Spaces:
Runtime error
Runtime error
Commit
Β·
0edd243
1
Parent(s):
7968d81
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,166 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import yt_dlp
|
3 |
+
import os
|
4 |
+
import time
|
5 |
+
import torch
|
6 |
+
from multilingual_clip import pt_multilingual_clip
|
7 |
+
import transformers
|
8 |
+
import clip
|
9 |
+
import numpy as np
|
10 |
+
import cv2
|
11 |
+
import random
|
12 |
+
from PIL import Image
|
13 |
+
|
14 |
+
os.system('%cd /Multilingual-CLIP && bash get-weights.sh')
|
15 |
+
|
16 |
+
class SearchVideo:
|
17 |
+
|
18 |
+
def __init__(
|
19 |
+
self,
|
20 |
+
clip_model: str,
|
21 |
+
text_model: str,
|
22 |
+
tokenizer,
|
23 |
+
compose,
|
24 |
+
) -> None:
|
25 |
+
"""
|
26 |
+
clip_model: CLIP model to use for image embeddings
|
27 |
+
text_model: text encoder model
|
28 |
+
"""
|
29 |
+
self.text_model = text_model
|
30 |
+
self.tokenizer = tokenizer
|
31 |
+
self.clip_model = clip_model
|
32 |
+
self.compose = compose
|
33 |
+
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
34 |
+
|
35 |
+
|
36 |
+
def __call__(self, video: str, text: str) -> list:
|
37 |
+
torch.cuda.empty_cache()
|
38 |
+
img_list = []
|
39 |
+
text_list = []
|
40 |
+
frames = self.video2frames_ffmpeg(video)
|
41 |
+
|
42 |
+
|
43 |
+
img_embs = self.get_img_embs(frames)
|
44 |
+
txt_emb = self.get_txt_embs(text)
|
45 |
+
# txt_emb = [[t]*len(frames) for t in txt_emb]
|
46 |
+
txt_emb = txt_emb*len(frames)
|
47 |
+
|
48 |
+
logits_per_image = self.compare_embeddings(img_embs, txt_emb)
|
49 |
+
logits_per_image = [logit.numpy()[0] for logit in logits_per_image]
|
50 |
+
ind = np.argmax(logits_per_image)
|
51 |
+
seg_path = self.extract_seg(video, ind)
|
52 |
+
return ind, seg_path, frames[ind]
|
53 |
+
|
54 |
+
|
55 |
+
def extract_seg(self, video:str, start:int):
|
56 |
+
start = start if start > 5 else start-5
|
57 |
+
start = time.strftime('%H:%M:%S', time.gmtime(start))
|
58 |
+
cmd = f'ffmpeg -ss {start} -i "{video}" -t 00:00:05 -vcodec copy -acodec copy -y segment_{start}.mp4'
|
59 |
+
os.system(cmd)
|
60 |
+
return f'segment_{start}.mp4'
|
61 |
+
|
62 |
+
def video2frames_ffmpeg(self, video: str) -> list:
|
63 |
+
frames_dir = 'frames'
|
64 |
+
if not os.path.exists(frames_dir):
|
65 |
+
os.makedirs(frames_dir)
|
66 |
+
|
67 |
+
select = "select='if(eq(n\,0),1,floor(t)-floor(prev_selected_t))'"
|
68 |
+
os.system(f'ffmpeg -i {video} -r 1 {frames_dir}/output-%04d.jpg')
|
69 |
+
|
70 |
+
images = [Image.open(f'{frames_dir}/{f}') for f in sorted(os.listdir(frames_dir))]
|
71 |
+
os.system(f'rm -rf {frames_dir}')
|
72 |
+
return images
|
73 |
+
|
74 |
+
def video2frames(self, video: str) -> list:
|
75 |
+
cap = cv2.VideoCapture(video)
|
76 |
+
num_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
77 |
+
images = []
|
78 |
+
frames_sec = [i for i in range(0, num_frames, 24*1)]
|
79 |
+
has_frames,image = cap.read()
|
80 |
+
frame_count = 0
|
81 |
+
while has_frames:
|
82 |
+
has_frames,image = cap.read()
|
83 |
+
frame_count += 1
|
84 |
+
if has_frames:
|
85 |
+
if frame_count in frames_sec:
|
86 |
+
image = Image.fromarray(image)
|
87 |
+
images.append(image)
|
88 |
+
return images
|
89 |
+
|
90 |
+
def get_img_embs(self, img_list: list) -> list:
|
91 |
+
"""
|
92 |
+
takes list of image and calculates clip embeddings with model specified by clip_model
|
93 |
+
"""
|
94 |
+
img_input = torch.stack([self.compose(img).to(self.device)
|
95 |
+
for img in img_list])
|
96 |
+
with torch.no_grad():
|
97 |
+
image_embs = self.clip_model.encode_image(img_input).float().cpu()
|
98 |
+
return image_embs
|
99 |
+
|
100 |
+
def get_txt_embs(self, text: str) -> torch.Tensor:
|
101 |
+
"calculates clip emebdding for the text "
|
102 |
+
with torch.no_grad():
|
103 |
+
return self.text_model(text, self.tokenizer)
|
104 |
+
|
105 |
+
def compare_embeddings(self, img_embs, txt_embs):
|
106 |
+
# normalized features
|
107 |
+
image_features = img_embs / img_embs.norm(dim=-1, keepdim=True)
|
108 |
+
text_features = txt_embs / txt_embs.norm(dim=-1, keepdim=True)
|
109 |
+
|
110 |
+
# cosine similarity as logits
|
111 |
+
logits_per_image = []
|
112 |
+
for image_feature in image_features:
|
113 |
+
logits_per_image.append(image_feature @ text_features.t())
|
114 |
+
|
115 |
+
return logits_per_image
|
116 |
+
|
117 |
+
def download_yt_video(url):
|
118 |
+
ydl_opts = {
|
119 |
+
'quiet': True,
|
120 |
+
"outtmpl": "%(id)s.%(ext)s",
|
121 |
+
'format': 'bv*[height<=360][ext=mp4]+ba/b[height<=360] / wv*+ba/w'
|
122 |
+
}
|
123 |
+
|
124 |
+
with yt_dlp.YoutubeDL(ydl_opts) as ydl:
|
125 |
+
ydl.download([url])
|
126 |
+
return url.split('/')[-1].replace('watch?v=', '')+'.mp4'
|
127 |
+
|
128 |
+
|
129 |
+
clip_model='ViT-B/32'
|
130 |
+
text_model='M-CLIP/XLM-Roberta-Large-Vit-B-32'
|
131 |
+
clip_model, compose = clip.load(clip_model)
|
132 |
+
tokenizer = transformers.AutoTokenizer.from_pretrained(text_model)
|
133 |
+
text_model = pt_multilingual_clip.MultilingualCLIP.from_pretrained(text_model)
|
134 |
+
|
135 |
+
def search_video(video_url, text, video=None):
|
136 |
+
search = SearchVideo(
|
137 |
+
clip_model=clip_model,
|
138 |
+
text_model=text_model,
|
139 |
+
tokenizer=tokenizer,
|
140 |
+
compose=compose
|
141 |
+
)
|
142 |
+
if video !=None:
|
143 |
+
video_url = None
|
144 |
+
if video_url:
|
145 |
+
video = download_yt_video(video_url)
|
146 |
+
ind, seg_path, img = search(video, text)
|
147 |
+
start = time.strftime('%H:%M:%S', time.gmtime(ind))
|
148 |
+
return f'"{text}" found at {start}', seg_path
|
149 |
+
|
150 |
+
title = 'πποΈπ Search inside a video'
|
151 |
+
description = '''Just enter a search query, a video URL or upload your video and get a 5-sec fragment from the video which is visually closest to you query.'''
|
152 |
+
|
153 |
+
examples = [["https://www.youtube.com/watch?v=M93w3TjzVUE", "A dog"]]
|
154 |
+
|
155 |
+
iface = gr.Interface(
|
156 |
+
search_video,
|
157 |
+
inputs=[gr.Textbox(value="https://www.youtube.com/watch?v=M93w3TjzVUE", label='Video URL'), gr.Textbox(value="a dog", label='Text query'), gr.Video()],
|
158 |
+
outputs=[gr.Textbox(label="Output"), gr.Video(label="Video segment")],
|
159 |
+
allow_flagging="never",
|
160 |
+
title=title,
|
161 |
+
description=description,
|
162 |
+
examples=examples
|
163 |
+
)
|
164 |
+
|
165 |
+
if __name__ == "__main__":
|
166 |
+
iface.launch(show_error=True)
|