Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,104 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import io
|
5 |
+
from PIL import Image
|
6 |
+
from transformers import (
|
7 |
+
AutoImageProcessor,
|
8 |
+
AutoTokenizer,
|
9 |
+
AutoModelForCausalLM,
|
10 |
+
)
|
11 |
+
import numpy as np
|
12 |
+
model_root = "qihoo360/fg-clip-base"
|
13 |
+
|
14 |
+
model = AutoModelForCausalLM.from_pretrained(model_root,trust_remote_code=True)
|
15 |
+
device = model.device
|
16 |
+
tokenizer = AutoTokenizer.from_pretrained(model_root)
|
17 |
+
image_processor = AutoImageProcessor.from_pretrained(model_root)
|
18 |
+
|
19 |
+
import math
|
20 |
+
import matplotlib
|
21 |
+
matplotlib.use('Agg')
|
22 |
+
import matplotlib.pyplot as plt
|
23 |
+
|
24 |
+
|
25 |
+
|
26 |
+
|
27 |
+
|
28 |
+
|
29 |
+
def Get_Densefeature(image, candidate_labels):
|
30 |
+
"""
|
31 |
+
Takes an image and a comma-separated string of candidate labels,
|
32 |
+
and returns the classification scores.
|
33 |
+
"""
|
34 |
+
candidate_labels = [label.lstrip(" ") for label in candidate_labels.split(",") if label !=""]
|
35 |
+
# print(candidate_labels)
|
36 |
+
|
37 |
+
image_size=224
|
38 |
+
image = image.convert("RGB")
|
39 |
+
image = image.resize((image_size,image_size))
|
40 |
+
image_input = image_processor.preprocess(image, return_tensors='pt')['pixel_values'].to(device)
|
41 |
+
|
42 |
+
with torch.no_grad():
|
43 |
+
dense_image_feature = model.get_image_dense_features(image_input)
|
44 |
+
captions = [candidate_labels[0]]
|
45 |
+
caption_input = torch.tensor(tokenizer(captions, max_length=77, padding="max_length", truncation=True).input_ids, dtype=torch.long, device=device)
|
46 |
+
text_feature = model.get_text_features(caption_input,walk_short_pos=True)
|
47 |
+
text_feature = text_feature / text_feature.norm(p=2, dim=-1, keepdim=True)
|
48 |
+
dense_image_feature = dense_image_feature / dense_image_feature.norm(p=2, dim=-1, keepdim=True)
|
49 |
+
|
50 |
+
similarity = dense_image_feature.squeeze() @ text_feature.squeeze().T
|
51 |
+
similarity = similarity.cpu().numpy()
|
52 |
+
patch_size = int(math.sqrt(similarity.shape[0]))
|
53 |
+
|
54 |
+
|
55 |
+
original_shape = (patch_size, patch_size)
|
56 |
+
show_image = similarity.reshape(original_shape)
|
57 |
+
|
58 |
+
|
59 |
+
fig = plt.figure(figsize=(6, 6))
|
60 |
+
plt.imshow(show_image)
|
61 |
+
plt.title('similarity Visualization')
|
62 |
+
plt.axis('off')
|
63 |
+
|
64 |
+
buf = io.BytesIO()
|
65 |
+
plt.savefig(buf, format='png')
|
66 |
+
buf.seek(0)
|
67 |
+
plt.close(fig)
|
68 |
+
|
69 |
+
pil_img = Image.open(buf)
|
70 |
+
# buf.close()
|
71 |
+
return pil_img
|
72 |
+
|
73 |
+
|
74 |
+
|
75 |
+
|
76 |
+
|
77 |
+
with gr.Blocks() as demo:
|
78 |
+
gr.Markdown("# FG-CLIP Densefeature")
|
79 |
+
gr.Markdown(
|
80 |
+
|
81 |
+
"This app uses the FG-CLIP model (qihoo360/fg-clip-base) for Densefeature show on CPU :"
|
82 |
+
)
|
83 |
+
gr.Markdown(
|
84 |
+
"<span style='color: red; font-weight: bold;'>⚠️ (Run DenseFeature) only support one class</span>"
|
85 |
+
)
|
86 |
+
|
87 |
+
with gr.Row():
|
88 |
+
with gr.Column():
|
89 |
+
image_input = gr.Image(type="pil")
|
90 |
+
text_input = gr.Textbox(label="Input a label")
|
91 |
+
dfs_button = gr.Button("Run Densefeature", visible=True)
|
92 |
+
with gr.Column():
|
93 |
+
dfs_output = gr.Image(label="Similarity Visualization", type="pil")
|
94 |
+
|
95 |
+
examples = [
|
96 |
+
["./cat_dfclor.jpg", "white cat,"],
|
97 |
+
]
|
98 |
+
gr.Examples(
|
99 |
+
examples=examples,
|
100 |
+
inputs=[image_input, text_input],
|
101 |
+
|
102 |
+
)
|
103 |
+
dfs_button.click(fn=Get_Densefeature, inputs=[image_input, text_input], outputs=dfs_output)
|
104 |
+
demo.launch()
|