Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gradio as gr
|
| 2 |
+
import spaces
|
| 3 |
+
from PIL import Image
|
| 4 |
+
import requests
|
| 5 |
+
from transformers import AutoModelForCausalLM, AutoProcessor
|
| 6 |
+
import torch
|
| 7 |
+
|
| 8 |
+
# Load the model and processor
|
| 9 |
+
model_id = "microsoft/Phi-3.5-vision-instruct"
|
| 10 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 11 |
+
model_id,
|
| 12 |
+
trust_remote_code=True,
|
| 13 |
+
torch_dtype=torch.float16,
|
| 14 |
+
)
|
| 15 |
+
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True, num_crops=16)
|
| 16 |
+
|
| 17 |
+
@spaces.GPU(duration=120) # Adjust the duration as needed
|
| 18 |
+
def solve_math_problem(image):
|
| 19 |
+
# Move model to GPU for this function call
|
| 20 |
+
model.to('cuda')
|
| 21 |
+
|
| 22 |
+
# Prepare the input
|
| 23 |
+
messages = [
|
| 24 |
+
{"role": "user", "content": "<|image_1|>\nSolve this math problem step by step. Explain your reasoning clearly."},
|
| 25 |
+
]
|
| 26 |
+
prompt = processor.tokenizer.apply_chat_template(
|
| 27 |
+
messages,
|
| 28 |
+
tokenize=False,
|
| 29 |
+
add_generation_prompt=True
|
| 30 |
+
)
|
| 31 |
+
|
| 32 |
+
# Process the input
|
| 33 |
+
inputs = processor(prompt, image, return_tensors="pt").to("cuda")
|
| 34 |
+
|
| 35 |
+
# Generate the response
|
| 36 |
+
generation_args = {
|
| 37 |
+
"max_new_tokens": 1000,
|
| 38 |
+
"temperature": 0.2,
|
| 39 |
+
"do_sample": True,
|
| 40 |
+
}
|
| 41 |
+
generate_ids = model.generate(**inputs,
|
| 42 |
+
eos_token_id=processor.tokenizer.eos_token_id,
|
| 43 |
+
**generation_args
|
| 44 |
+
)
|
| 45 |
+
|
| 46 |
+
# Decode the response
|
| 47 |
+
generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
|
| 48 |
+
response = processor.batch_decode(generate_ids,
|
| 49 |
+
skip_special_tokens=True,
|
| 50 |
+
clean_up_tokenization_spaces=False
|
| 51 |
+
)[0]
|
| 52 |
+
|
| 53 |
+
# Move model back to CPU to free up GPU memory
|
| 54 |
+
model.to('cpu')
|
| 55 |
+
|
| 56 |
+
return response
|
| 57 |
+
|
| 58 |
+
# Create the Gradio interface
|
| 59 |
+
iface = gr.Interface(
|
| 60 |
+
fn=solve_math_problem,
|
| 61 |
+
inputs=gr.Image(type="pil"),
|
| 62 |
+
outputs="text",
|
| 63 |
+
title="Visual Math Problem Solver",
|
| 64 |
+
description="Upload an image of a math problem, and I'll try to solve it step by step!",
|
| 65 |
+
examples=[
|
| 66 |
+
["example_math_problem1.jpg"],
|
| 67 |
+
["example_math_problem2.jpg"]
|
| 68 |
+
]
|
| 69 |
+
)
|
| 70 |
+
|
| 71 |
+
# Launch the app
|
| 72 |
+
iface.launch()
|