app.py

import torch
from src.model import CRNN
from PIL import Image
import torchvision.transforms as transforms
import gradio as gr
import os

# ----------------------------
# 1️⃣ Load CRNN model
# ----------------------------
MODEL_PATH = "crnn_gpu.pt"
if not os.path.exists(MODEL_PATH):
    raise FileNotFoundError(f"{MODEL_PATH} not found! Make sure it's in the Space root.")

model = CRNN(img_height=32, img_width=100, img_channel=1, num_class=37, rnn_hidden=256)
model.load_state_dict(torch.load(MODEL_PATH, map_location="cpu"))
model.eval()

# ----------------------------
# 2️⃣ Characters and CTC decoding
# ----------------------------
alphabet = '0123456789abcdefghijklmnopqrstuvwxyz'

def ctc_decode(preds):
    """Greedy CTC decoder"""
    preds = preds.argmax(2).transpose(1,0).contiguous().view(-1)
    decoded = []
    prev_idx = -1
    for idx in preds:
        if idx != prev_idx and idx != 0:  # skip duplicates & blank
            decoded.append(alphabet[idx-1])
        prev_idx = idx
    return ''.join(decoded)

# ----------------------------
# 3️⃣ Preprocessing
# ----------------------------
def to_grayscale(img: Image.Image):
    """Convert any image type to grayscale"""
    if img.mode != "L":
        return img.convert("L")
    return img

transform = transforms.Compose([
    transforms.Lambda(to_grayscale),          # convert any input image to grayscale
    transforms.Resize((32, 100)),             # match CRNN input
    transforms.ToTensor(),
    transforms.Normalize((0.5,), (0.5,))
])

# ----------------------------
# 4️⃣ OCR function
# ----------------------------
def ocr(image: Image.Image):
    try:
        img_tensor = transform(image).unsqueeze(0)  # add batch dimension
        with torch.no_grad():
            preds = model(img_tensor)
            text = ctc_decode(preds)
        return text
    except Exception as e:
        return f"Error during inference: {e}"

# ----------------------------
# 5️⃣ Gradio interface
# ----------------------------
iface = gr.Interface(
    fn=ocr,
    inputs=gr.Image(type="pil", label="Upload any image (RGB, RGBA, etc.)"),
    outputs="text",
    title="CRNN OCR",
    description="Upload an image and get the OCR text prediction."
)

# Launch
iface.launch(share=True)