Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -3,7 +3,7 @@ from transformers import (
|
|
3 |
AutoModelForSpeechSeq2Seq,
|
4 |
AutoProcessor,
|
5 |
AutoModelForCausalLM,
|
6 |
-
AutoTokenizer,
|
7 |
BitsAndBytesConfig,
|
8 |
)
|
9 |
import torch
|
@@ -17,16 +17,12 @@ def yt2mp3(url, outputMp3F):
|
|
17 |
|
18 |
|
19 |
def speech2text(mp3_file):
|
20 |
-
# Set the computation device to GPU (if available) or CPU
|
21 |
device = 'cuda:0'
|
22 |
|
23 |
-
# Choose data type based on CUDA availability (float16 for GPU, float32 for CPU)
|
24 |
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
|
25 |
|
26 |
-
# Model identifier for the speech-to-text model
|
27 |
model_id = "distil-whisper/distil-large-v2"
|
28 |
|
29 |
-
# Load the model with specified configurations for efficient processing
|
30 |
model = AutoModelForSpeechSeq2Seq.from_pretrained(
|
31 |
model_id,
|
32 |
torch_dtype=torch_dtype,
|
@@ -35,13 +31,10 @@ def speech2text(mp3_file):
|
|
35 |
use_flash_attention_2=True
|
36 |
)
|
37 |
|
38 |
-
# Move the model to the specified device (GPU/CPU)
|
39 |
model.to(device)
|
40 |
|
41 |
-
# Load the processor for the model (handling tokenization and feature extraction)
|
42 |
processor = AutoProcessor.from_pretrained(model_id)
|
43 |
|
44 |
-
# Set up a speech recognition pipeline with the model and processor
|
45 |
pipe = pipeline(
|
46 |
"automatic-speech-recognition",
|
47 |
model=model,
|
@@ -54,35 +47,23 @@ def speech2text(mp3_file):
|
|
54 |
device=device,
|
55 |
)
|
56 |
|
57 |
-
# Process the MP3 file through the pipeline to get the speech recognition result
|
58 |
result = pipe(mp3_file)
|
59 |
|
60 |
-
# Extract the text from the recognition result
|
61 |
text_from_video = result["text"]
|
62 |
|
63 |
-
# Return the extracted text
|
64 |
return text_from_video
|
65 |
|
66 |
|
67 |
def chat(system_prompt, text):
|
68 |
-
"""
|
69 |
-
It is not a good practice to load the model again and again,
|
70 |
-
but for the sake of simlicity for demo, let's keep as it is
|
71 |
-
"""
|
72 |
-
|
73 |
-
# Define the model name to be used for the chat function
|
74 |
model_name = "meta-llama/Llama-2-7b-chat-hf"
|
75 |
-
# Authentication token for Hugging Face API
|
76 |
token = os.environ['HUGGINGFACE_TOKEN']
|
77 |
|
78 |
-
# Configure the model to load in a quantized 8-bit format for efficiency
|
79 |
bnb_config = BitsAndBytesConfig(
|
80 |
load_in_8bit=True
|
81 |
)
|
82 |
|
83 |
-
# Set the device map to load the model on GPU 0
|
84 |
device_map = {"": 0}
|
85 |
-
|
86 |
model = AutoModelForCausalLM.from_pretrained(
|
87 |
model_name,
|
88 |
quantization_config=bnb_config,
|
@@ -90,13 +71,10 @@ def chat(system_prompt, text):
|
|
90 |
use_auth_token=token
|
91 |
)
|
92 |
|
93 |
-
# Load the tokenizer for the model
|
94 |
tokenizer = AutoTokenizer.from_pretrained(model_name, use_auth_token=token)
|
95 |
|
96 |
-
# Create a text-generation pipeline with the loaded model and tokenizer
|
97 |
llama_pipeline = pipeline(task="text-generation", model=model, tokenizer=tokenizer)
|
98 |
|
99 |
-
# Format the input text with special tokens for the model
|
100 |
text = f"""
|
101 |
<s>[INST] <<SYS>>
|
102 |
{system_prompt}
|
@@ -104,7 +82,6 @@ def chat(system_prompt, text):
|
|
104 |
{text}[/INST]
|
105 |
"""
|
106 |
|
107 |
-
# Generate sequences using the pipeline with specified parameters
|
108 |
sequences = llama_pipeline(
|
109 |
text,
|
110 |
do_sample=True,
|
@@ -114,54 +91,38 @@ def chat(system_prompt, text):
|
|
114 |
max_length=32000
|
115 |
)
|
116 |
|
117 |
-
# Extract the generated text from the sequences
|
118 |
generated_text = sequences[0]["generated_text"]
|
119 |
-
# Trim the generated text to remove the instruction part
|
120 |
generated_text = generated_text[generated_text.find('[/INST]')+len('[/INST]'):]
|
121 |
|
122 |
-
# Return the processed generated text
|
123 |
return generated_text
|
124 |
|
125 |
def summarize(text):
|
126 |
-
# Define the maximum input length for each iteration of summarization
|
127 |
input_len = 10000
|
128 |
|
129 |
-
# Start an infinite loop to repeatedly summarize the text
|
130 |
while True:
|
131 |
-
# Print the current length of the text
|
132 |
-
print(len(text))
|
133 |
-
# Call the chat function to summarize the text. Only the first 'input_len' characters are considered for summarization
|
134 |
summary = chat("", "Summarize the following: " + text[0:input_len])
|
135 |
|
136 |
if len(text) < input_len:
|
137 |
return summary
|
138 |
|
139 |
-
# Concatenate the current summary with the remaining part of the text for the next iteration
|
140 |
text = summary + " " + text[input_len:]
|
141 |
|
142 |
import gradio as gr
|
143 |
|
144 |
-
# Fungsi dan impor yang sudah Anda miliki sebelumnya
|
145 |
-
|
146 |
-
# Fungsi untuk merangkum teks dari URL YouTube
|
147 |
def summarize_from_youtube(url):
|
148 |
-
# Unduh audio dari URL YouTube dan transkripsi ucapan menjadi teks
|
149 |
outputMp3F = "./files/audio.mp3"
|
150 |
yt2mp3(url=url, outputMp3F=outputMp3F)
|
151 |
transcribed = speech2text(mp3_file=outputMp3F)
|
152 |
-
# Rangkum teks yang telah ditranskripsi
|
153 |
summary = summarize(transcribed)
|
154 |
return summary
|
155 |
|
156 |
-
|
157 |
-
|
158 |
-
output_text = gr.outputs.Textbox(label="Ringkasan")
|
159 |
|
160 |
-
# Membuat antarmuka Gradio
|
161 |
gr.Interface(
|
162 |
fn=summarize_from_youtube,
|
163 |
inputs=youtube_url,
|
164 |
outputs=output_text,
|
165 |
-
title="
|
166 |
-
description="
|
167 |
).launch()
|
|
|
3 |
AutoModelForSpeechSeq2Seq,
|
4 |
AutoProcessor,
|
5 |
AutoModelForCausalLM,
|
6 |
+
AutoTokenizer,
|
7 |
BitsAndBytesConfig,
|
8 |
)
|
9 |
import torch
|
|
|
17 |
|
18 |
|
19 |
def speech2text(mp3_file):
|
|
|
20 |
device = 'cuda:0'
|
21 |
|
|
|
22 |
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
|
23 |
|
|
|
24 |
model_id = "distil-whisper/distil-large-v2"
|
25 |
|
|
|
26 |
model = AutoModelForSpeechSeq2Seq.from_pretrained(
|
27 |
model_id,
|
28 |
torch_dtype=torch_dtype,
|
|
|
31 |
use_flash_attention_2=True
|
32 |
)
|
33 |
|
|
|
34 |
model.to(device)
|
35 |
|
|
|
36 |
processor = AutoProcessor.from_pretrained(model_id)
|
37 |
|
|
|
38 |
pipe = pipeline(
|
39 |
"automatic-speech-recognition",
|
40 |
model=model,
|
|
|
47 |
device=device,
|
48 |
)
|
49 |
|
|
|
50 |
result = pipe(mp3_file)
|
51 |
|
|
|
52 |
text_from_video = result["text"]
|
53 |
|
|
|
54 |
return text_from_video
|
55 |
|
56 |
|
57 |
def chat(system_prompt, text):
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
model_name = "meta-llama/Llama-2-7b-chat-hf"
|
|
|
59 |
token = os.environ['HUGGINGFACE_TOKEN']
|
60 |
|
|
|
61 |
bnb_config = BitsAndBytesConfig(
|
62 |
load_in_8bit=True
|
63 |
)
|
64 |
|
|
|
65 |
device_map = {"": 0}
|
66 |
+
|
67 |
model = AutoModelForCausalLM.from_pretrained(
|
68 |
model_name,
|
69 |
quantization_config=bnb_config,
|
|
|
71 |
use_auth_token=token
|
72 |
)
|
73 |
|
|
|
74 |
tokenizer = AutoTokenizer.from_pretrained(model_name, use_auth_token=token)
|
75 |
|
|
|
76 |
llama_pipeline = pipeline(task="text-generation", model=model, tokenizer=tokenizer)
|
77 |
|
|
|
78 |
text = f"""
|
79 |
<s>[INST] <<SYS>>
|
80 |
{system_prompt}
|
|
|
82 |
{text}[/INST]
|
83 |
"""
|
84 |
|
|
|
85 |
sequences = llama_pipeline(
|
86 |
text,
|
87 |
do_sample=True,
|
|
|
91 |
max_length=32000
|
92 |
)
|
93 |
|
|
|
94 |
generated_text = sequences[0]["generated_text"]
|
|
|
95 |
generated_text = generated_text[generated_text.find('[/INST]')+len('[/INST]'):]
|
96 |
|
|
|
97 |
return generated_text
|
98 |
|
99 |
def summarize(text):
|
|
|
100 |
input_len = 10000
|
101 |
|
|
|
102 |
while True:
|
|
|
|
|
|
|
103 |
summary = chat("", "Summarize the following: " + text[0:input_len])
|
104 |
|
105 |
if len(text) < input_len:
|
106 |
return summary
|
107 |
|
|
|
108 |
text = summary + " " + text[input_len:]
|
109 |
|
110 |
import gradio as gr
|
111 |
|
|
|
|
|
|
|
112 |
def summarize_from_youtube(url):
|
|
|
113 |
outputMp3F = "./files/audio.mp3"
|
114 |
yt2mp3(url=url, outputMp3F=outputMp3F)
|
115 |
transcribed = speech2text(mp3_file=outputMp3F)
|
|
|
116 |
summary = summarize(transcribed)
|
117 |
return summary
|
118 |
|
119 |
+
youtube_url = gr.inputs.Textbox(lines=1, label="Enter YouTube URL")
|
120 |
+
output_text = gr.outputs.Textbox(label="Summary")
|
|
|
121 |
|
|
|
122 |
gr.Interface(
|
123 |
fn=summarize_from_youtube,
|
124 |
inputs=youtube_url,
|
125 |
outputs=output_text,
|
126 |
+
title="YouTube Summarizer",
|
127 |
+
description="Enter a YouTube URL to summarize its content."
|
128 |
).launch()
|