Spaces:
Sleeping
Sleeping
add first files
Browse files- app.py +222 -5
- fonts/Muller-Trial-Medium.ttf +0 -0
- images/icon.png +0 -0
- movies/.__init__.py +0 -0
- requirements.txt +95 -0
app.py
CHANGED
@@ -1,7 +1,224 @@
|
|
1 |
-
|
2 |
|
3 |
-
|
4 |
-
|
|
|
|
|
|
|
5 |
|
6 |
-
|
7 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import copy
|
2 |
|
3 |
+
from pytube import YouTube
|
4 |
+
from scipy.signal import resample
|
5 |
+
import gradio as gr
|
6 |
+
import numpy as np
|
7 |
+
import pytsmod as tsm
|
8 |
|
9 |
+
from moviepy.audio.AudioClip import AudioArrayClip
|
10 |
+
from moviepy.editor import *
|
11 |
+
from moviepy.video.fx.speedx import speedx
|
12 |
+
|
13 |
+
from sentence_transformers import SentenceTransformer, util
|
14 |
+
from transformers import pipeline, BertTokenizer, BertForNextSentencePrediction
|
15 |
+
import torch
|
16 |
+
import whisper
|
17 |
+
|
18 |
+
|
19 |
+
transcriber = whisper.load_model("medium")
|
20 |
+
sentence_transformer = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
|
21 |
+
tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
|
22 |
+
next_sentence_predict = BertForNextSentencePrediction.from_pretrained("bert-base-cased").eval()
|
23 |
+
summarizer = pipeline("summarization", model="philschmid/bart-large-cnn-samsum")
|
24 |
+
|
25 |
+
def get_youtube(video_url):
|
26 |
+
# YouTubeの動画をダウンロード
|
27 |
+
print("Start download video")
|
28 |
+
yt = YouTube(video_url)
|
29 |
+
abs_video_path = yt.streams.filter(progressive=True, file_extension='mp4').order_by('resolution').desc().first().download(filename='download.mp4', output_path='./movies/')
|
30 |
+
print("Success download video")
|
31 |
+
print(abs_video_path)
|
32 |
+
|
33 |
+
return abs_video_path
|
34 |
+
|
35 |
+
def two_chnnel_to_one_channel(sample):
|
36 |
+
# 音声を2チャンネルから1チャンネルに変換
|
37 |
+
left_channel = sample[:, 0]
|
38 |
+
right_channel = sample[:, 1]
|
39 |
+
mono_sample = (left_channel + right_channel) / 2
|
40 |
+
return mono_sample
|
41 |
+
|
42 |
+
def convert_sample_rate(data, original_sr, target_sr):
|
43 |
+
# 音声データのサンプリング周波数を変更
|
44 |
+
target_length = int(len(data) * target_sr / original_sr)
|
45 |
+
return resample(data, target_length)
|
46 |
+
|
47 |
+
def summarize_video(video_path, ratio_sum, playback_speed):
|
48 |
+
print("Start summarize video")
|
49 |
+
output_path = "./movies/output.mp4"
|
50 |
+
|
51 |
+
movie_clip = VideoFileClip(video_path)
|
52 |
+
|
53 |
+
audio_sampling_rate = movie_clip.audio.fps
|
54 |
+
clip_audio = np.array(movie_clip.audio.to_soundarray())
|
55 |
+
|
56 |
+
# 文字の書き起こし
|
57 |
+
audio_fp32 = convert_sample_rate(clip_audio, audio_sampling_rate, 16000)
|
58 |
+
audio_fp32 = two_chnnel_to_one_channel(audio_fp32).astype(np.float32)
|
59 |
+
transcription_results = transcriber.transcribe(audio_fp32)
|
60 |
+
|
61 |
+
# 文の句切れごとにテキスト/発話時間をまとめる
|
62 |
+
periods = ('.', '!', '?')
|
63 |
+
clip_sentences = []
|
64 |
+
head_sentence = True
|
65 |
+
for r in transcription_results['segments']:
|
66 |
+
if head_sentence:
|
67 |
+
start_time = r['start']
|
68 |
+
clip_sentences.append({'sentence':'', 'sentences':[], 'duration':[r['start'], None], 'durations':[]})
|
69 |
+
head_sentence = False
|
70 |
+
clip_sentences[-1]['sentence'] += r['text']
|
71 |
+
clip_sentences[-1]['sentences'].append(r['text'])
|
72 |
+
clip_sentences[-1]['durations'].append([r['start'], r['end']])
|
73 |
+
if r['text'].endswith(periods):
|
74 |
+
clip_sentences[-1]['duration'][1] = r['end']
|
75 |
+
head_sentence = True
|
76 |
+
|
77 |
+
# 文字の要約
|
78 |
+
transcription = transcription_results['text']
|
79 |
+
summary_text = summarizer(transcription, max_length=int(len(transcription)*0.1), min_length=int(len(transcription)*0.05), do_sample=False)[0]['summary_text']
|
80 |
+
print(summary_text)
|
81 |
+
|
82 |
+
# 要約文と一致する文を判別
|
83 |
+
summary_embedings = [sentence_transformer.encode(s, convert_to_tensor=True) for s in summary_text.split('.')]
|
84 |
+
important_sentence_idxs = [False]*len(clip_sentences)
|
85 |
+
for s, clip_sentence in enumerate(clip_sentences):
|
86 |
+
embedding = sentence_transformer.encode(clip_sentence['sentence'], convert_to_tensor=True)
|
87 |
+
for s_e in summary_embedings:
|
88 |
+
if util.pytorch_cos_sim(embedding, s_e) > ratio_sum:
|
89 |
+
important_sentence_idxs[s] = True
|
90 |
+
|
91 |
+
# となりの文と接続する文を判別
|
92 |
+
def next_prob(prompt, next_sentence, b=1.2):
|
93 |
+
encoding = tokenizer(prompt, next_sentence, return_tensors="pt")
|
94 |
+
logits = next_sentence_predict(**encoding, labels=torch.LongTensor([1])).logits
|
95 |
+
pos = b ** logits[0, 0]
|
96 |
+
neg = b ** logits[0, 1]
|
97 |
+
return float(pos / (pos + neg))
|
98 |
+
|
99 |
+
connection_idxs = [False]*(len(clip_sentences)-1)
|
100 |
+
for s in range(len(clip_sentences)-1):
|
101 |
+
if next_prob(clip_sentences[s]['sentence'], clip_sentences[s+1]['sentence']) > 0.88:
|
102 |
+
connection_idxs[s] = True
|
103 |
+
|
104 |
+
# 要約後の文章のみ残す
|
105 |
+
def combine_arrays(A, B):
|
106 |
+
C = copy.deepcopy(A)
|
107 |
+
for i in range(len(A)):
|
108 |
+
if A[i]:
|
109 |
+
j = i
|
110 |
+
while j < len(B) and B[j]:
|
111 |
+
C[j+1] = True
|
112 |
+
j += 1
|
113 |
+
j = i
|
114 |
+
while j > 0 and B[j-1]:
|
115 |
+
C[j] = True
|
116 |
+
j -= 1
|
117 |
+
return C
|
118 |
+
|
119 |
+
important_idxs = combine_arrays(important_sentence_idxs, connection_idxs)
|
120 |
+
|
121 |
+
# 要約後の文章がどこかを可視化
|
122 |
+
html_text = "<h1 class='title'>Full Transcription</h1>"
|
123 |
+
for idx in range(len(important_sentence_idxs)):
|
124 |
+
seconds = clip_sentences[idx]['duration'][0] * (1/playback_speed)
|
125 |
+
minutes = int(seconds // 60)
|
126 |
+
remaining_seconds = str(seconds % 60)
|
127 |
+
if important_idxs[idx]:
|
128 |
+
html_text += '<p> <font color="#dc974e">' + f"{minutes}:{remaining_seconds[0]} | {clip_sentences[idx]['sentence']}</font> </p>"
|
129 |
+
else:
|
130 |
+
html_text += f"<p>{minutes}:{remaining_seconds[0]} | {clip_sentences[idx]['sentence']}</p>"
|
131 |
+
|
132 |
+
# 動画を結合
|
133 |
+
clips = []
|
134 |
+
for i in range(len(important_idxs)):
|
135 |
+
if important_idxs[i]:
|
136 |
+
tmp_clips = []
|
137 |
+
for j in range(len(clip_sentences[i]['sentences'])):
|
138 |
+
start_time, end_time = clip_sentences[i]['durations'][j][0], clip_sentences[i]['durations'][j][1]
|
139 |
+
if end_time > movie_clip.duration:
|
140 |
+
end_time = movie_clip.duration
|
141 |
+
clip = movie_clip.subclip(start_time, end_time)
|
142 |
+
clip = clip.set_pos("center").set_duration(end_time-start_time)
|
143 |
+
txt_clip = TextClip(clip_sentences[i]['sentences'][j], fontsize=int(movie_clip.w/40), color='white', bg_color='black', font='./fonts/Muller-Trial-Medium.ttf')
|
144 |
+
txt_clip = txt_clip.set_duration(end_time-start_time).set_position(("center", "bottom"))
|
145 |
+
clip = CompositeVideoClip([clip, txt_clip])
|
146 |
+
tmp_clips.append(clip)
|
147 |
+
clips.append(concatenate_videoclips(tmp_clips))
|
148 |
+
|
149 |
+
# クリップをクロスディゾルブで結合
|
150 |
+
# for c in range(len(clips)-1):
|
151 |
+
# fade_duration = 2
|
152 |
+
# clips[c] = clips[c].crossfadeout(fade_duration).audio_fadeout(fade_duration)
|
153 |
+
# clips[c+1] = clips[c+1].crossfadein(fade_duration).audio_fadein(fade_duration)
|
154 |
+
|
155 |
+
# 動画を結合し再生速度を変化させる
|
156 |
+
final_video = concatenate_videoclips(clips, method="chain")
|
157 |
+
final_video_audio = np.array(final_video.audio.to_soundarray(fps=audio_sampling_rate))
|
158 |
+
if playback_speed != 1:
|
159 |
+
final_video_audio_fixed = tsm.wsola(final_video_audio, 1/playback_speed).T
|
160 |
+
else:
|
161 |
+
final_video_audio_fixed = final_video_audio
|
162 |
+
final_video = speedx(final_video, factor=playback_speed)
|
163 |
+
final_video = final_video.set_audio(AudioArrayClip(final_video_audio_fixed, fps=audio_sampling_rate))
|
164 |
+
# if final_video.duration > 30:
|
165 |
+
# final_video = final_video.subclip(0, 30)
|
166 |
+
final_video.write_videofile(output_path)
|
167 |
+
print(output_path)
|
168 |
+
print("Success summarize video")
|
169 |
+
return output_path, summary_text, html_text
|
170 |
+
|
171 |
+
|
172 |
+
# ---- Gradio Layout -----
|
173 |
+
youtube_url_in = gr.Textbox(label="Youtube url", lines=1, interactive=True)
|
174 |
+
video_in = gr.Video(label="Input Video", mirror_webcam=False, interactive=True)
|
175 |
+
video_out = gr.Video(label="Output Video")
|
176 |
+
summary_text = gr.Textbox(label="Video Transcription Summary")
|
177 |
+
transcription_text = gr.HTML(label="Full Transcription")
|
178 |
+
demo = gr.Blocks()
|
179 |
+
demo.encrypt = False
|
180 |
+
|
181 |
+
with demo:
|
182 |
+
gr.Markdown('''
|
183 |
+
<div style="text-align: center">
|
184 |
+
<h1 style='text-align: center'>FastPerson: Video summarization applied with transcription and text summarization</h1>
|
185 |
+
<img src="https://user-images.githubusercontent.com/33136532/215362410-97727904-e1ca-408d-967e-f5798671405e.png" alt="Video Summarization">
|
186 |
+
</div>
|
187 |
+
''')
|
188 |
+
with gr.Row():
|
189 |
+
gr.Markdown('''
|
190 |
+
### Summarize video
|
191 |
+
##### Step 1a. Download video from youtube
|
192 |
+
##### Step 1b. You also can upload video directly
|
193 |
+
##### Step 2. Enter summary rate and playback speed
|
194 |
+
##### Step 3. Generating summarized video.
|
195 |
+
''')
|
196 |
+
with gr.Row():
|
197 |
+
gr.Markdown('''
|
198 |
+
### You can test by following examples:
|
199 |
+
''')
|
200 |
+
examples = gr.Examples(examples=
|
201 |
+
[ "https://www.youtube.com/watch?v=QghjaS0WQQU",
|
202 |
+
"https://www.youtube.com/watch?v=cUS_22_lDiM",
|
203 |
+
"https://www.youtube.com/watch?v=80yqL2KzBVw"],
|
204 |
+
label="Examples", inputs=[youtube_url_in])
|
205 |
+
with gr.Column():
|
206 |
+
youtube_url_in.render()
|
207 |
+
download_youtube_btn = gr.Button("Download Youtube video")
|
208 |
+
download_youtube_btn.click(get_youtube, [youtube_url_in], [video_in])
|
209 |
+
print(video_in)
|
210 |
+
with gr.Row():
|
211 |
+
ratio_sum = gr.Slider(label="Summarize Ratio", minimum=0.3, maximum=0.8, step=0.05, value=0.6)
|
212 |
+
playback_speed = gr.Slider(label="Playback Speed", minimum=0.5, maximum=2.0, step=0.25, value=1.0)
|
213 |
+
with gr.Row():
|
214 |
+
upload_output_video_btn = gr.Button("Summarize Video")
|
215 |
+
upload_output_video_btn.click(summarize_video, [video_in, ratio_sum, playback_speed], [video_out, summary_text, transcription_text])
|
216 |
+
with gr.Row():
|
217 |
+
video_in.render()
|
218 |
+
video_out.render()
|
219 |
+
with gr.Row():
|
220 |
+
summary_text.render()
|
221 |
+
with gr.Row():
|
222 |
+
transcription_text.render()
|
223 |
+
|
224 |
+
demo.launch(debug=True, share=True)
|
fonts/Muller-Trial-Medium.ttf
ADDED
Binary file (870 kB). View file
|
|
images/icon.png
ADDED
movies/.__init__.py
ADDED
File without changes
|
requirements.txt
ADDED
@@ -0,0 +1,95 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
aiofiles==22.1.0
|
2 |
+
aiohttp==3.8.3
|
3 |
+
aiosignal==1.3.1
|
4 |
+
altair==4.2.2
|
5 |
+
anyio==3.6.2
|
6 |
+
async-timeout==4.0.2
|
7 |
+
attrs==22.2.0
|
8 |
+
certifi @ file:///croot/certifi_1671487769961/work/certifi
|
9 |
+
cffi==1.15.1
|
10 |
+
charset-normalizer==2.1.1
|
11 |
+
click==8.1.3
|
12 |
+
contourpy==1.0.7
|
13 |
+
cycler==0.11.0
|
14 |
+
decorator==4.4.2
|
15 |
+
entrypoints==0.4
|
16 |
+
fastapi==0.89.1
|
17 |
+
ffmpeg-python==0.2.0
|
18 |
+
ffmpy==0.3.0
|
19 |
+
filelock==3.9.0
|
20 |
+
fonttools==4.38.0
|
21 |
+
frozenlist==1.3.3
|
22 |
+
fsspec==2023.1.0
|
23 |
+
future==0.18.3
|
24 |
+
gradio==3.16.2
|
25 |
+
h11==0.14.0
|
26 |
+
httpcore==0.16.3
|
27 |
+
httpx==0.23.3
|
28 |
+
huggingface-hub==0.12.0
|
29 |
+
idna==3.4
|
30 |
+
imageio==2.25.0
|
31 |
+
imageio-ffmpeg==0.4.8
|
32 |
+
Jinja2==3.1.2
|
33 |
+
joblib==1.2.0
|
34 |
+
jsonschema==4.17.3
|
35 |
+
kiwisolver==1.4.4
|
36 |
+
linkify-it-py==1.0.3
|
37 |
+
markdown-it-py==2.1.0
|
38 |
+
MarkupSafe==2.1.2
|
39 |
+
matplotlib==3.6.3
|
40 |
+
mdit-py-plugins==0.3.3
|
41 |
+
mdurl==0.1.2
|
42 |
+
more-itertools==9.0.0
|
43 |
+
moviepy==1.0.3
|
44 |
+
multidict==6.0.4
|
45 |
+
nltk==3.8.1
|
46 |
+
numpy==1.24.1
|
47 |
+
nvidia-cublas-cu11==11.10.3.66
|
48 |
+
nvidia-cuda-nvrtc-cu11==11.7.99
|
49 |
+
nvidia-cuda-runtime-cu11==11.7.99
|
50 |
+
nvidia-cudnn-cu11==8.5.0.96
|
51 |
+
openai-whisper @ git+https://github.com/openai/whisper.git@5c1a8c10e762bf9c29fcf6b3e40f17bc8ab09864
|
52 |
+
orjson==3.8.5
|
53 |
+
packaging==23.0
|
54 |
+
pandas==1.5.3
|
55 |
+
Pillow==9.4.0
|
56 |
+
proglog==0.1.10
|
57 |
+
pycparser==2.21
|
58 |
+
pycryptodome==3.17
|
59 |
+
pydantic==1.10.4
|
60 |
+
pydub==0.25.1
|
61 |
+
pyparsing==3.0.9
|
62 |
+
pyrsistent==0.19.3
|
63 |
+
python-dateutil==2.8.2
|
64 |
+
python-multipart==0.0.5
|
65 |
+
pytsmod==0.3.6
|
66 |
+
pytube==12.1.0
|
67 |
+
pytz==2022.7.1
|
68 |
+
PyYAML==6.0
|
69 |
+
regex==2022.10.31
|
70 |
+
requests==2.28.2
|
71 |
+
rfc3986==1.5.0
|
72 |
+
scikit-learn==1.2.1
|
73 |
+
scipy==1.10.0
|
74 |
+
sentence-transformers==2.2.2
|
75 |
+
sentencepiece==0.1.97
|
76 |
+
six==1.16.0
|
77 |
+
sniffio==1.3.0
|
78 |
+
soundfile==0.11.0
|
79 |
+
starlette==0.22.0
|
80 |
+
threadpoolctl==3.1.0
|
81 |
+
tokenizers==0.13.2
|
82 |
+
toolz==0.12.0
|
83 |
+
torch==1.13.1
|
84 |
+
torchaudio==0.13.1
|
85 |
+
torchvision==0.14.1
|
86 |
+
tqdm==4.64.1
|
87 |
+
transformers==4.26.0
|
88 |
+
typing_extensions==4.4.0
|
89 |
+
uc-micro-py==1.0.1
|
90 |
+
urllib3==1.26.14
|
91 |
+
uvicorn==0.20.0
|
92 |
+
watchdog==2.2.1
|
93 |
+
websockets==10.4
|
94 |
+
whisper==1.1.10
|
95 |
+
yarl==1.8.2
|