Spaces:
Sleeping
Sleeping
Initial commit :tada:
Browse files
app.py
ADDED
@@ -0,0 +1,120 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import requests
|
2 |
+
import numpy as np
|
3 |
+
import cv2 as cv2
|
4 |
+
from PIL import Image
|
5 |
+
|
6 |
+
import torch
|
7 |
+
import torch.nn.functional as F
|
8 |
+
import open_clip
|
9 |
+
|
10 |
+
import gradio as gr
|
11 |
+
|
12 |
+
from legrad import LeWrapper, LePreprocess
|
13 |
+
|
14 |
+
|
15 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
16 |
+
layer_index = -2 # will run on cpu
|
17 |
+
image_size = 448
|
18 |
+
# ---------- Init CLIP Model ----------
|
19 |
+
model_name = 'ViT-B-16'
|
20 |
+
pretrained = 'laion2b_s34b_b88k'
|
21 |
+
patch_size = 16
|
22 |
+
|
23 |
+
model, _, preprocess = open_clip.create_model_and_transforms(model_name, pretrained=pretrained, device=device)
|
24 |
+
tokenizer = open_clip.get_tokenizer(model_name)
|
25 |
+
|
26 |
+
# ---------- Apply LeGrad's wrappers ----------
|
27 |
+
model = LeWrapper(model)
|
28 |
+
preprocess = LePreprocess(preprocess=preprocess, image_size=image_size)
|
29 |
+
|
30 |
+
|
31 |
+
# ---------- Function to load image from URL ----------
|
32 |
+
def change_to_url(url):
|
33 |
+
img_pil = Image.open(requests.get(url, stream=True).raw).convert('RGB')
|
34 |
+
return img_pil
|
35 |
+
|
36 |
+
|
37 |
+
def _get_text_embedding(model, tokenizer, classes: list, device):
|
38 |
+
prompts = [f'a photo of a {cls}.' for cls in classes]
|
39 |
+
|
40 |
+
tokenized_prompts = tokenizer(prompts).to(device)
|
41 |
+
|
42 |
+
text_embedding = model.encode_text(tokenized_prompts)
|
43 |
+
text_embedding = F.normalize(text_embedding, dim=-1)
|
44 |
+
return text_embedding.unsqueeze(0)
|
45 |
+
|
46 |
+
# ---------- Function to convert logits to heatmaps ----------
|
47 |
+
def logits_to_heatmaps(logits, image_cv):
|
48 |
+
logits = logits[0, 0].detach().cpu().numpy()
|
49 |
+
logits = (logits * 255).astype('uint8')
|
50 |
+
heat_map = cv2.applyColorMap(logits, cv2.COLORMAP_JET)
|
51 |
+
viz = 0.4 * image_cv + 0.6 * heat_map
|
52 |
+
viz = cv2.cvtColor(viz.astype('uint8'), cv2.COLOR_BGR2RGB)
|
53 |
+
return viz
|
54 |
+
|
55 |
+
|
56 |
+
# ---------- Main visualization function ----------
|
57 |
+
def viz_func(url, image, text_query):
|
58 |
+
image_torch = preprocess(image).unsqueeze(0).to(device)
|
59 |
+
text_emb = _get_text_embedding(model, tokenizer, classes=[text_query], device=device)
|
60 |
+
|
61 |
+
# ------- Get LeGrad output -------
|
62 |
+
logits_legrad = model.compute_legrad(image=image_torch, text_embedding=text_emb)
|
63 |
+
# ------- Get Heatmpas -------
|
64 |
+
image_cv = cv2.cvtColor(np.array(image.resize((image_size, image_size))), cv2.COLOR_RGB2BGR)
|
65 |
+
|
66 |
+
viz_legrad = logits_to_heatmaps(logits=logits_legrad, image_cv=image_cv)
|
67 |
+
return viz_legrad
|
68 |
+
|
69 |
+
inputs = [
|
70 |
+
gr.Textbox(label="Paste the url to the selected image"),
|
71 |
+
gr.Image(type="pil", interactive=True, label='Select An Image'),
|
72 |
+
gr.Textbox(label="Text query"),
|
73 |
+
]
|
74 |
+
|
75 |
+
|
76 |
+
with gr.Blocks(css="#gradio-app-title { text-align: center; }") as demo:
|
77 |
+
gr.Markdown(
|
78 |
+
"""
|
79 |
+
# **LeGrad: An Explainability Method for Vision Transformers via Feature Formation Sensitivity**
|
80 |
+
### This demo that showcases LeGrad method to visualize the important regions in an image that correspond to a given text query.
|
81 |
+
The model used is OpenCLIP-ViT-B-16 (weights: `laion2b_s34b_b88k`)
|
82 |
+
"""
|
83 |
+
)
|
84 |
+
with gr.Row():
|
85 |
+
with gr.Column():
|
86 |
+
gr.Markdown('# Select An Image')
|
87 |
+
selected_image = gr.Image(type="pil", interactive=True, label='')
|
88 |
+
gr.Markdown('## Paste the url to the selected image')
|
89 |
+
url_query = gr.Textbox(label="")
|
90 |
+
gr.Markdown('# Create your Own query')
|
91 |
+
text_query = gr.Textbox(label='')
|
92 |
+
run_button = gr.Button(icon='https://cdn-icons-png.flaticon.com/512/3348/3348036.png')
|
93 |
+
|
94 |
+
inputs[0].change(fn=change_to_url, outputs=inputs[1], inputs=inputs[0])
|
95 |
+
gr.Markdown('## LeGrad Explanation')
|
96 |
+
le_grad_output = gr.Image(label='LeGrad')
|
97 |
+
|
98 |
+
run_button.click(fn=viz_func,
|
99 |
+
inputs=[url_query, selected_image, text_query],
|
100 |
+
outputs=[le_grad_output])
|
101 |
+
|
102 |
+
with gr.Column():
|
103 |
+
gr.Markdown('# Select a Premade Example')
|
104 |
+
gr.Examples(
|
105 |
+
examples=[
|
106 |
+
["gradio_app/assets/cats_remote_control.jpeg", "cat"],
|
107 |
+
["gradio_app/assets/cats_remote_control.jpeg", "remote control"],
|
108 |
+
["gradio_app/assets/la_baguette.webp", "la baguette"],
|
109 |
+
["gradio_app/assets/la_baguette.webp", "beret"],
|
110 |
+
["gradio_app/assets/pokemons.jpeg", "Pikachu"],
|
111 |
+
["gradio_app/assets/pokemons.jpeg", "Bulbasaur"],
|
112 |
+
["gradio_app/assets/pokemons.jpeg", "Charmander"],
|
113 |
+
["gradio_app/assets/pokemons.jpeg", "Pokemons"],
|
114 |
+
],
|
115 |
+
inputs=[selected_image, text_query],
|
116 |
+
label=''
|
117 |
+
)
|
118 |
+
|
119 |
+
demo.queue()
|
120 |
+
demo.launch()
|