Spaces:
Runtime error
Runtime error
RelaxxOfficial
commited on
Commit
•
07dadcd
1
Parent(s):
4e24657
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,149 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
|
3 |
+
from diffusers import StableDiffusionPipeline, EulerDiscreteScheduler
|
4 |
+
import torch
|
5 |
+
import cv2
|
6 |
+
import os
|
7 |
+
import base64
|
8 |
+
import soundfile as sf
|
9 |
+
import time
|
10 |
+
|
11 |
+
# --- Set up Models ---
|
12 |
+
|
13 |
+
# Stable Diffusion for image generation
|
14 |
+
scheduler = EulerDiscreteScheduler.from_pretrained("stabilityai/stable-diffusion-2-1", subfolder="scheduler")
|
15 |
+
pipe = StableDiffusionPipeline.from_pretrained(
|
16 |
+
"stabilityai/stable-diffusion-2-1",
|
17 |
+
scheduler=scheduler,
|
18 |
+
torch_dtype=torch.float16
|
19 |
+
).to("cuda")
|
20 |
+
|
21 |
+
# LLaVA for vision-based language understanding
|
22 |
+
tokenizer = AutoTokenizer.from_pretrained("xtuner/llava-llama-3-8b-v1_1-transformers")
|
23 |
+
model = AutoModelForCausalLM.from_pretrained("xtuner/llava-llama-3-8b-v1_1-transformers").to("cuda")
|
24 |
+
|
25 |
+
# Open-source language model for text generation (e.g., GPT-Neo)
|
26 |
+
gpt_neo_pipe = pipeline("text-generation", model="EleutherAI/gpt-neo-1.3B")
|
27 |
+
|
28 |
+
# Text-to-Speech
|
29 |
+
text_to_speech = pipeline(
|
30 |
+
"text-to-speech", model="espnet/fastspeech2_en_ljspeech"
|
31 |
+
)
|
32 |
+
|
33 |
+
# --- Functions ---
|
34 |
+
|
35 |
+
def process_image(image_base64, chat_history):
|
36 |
+
"""Processes an image, sends it to LLaVA, and generates a response."""
|
37 |
+
# Prepare LLaVA input
|
38 |
+
input_text = f"""<image> {image_base64} </image>\n\nWhat do you see in this image?"""
|
39 |
+
inputs = tokenizer(input_text, return_tensors="pt").to("cuda")
|
40 |
+
|
41 |
+
# Generate response using LLaVA
|
42 |
+
with torch.no_grad():
|
43 |
+
outputs = model(**inputs)
|
44 |
+
response = tokenizer.decode(outputs.logits.argmax(-1)[0], skip_special_tokens=True)
|
45 |
+
|
46 |
+
# Generate speech from the response
|
47 |
+
audio = text_to_speech(response)
|
48 |
+
audio_path = "generated_audio.wav"
|
49 |
+
sf.write(audio_path, audio[0].numpy(), samplerate=22050)
|
50 |
+
|
51 |
+
# Update chat history
|
52 |
+
chat_history += "You: Image\n"
|
53 |
+
chat_history += "Model: " + response + "\n"
|
54 |
+
|
55 |
+
return chat_history, audio_path
|
56 |
+
|
57 |
+
def generate_image(prompt, chat_history):
|
58 |
+
"""Generates an image using Stable Diffusion based on a prompt."""
|
59 |
+
image = pipe(
|
60 |
+
prompt=prompt,
|
61 |
+
guidance_scale=7.5,
|
62 |
+
num_inference_steps=50,
|
63 |
+
).images[0]
|
64 |
+
|
65 |
+
# Update chat history
|
66 |
+
chat_history += "You: " + prompt + "\n"
|
67 |
+
chat_history += "Model: Image\n"
|
68 |
+
|
69 |
+
return chat_history, image
|
70 |
+
|
71 |
+
def process_text(text, chat_history):
|
72 |
+
"""Processes text, generates a response using GPT-Neo, and generates speech."""
|
73 |
+
# Generate response using GPT-Neo
|
74 |
+
response = gpt_neo_pipe(
|
75 |
+
text,
|
76 |
+
max_length=100,
|
77 |
+
num_return_sequences=1,
|
78 |
+
)[0]["generated_text"]
|
79 |
+
|
80 |
+
# Generate speech from the response
|
81 |
+
audio = text_to_speech(response)
|
82 |
+
audio_path = "generated_audio.wav"
|
83 |
+
sf.write(audio_path, audio[0].numpy(), samplerate=22050)
|
84 |
+
|
85 |
+
# Update chat history
|
86 |
+
chat_history += "You: " + text + "\n"
|
87 |
+
chat_history += "Model: " + response + "\n"
|
88 |
+
|
89 |
+
return chat_history, audio_path
|
90 |
+
|
91 |
+
# --- Webcam Capture ---
|
92 |
+
|
93 |
+
def capture_image():
|
94 |
+
"""Captures a screenshot from the webcam."""
|
95 |
+
cap = cv2.VideoCapture(0)
|
96 |
+
ret, frame = cap.read()
|
97 |
+
cap.release()
|
98 |
+
image = Image.fromarray(frame)
|
99 |
+
image_bytes = image.convert("RGB").save("captured_image.jpg", "JPEG")
|
100 |
+
with open("captured_image.jpg", "rb") as f:
|
101 |
+
image_base64 = base64.b64encode(f.read()).decode("utf-8")
|
102 |
+
return image_base64
|
103 |
+
|
104 |
+
# --- Gradio Interface ---
|
105 |
+
|
106 |
+
with gr.Blocks() as demo:
|
107 |
+
gr.Markdown("## Llama-LLaVA Vision Speech Assistant")
|
108 |
+
chat_history = gr.Textbox(label="Chat History", lines=10, interactive=False)
|
109 |
+
webcam_output = gr.Image(label="Webcam Feed", interactive=False)
|
110 |
+
image_input = gr.Image(label="Uploaded Image")
|
111 |
+
text_input = gr.Textbox(label="Enter Text")
|
112 |
+
audio_output = gr.Audio(label="Audio Response")
|
113 |
+
|
114 |
+
# Screenshot button
|
115 |
+
screenshot_button = gr.Button("Capture Screenshot")
|
116 |
+
screenshot_button.click(fn=capture_image, outputs=image_input)
|
117 |
+
|
118 |
+
# Image processing (LLaVA)
|
119 |
+
image_input.change(fn=process_image, inputs=[image_input, chat_history], outputs=[chat_history, audio_output])
|
120 |
+
|
121 |
+
# Text processing (GPT-Neo)
|
122 |
+
text_input.submit(fn=process_text, inputs=[text_input, chat_history], outputs=[chat_history, audio_output])
|
123 |
+
|
124 |
+
# Image generation (Stable Diffusion)
|
125 |
+
with gr.Tab("Image Generation"):
|
126 |
+
image_prompt = gr.Textbox(label="Enter image prompt:")
|
127 |
+
image_generation_output = gr.Image(label="Generated Image")
|
128 |
+
generate_image_button = gr.Button("Generate Image")
|
129 |
+
generate_image_button.click(
|
130 |
+
fn=generate_image, inputs=[image_prompt, chat_history], outputs=[chat_history, image_generation_output]
|
131 |
+
)
|
132 |
+
|
133 |
+
# Webcam stream
|
134 |
+
with gr.Tab("Webcam"):
|
135 |
+
webcam_output = gr.Image(label="Webcam Feed", source="webcam", interactive=False)
|
136 |
+
# Update webcam image every second
|
137 |
+
def update_webcam():
|
138 |
+
cap = cv2.VideoCapture(0)
|
139 |
+
while True:
|
140 |
+
ret, frame = cap.read()
|
141 |
+
if not ret:
|
142 |
+
break
|
143 |
+
image = Image.fromarray(frame)
|
144 |
+
yield image
|
145 |
+
time.sleep(1) # Update every second
|
146 |
+
|
147 |
+
webcam_output.source = update_webcam()
|
148 |
+
|
149 |
+
demo.launch(share=True)
|