - att_map.min()) / (att_map.max() - att_map.min())
|
114 |
+
|
115 |
+
att_map = att_map.repeat_interleave(32, dim=0).repeat_interleave(32, dim=1)
|
116 |
+
|
117 |
+
att_map_resized = ToPILImage()(
|
118 |
+
att_map.unsqueeze(0).repeat(3, 1, 1)
|
119 |
+
).resize(image.size[::])
|
120 |
+
|
121 |
+
blended = Image.blend(image, att_map_resized, alpha=0.75)
|
122 |
+
gallery_output.append((blended, token))
|
123 |
+
|
124 |
+
return gallery_output
|
125 |
+
|
126 |
+
################################### PAGE ####################################
|
127 |
+
|
128 |
+
# Define UI
|
129 |
+
with gr.Blocks(theme=gr.themes.Citrus(primary_hue="blue", secondary_hue="orange")) as interface:
|
130 |
+
gr.Markdown("""
|
131 |
+
# Welcome to the LAICSI-IFES Vision Encoder-Decoder Demo
|
132 |
+
---
|
133 |
+
### Select a pretrained model and upload an image to visualize attention maps.
|
134 |
+
""")
|
135 |
+
|
136 |
+
with gr.Row(variant='panel'):
|
137 |
+
model_selector = gr.Dropdown(
|
138 |
+
choices=list(models.keys()),
|
139 |
+
value="laicsiifes/swin-distilbertimbau",
|
140 |
+
label="Select Model"
|
141 |
+
)
|
142 |
+
|
143 |
+
gr.Markdown("""---\n### Upload or select an image and click 'Generate' to view attention maps.""")
|
144 |
+
|
145 |
+
with gr.Row(variant='panel'):
|
146 |
+
with gr.Column():
|
147 |
+
image_display = gr.Image(type="pil", label="Image Preview", image_mode="RGB", height=400)
|
148 |
+
with gr.Column():
|
149 |
+
output_gallery = gr.Gallery(label="Attention Maps", columns=4, rows=3, height=600)
|
150 |
+
generate_button = gr.Button("Generate")
|
151 |
+
|
152 |
+
gr.Markdown("""---""")
|
153 |
+
|
154 |
+
with gr.Row(variant='panel'):
|
155 |
+
examples = gr.Examples(
|
156 |
+
examples=predefined_images,
|
157 |
+
fn=preprocess_image,
|
158 |
+
inputs=[image_display],
|
159 |
+
outputs=[image_display, output_gallery],
|
160 |
+
label="Examples"
|
161 |
+
)
|
162 |
+
|
163 |
+
# Actions
|
164 |
+
model_selector.change(fn=lambda: (None, []), outputs=[image_display, output_gallery])
|
165 |
+
image_display.upload(fn=preprocess_image, inputs=[image_display], outputs=[image_display, output_gallery])
|
166 |
+
image_display.clear(fn=lambda: None, outputs=[output_gallery])
|
167 |
+
generate_button.click(fn=generate_attention_gallery, inputs=[image_display, model_selector], outputs=output_gallery)
|
168 |
+
|
169 |
+
interface.launch(share=False)
|
requirements.txt
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
transformers==4.33.0
|
2 |
+
Pillow==9.5.0
|
3 |
+
requests==2.31.0
|
4 |
+
gradio==3.29.0
|
5 |
+
torch==2.0.1
|
6 |
+
numpy==1.26.4
|