Spaces:
Runtime error
Runtime error
Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,163 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
# Transcript
|
3 |
+
from youtube_transcript_api import YouTubeTranscriptApi
|
4 |
+
import os
|
5 |
+
# Summarization
|
6 |
+
from transformers import (
|
7 |
+
pipeline,
|
8 |
+
AutoModelForSpeechSeq2Seq,
|
9 |
+
AutoProcessor,
|
10 |
+
AutoModelForCausalLM,
|
11 |
+
AutoTokenizer,
|
12 |
+
BitsAndBytesConfig,
|
13 |
+
)
|
14 |
+
import torch
|
15 |
+
import re
|
16 |
+
|
17 |
+
|
18 |
+
def fetch_transcript(video_url):
|
19 |
+
try:
|
20 |
+
# Extract the video ID from the URL
|
21 |
+
video_id = video_url.split("v=")[1]
|
22 |
+
# Fetch the transcript for the video
|
23 |
+
transcript = YouTubeTranscriptApi.get_transcript(video_id)
|
24 |
+
# Process the transcript data
|
25 |
+
text_transcript = "\n".join([entry['text'] for entry in transcript])
|
26 |
+
return text_transcript
|
27 |
+
except Exception as e:
|
28 |
+
return str(e)
|
29 |
+
|
30 |
+
def clean_transcript(transcript):
|
31 |
+
# Remove non-speech elements (e.g., laughter, background noises)
|
32 |
+
transcript = re.sub(r'\[.*?\]', '', transcript)
|
33 |
+
|
34 |
+
# Correct spelling and grammar (you can use libraries like NLTK or spaCy for this)
|
35 |
+
# Example:
|
36 |
+
# import nltk
|
37 |
+
# transcript = ' '.join(nltk.word_tokenize(transcript))
|
38 |
+
|
39 |
+
# Normalize punctuation and formatting
|
40 |
+
transcript = transcript.replace('\n', ' ') # Remove line breaks
|
41 |
+
transcript = re.sub(r'\s+', ' ', transcript) # Remove extra whitespaces
|
42 |
+
|
43 |
+
# Remove timestamps and annotations
|
44 |
+
transcript = re.sub(r'\[\d+:\d+:\d+\]', '', transcript)
|
45 |
+
|
46 |
+
# Handle speaker identification (if present)
|
47 |
+
# Example: transcript = re.sub(r'Speaker\d+:', '', transcript)
|
48 |
+
|
49 |
+
# Remove filler words and phrases
|
50 |
+
filler_words = ['like', 'you know', 'sort of'] # Add more as needed
|
51 |
+
for word in filler_words:
|
52 |
+
transcript = transcript.replace(word, '')
|
53 |
+
|
54 |
+
# Replace common contractions with their expanded forms
|
55 |
+
transcript = transcript.replace("won't", "will not")
|
56 |
+
transcript = transcript.replace("can't", "cannot")
|
57 |
+
transcript = transcript.replace("n't", " not")
|
58 |
+
transcript = transcript.replace("'ll", " will")
|
59 |
+
transcript = transcript.replace("'ve", " have")
|
60 |
+
transcript = transcript.replace("'re", " are")
|
61 |
+
transcript = transcript.replace("'d", " would")
|
62 |
+
transcript = transcript.replace("'s", " is")
|
63 |
+
|
64 |
+
return transcript.strip() # Trim leading/trailing whitespaces
|
65 |
+
|
66 |
+
def extract_video_id(url):
|
67 |
+
"""Extracts the YouTube video ID from the URL."""
|
68 |
+
match = re.search(r"(?<=v=)[\w-]+", url)
|
69 |
+
if match:
|
70 |
+
return match.group(0)
|
71 |
+
else:
|
72 |
+
return None
|
73 |
+
|
74 |
+
|
75 |
+
def summarize_transcript(text, llama_pipeline):
|
76 |
+
def summarize_text(llama_pipeline, system_prompt, text):
|
77 |
+
# Format the input text with special tokens for the model
|
78 |
+
text = f"""
|
79 |
+
<s>[INST] <<SYS>>
|
80 |
+
{system_prompt}
|
81 |
+
<</SYS>>
|
82 |
+
{text}[/INST]
|
83 |
+
"""
|
84 |
+
# Generate sequences using the pipeline with specified parameters
|
85 |
+
sequences = llama_pipeline(text)
|
86 |
+
# Extract the generated text from the sequences
|
87 |
+
generated_text = sequences[0]["generated_text"]
|
88 |
+
# Trim the generated text to remove the instruction part
|
89 |
+
generated_text = generated_text[generated_text.find('[/INST]')+len('[/INST]'):]
|
90 |
+
# Return the processed generated text
|
91 |
+
return generated_text
|
92 |
+
# Define the maximum input length for each iteration of summarization
|
93 |
+
input_len = 1000
|
94 |
+
# Start an infinite loop to repeatedly summarize the text
|
95 |
+
while True:
|
96 |
+
# Print the current length of the text
|
97 |
+
print(len(text))
|
98 |
+
# Call the chat function to summarize the text. Only the first 'input_len' characters are considered for summarization
|
99 |
+
summary = summarize_text(llama_pipeline, "", "Summarize the following: " + text[0:input_len])
|
100 |
+
if len(summary) < input_len:
|
101 |
+
return summary
|
102 |
+
# Concatenate the current summary with the remaining part of the text for the next iteration
|
103 |
+
text = summary + " " + text[input_len:]
|
104 |
+
|
105 |
+
# Load the model and tokenizer
|
106 |
+
@st.cache_resource()
|
107 |
+
def load_model():
|
108 |
+
# Define the model name to be used for the chat function
|
109 |
+
model_name = "meta-llama/Llama-2-7b-chat-hf"
|
110 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
111 |
+
pipeline_llama2 = pipeline(
|
112 |
+
"text-generation", #task
|
113 |
+
model=model_name,
|
114 |
+
tokenizer=tokenizer,
|
115 |
+
torch_dtype=torch.bfloat16,
|
116 |
+
trust_remote_code=True,
|
117 |
+
device_map="auto",
|
118 |
+
# max_length=max_token_length,
|
119 |
+
do_sample=True,
|
120 |
+
top_k=10,
|
121 |
+
num_return_sequences=1,
|
122 |
+
eos_token_id=tokenizer.eos_token_id
|
123 |
+
)
|
124 |
+
return pipeline_llama2
|
125 |
+
|
126 |
+
def main():
|
127 |
+
st.title("YouTube Video Preview")
|
128 |
+
|
129 |
+
with st.spinner('Loading checkpoint shards of LLAMA-2'):
|
130 |
+
pipeline_llama2 = load_model()
|
131 |
+
st.success('Done!')
|
132 |
+
|
133 |
+
# Input field for the YouTube video link
|
134 |
+
youtube_url = st.text_input("Paste YouTube Video Link:")
|
135 |
+
|
136 |
+
# Extract video ID from the URL
|
137 |
+
video_id = extract_video_id(youtube_url)
|
138 |
+
|
139 |
+
# Display video preview if video ID is found
|
140 |
+
if video_id:
|
141 |
+
video_url = f"https://www.youtube.com/watch?v={video_id}"
|
142 |
+
st.video(video_url, format='video/mp4')
|
143 |
+
video_transcript = clean_transcript(fetch_transcript(video_url))
|
144 |
+
if video_transcript:
|
145 |
+
# Display transcript and summary side by side
|
146 |
+
col1, col2 = st.columns(2)
|
147 |
+
with col1:
|
148 |
+
st.subheader("Transcript:")
|
149 |
+
st.text_area(" ", video_transcript, height=400)
|
150 |
+
|
151 |
+
with col2:
|
152 |
+
st.subheader("Summary:")
|
153 |
+
video_summary = summarize_transcript(video_transcript, pipeline_llama2)
|
154 |
+
st.text_area(" ", video_summary, height=400)
|
155 |
+
print(f"Summary:{video_summary}")
|
156 |
+
else:
|
157 |
+
st.error("Failed to fetch video transcript. Please check the video ID or try again later.")
|
158 |
+
|
159 |
+
elif youtube_url:
|
160 |
+
st.warning("Invalid YouTube Video Link")
|
161 |
+
|
162 |
+
if __name__ == "__main__":
|
163 |
+
main()
|