File size: 1,653 Bytes
9497244
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
from fastapi import FastAPI, UploadFile, File
from fastapi.middleware.cors import CORSMiddleware
from transformers import TrOCRProcessor, VisionEncoderDecoderModel
from PIL import Image
import torch
import io

app = FastAPI()

# Enable CORS so the React frontend can communicate with this API
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],  # In production, change this to your frontend URL
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

device = torch.device("cpu") # Use CPU for standard web hosting unless paying for GPU servers
print("Loading model into server memory...")

# Point this to your fine-tuned local folder, or the base model if testing

model_path = "Vdv26/trocr-captcha-finetuned"
processor = TrOCRProcessor.from_pretrained(model_path)
model = VisionEncoderDecoderModel.from_pretrained(model_path).to(device)

@app.post("/api/predict")
async def predict_captcha(file: UploadFile = File(...)):
    # 1. Read the uploaded image bytes from the internet
    contents = await file.read()
    image = Image.open(io.BytesIO(contents)).convert("RGB")
    
    # 2. Run inference
    pixel_values = processor(images=image, return_tensors="pt").pixel_values.to(device)
    
    with torch.no_grad():
        generated_ids = model.generate(pixel_values, max_new_tokens=10)
        
    prediction = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
    
    # 3. Return the JSON response to the frontend
    return {"filename": file.filename, "prediction": prediction.replace(' ', '')}

# Run locally using: uvicorn main:app --reload