Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -1,7 +1,7 @@
|
|
1 |
##########################################
|
2 |
# Step 0: Import required libraries
|
3 |
##########################################
|
4 |
-
import streamlit as st
|
5 |
from transformers import (
|
6 |
pipeline,
|
7 |
SpeechT5Processor,
|
@@ -9,17 +9,17 @@ from transformers import (
|
|
9 |
SpeechT5HifiGan,
|
10 |
AutoModelForCausalLM,
|
11 |
AutoTokenizer
|
12 |
-
)
|
13 |
-
from datasets import load_dataset
|
14 |
-
import torch
|
15 |
-
import soundfile as sf
|
16 |
-
import sentencepiece
|
17 |
|
18 |
##########################################
|
19 |
# Initial configuration (MUST be first)
|
20 |
##########################################
|
21 |
st.set_page_config(
|
22 |
-
page_title="
|
23 |
page_icon="๐ฌ",
|
24 |
layout="centered",
|
25 |
initial_sidebar_state="collapsed"
|
@@ -32,22 +32,15 @@ st.set_page_config(
|
|
32 |
def load_models():
|
33 |
"""Load and cache all ML models"""
|
34 |
return {
|
35 |
-
# Emotion classifier
|
36 |
'emotion': pipeline(
|
37 |
"text-classification",
|
38 |
model="Thea231/jhartmann_emotion_finetuning"
|
39 |
),
|
40 |
-
|
41 |
-
# Text generation models
|
42 |
'textgen_tokenizer': AutoTokenizer.from_pretrained("Qwen/Qwen1.5-0.5B"),
|
43 |
'textgen_model': AutoModelForCausalLM.from_pretrained("Qwen/Qwen1.5-0.5B"),
|
44 |
-
|
45 |
-
# TTS components
|
46 |
'tts_processor': SpeechT5Processor.from_pretrained("microsoft/speecht5_tts"),
|
47 |
'tts_model': SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts"),
|
48 |
'tts_vocoder': SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan"),
|
49 |
-
|
50 |
-
# Speaker embeddings
|
51 |
'speaker_embeddings': torch.tensor(
|
52 |
load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")[7306]["xvector"]
|
53 |
).unsqueeze(0)
|
@@ -57,13 +50,13 @@ def load_models():
|
|
57 |
# UI Components
|
58 |
##########################################
|
59 |
def render_interface():
|
60 |
-
"""Create user interface
|
61 |
-
st.title("
|
62 |
-
st.
|
63 |
|
64 |
return st.text_area(
|
65 |
-
"๐
|
66 |
-
placeholder="
|
67 |
height=150,
|
68 |
key="user_input"
|
69 |
)
|
@@ -72,88 +65,80 @@ def render_interface():
|
|
72 |
# Core Logic Components
|
73 |
##########################################
|
74 |
def analyze_emotion(text, classifier):
|
75 |
-
"""Determine
|
76 |
results = classifier(text, return_all_scores=True)[0]
|
77 |
-
|
78 |
-
|
|
|
79 |
|
80 |
def generate_prompt(text, emotion):
|
81 |
-
"""
|
82 |
prompt_templates = {
|
83 |
-
"
|
84 |
-
"Customer
|
85 |
-
"Respond with:\n"
|
86 |
-
"
|
87 |
-
"Response:"
|
88 |
),
|
89 |
"joy": (
|
90 |
"Positive feedback: {input}\n"
|
91 |
-
"Respond with:\n"
|
92 |
-
"
|
93 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
94 |
),
|
95 |
-
"
|
96 |
-
"
|
97 |
-
"Respond with:\n"
|
98 |
-
"
|
99 |
-
"Response:"
|
100 |
)
|
101 |
}
|
102 |
-
return prompt_templates.get(emotion.lower(),
|
103 |
|
104 |
def process_response(output_text):
|
105 |
-
"""
|
106 |
-
|
107 |
-
|
108 |
-
output_text = output_text.rsplit('.', 1)[0] + '.'
|
109 |
-
|
110 |
-
# Length constraints
|
111 |
-
output_text = output_text[:300].strip() # Hard limit at 300 characters
|
112 |
-
|
113 |
-
# Fallback for short responses
|
114 |
-
if len(output_text) < 50:
|
115 |
-
return "Thank you for your feedback. We'll review this and contact you shortly."
|
116 |
-
|
117 |
-
return output_text
|
118 |
|
119 |
def generate_text_response(user_input, models):
|
120 |
-
"""
|
121 |
-
# Emotion analysis
|
122 |
emotion = analyze_emotion(user_input, models['emotion'])
|
123 |
-
|
124 |
-
# Prompt engineering
|
125 |
prompt = generate_prompt(user_input, emotion['label'])
|
126 |
|
127 |
-
# Text generation
|
128 |
inputs = models['textgen_tokenizer'](prompt, return_tensors="pt")
|
129 |
outputs = models['textgen_model'].generate(
|
130 |
inputs.input_ids,
|
131 |
-
max_new_tokens=
|
132 |
temperature=0.7,
|
133 |
do_sample=True,
|
134 |
top_p=0.9
|
135 |
)
|
136 |
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
|
141 |
def generate_audio_response(text, models):
|
142 |
-
"""
|
143 |
-
# Process text input
|
144 |
inputs = models['tts_processor'](text=text, return_tensors="pt")
|
145 |
-
|
146 |
-
# Generate spectrogram
|
147 |
spectrogram = models['tts_model'].generate_speech(
|
148 |
inputs["input_ids"],
|
149 |
models['speaker_embeddings']
|
150 |
)
|
151 |
-
|
152 |
-
# Generate waveform
|
153 |
with torch.no_grad():
|
154 |
waveform = models['tts_vocoder'](spectrogram)
|
155 |
-
|
156 |
-
# Save and return audio
|
157 |
sf.write("response.wav", waveform.numpy(), samplerate=16000)
|
158 |
return "response.wav"
|
159 |
|
@@ -161,25 +146,20 @@ def generate_audio_response(text, models):
|
|
161 |
# Main Application Flow
|
162 |
##########################################
|
163 |
def main():
|
164 |
-
# Load models once
|
165 |
ml_models = load_models()
|
166 |
-
|
167 |
-
# Render UI
|
168 |
user_input = render_interface()
|
169 |
|
170 |
-
# Process input
|
171 |
if user_input:
|
172 |
-
# Text
|
173 |
-
with st.
|
174 |
text_response = generate_text_response(user_input, ml_models)
|
175 |
-
status.update(label="โ
Analysis Complete", state="complete")
|
176 |
|
177 |
-
# Display
|
178 |
-
st.subheader("
|
179 |
-
st.
|
180 |
|
181 |
-
# Audio
|
182 |
-
with st.spinner("๐ Generating voice
|
183 |
audio_file = generate_audio_response(text_response, ml_models)
|
184 |
st.audio(audio_file, format="audio/wav")
|
185 |
|
|
|
1 |
##########################################
|
2 |
# Step 0: Import required libraries
|
3 |
##########################################
|
4 |
+
import streamlit as st
|
5 |
from transformers import (
|
6 |
pipeline,
|
7 |
SpeechT5Processor,
|
|
|
9 |
SpeechT5HifiGan,
|
10 |
AutoModelForCausalLM,
|
11 |
AutoTokenizer
|
12 |
+
)
|
13 |
+
from datasets import load_dataset
|
14 |
+
import torch
|
15 |
+
import soundfile as sf
|
16 |
+
import sentencepiece
|
17 |
|
18 |
##########################################
|
19 |
# Initial configuration (MUST be first)
|
20 |
##########################################
|
21 |
st.set_page_config(
|
22 |
+
page_title="Just Comment",
|
23 |
page_icon="๐ฌ",
|
24 |
layout="centered",
|
25 |
initial_sidebar_state="collapsed"
|
|
|
32 |
def load_models():
|
33 |
"""Load and cache all ML models"""
|
34 |
return {
|
|
|
35 |
'emotion': pipeline(
|
36 |
"text-classification",
|
37 |
model="Thea231/jhartmann_emotion_finetuning"
|
38 |
),
|
|
|
|
|
39 |
'textgen_tokenizer': AutoTokenizer.from_pretrained("Qwen/Qwen1.5-0.5B"),
|
40 |
'textgen_model': AutoModelForCausalLM.from_pretrained("Qwen/Qwen1.5-0.5B"),
|
|
|
|
|
41 |
'tts_processor': SpeechT5Processor.from_pretrained("microsoft/speecht5_tts"),
|
42 |
'tts_model': SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts"),
|
43 |
'tts_vocoder': SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan"),
|
|
|
|
|
44 |
'speaker_embeddings': torch.tensor(
|
45 |
load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")[7306]["xvector"]
|
46 |
).unsqueeze(0)
|
|
|
50 |
# UI Components
|
51 |
##########################################
|
52 |
def render_interface():
|
53 |
+
"""Create user interface"""
|
54 |
+
st.title("Just Comment")
|
55 |
+
st.markdown("### I'm listening to you, my friend๏ฝ")
|
56 |
|
57 |
return st.text_area(
|
58 |
+
"๐ Enter your comment:",
|
59 |
+
placeholder="Share your thoughts...",
|
60 |
height=150,
|
61 |
key="user_input"
|
62 |
)
|
|
|
65 |
# Core Logic Components
|
66 |
##########################################
|
67 |
def analyze_emotion(text, classifier):
|
68 |
+
"""Determine emotion with quick analysis"""
|
69 |
results = classifier(text, return_all_scores=True)[0]
|
70 |
+
valid_emotions = ['sadness', 'joy', 'love', 'anger', 'fear', 'surprise']
|
71 |
+
filtered = [e for e in results if e['label'].lower() in valid_emotions]
|
72 |
+
return max(filtered, key=lambda x: x['score'])
|
73 |
|
74 |
def generate_prompt(text, emotion):
|
75 |
+
"""Complete prompt templates for all 6 emotions"""
|
76 |
prompt_templates = {
|
77 |
+
"sadness": (
|
78 |
+
"Customer expressed sadness: {input}\n"
|
79 |
+
"Respond with:\n1. Empathetic acknowledgment\n"
|
80 |
+
"2. Supportive statement\n3. Concrete help offer\nResponse:"
|
|
|
81 |
),
|
82 |
"joy": (
|
83 |
"Positive feedback: {input}\n"
|
84 |
+
"Respond with:\n1. Enthusiastic thanks\n"
|
85 |
+
"2. Specific compliment\n3. Future engagement\nResponse:"
|
86 |
+
),
|
87 |
+
"love": (
|
88 |
+
"Customer showed affection: {input}\n"
|
89 |
+
"Respond with:\n1. Warm appreciation\n"
|
90 |
+
"2. Community building\n3. Exclusive offer\nResponse:"
|
91 |
+
),
|
92 |
+
"anger": (
|
93 |
+
"Angry complaint: {input}\n"
|
94 |
+
"Respond with:\n1. Sincere apology\n"
|
95 |
+
"2. Solution steps\n3. Compensation\nResponse:"
|
96 |
+
),
|
97 |
+
"fear": (
|
98 |
+
"Customer expressed concerns: {input}\n"
|
99 |
+
"Respond with:\n1. Reassurance\n"
|
100 |
+
"2. Safety measures\n3. Support channels\nResponse:"
|
101 |
),
|
102 |
+
"surprise": (
|
103 |
+
"Unexpected feedback: {input}\n"
|
104 |
+
"Respond with:\n1. Acknowledge uniqueness\n"
|
105 |
+
"2. Creative solution\n3. Follow-up plan\nResponse:"
|
|
|
106 |
)
|
107 |
}
|
108 |
+
return prompt_templates.get(emotion.lower(), "").format(input=text)
|
109 |
|
110 |
def process_response(output_text):
|
111 |
+
"""Optimized response processing"""
|
112 |
+
output_text = output_text.split("Response:")[-1].strip()
|
113 |
+
return output_text[:200] # Strict length control
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
114 |
|
115 |
def generate_text_response(user_input, models):
|
116 |
+
"""Efficient text generation"""
|
|
|
117 |
emotion = analyze_emotion(user_input, models['emotion'])
|
|
|
|
|
118 |
prompt = generate_prompt(user_input, emotion['label'])
|
119 |
|
|
|
120 |
inputs = models['textgen_tokenizer'](prompt, return_tensors="pt")
|
121 |
outputs = models['textgen_model'].generate(
|
122 |
inputs.input_ids,
|
123 |
+
max_new_tokens=150, # Reduced for speed
|
124 |
temperature=0.7,
|
125 |
do_sample=True,
|
126 |
top_p=0.9
|
127 |
)
|
128 |
|
129 |
+
return process_response(
|
130 |
+
models['textgen_tokenizer'].decode(outputs[0], skip_special_tokens=True)
|
131 |
+
)
|
132 |
|
133 |
def generate_audio_response(text, models):
|
134 |
+
"""Optimized TTS conversion"""
|
|
|
135 |
inputs = models['tts_processor'](text=text, return_tensors="pt")
|
|
|
|
|
136 |
spectrogram = models['tts_model'].generate_speech(
|
137 |
inputs["input_ids"],
|
138 |
models['speaker_embeddings']
|
139 |
)
|
|
|
|
|
140 |
with torch.no_grad():
|
141 |
waveform = models['tts_vocoder'](spectrogram)
|
|
|
|
|
142 |
sf.write("response.wav", waveform.numpy(), samplerate=16000)
|
143 |
return "response.wav"
|
144 |
|
|
|
146 |
# Main Application Flow
|
147 |
##########################################
|
148 |
def main():
|
|
|
149 |
ml_models = load_models()
|
|
|
|
|
150 |
user_input = render_interface()
|
151 |
|
|
|
152 |
if user_input:
|
153 |
+
# Text Generation
|
154 |
+
with st.spinner("๐ Analyzing emotions..."):
|
155 |
text_response = generate_text_response(user_input, ml_models)
|
|
|
156 |
|
157 |
+
# Display Results
|
158 |
+
st.subheader("๐ Generated Response")
|
159 |
+
st.success(text_response)
|
160 |
|
161 |
+
# Audio Generation
|
162 |
+
with st.spinner("๐ Generating voice..."):
|
163 |
audio_file = generate_audio_response(text_response, ml_models)
|
164 |
st.audio(audio_file, format="audio/wav")
|
165 |
|