server.py

import io
import os
import glob
import torch
import torch.nn as nn
from torchvision import models, transforms as T
from PIL import Image
from fastapi import FastAPI, UploadFile, File, HTTPException
import uvicorn

app = FastAPI()

# =================CONFIGURATION=================
CHECKPOINT_PATH = "checkpoints/best_age_model.pth"
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# ===============================================

# 1. Define the Model Architecture (Must match training EXACTLY)
def get_model():
    # Load standard ResNet18 structure
    model = models.resnet18(weights=None) 
    
    # Reconstruct the final layer exactly as we defined in training
    num_features = model.fc.in_features
    model.fc = nn.Sequential(
        nn.Linear(num_features, 128),
        nn.ReLU(),
        nn.Dropout(0.2),
        nn.Linear(128, 1)
    )
    return model

# 2. Load Weights
print(f"Initializing model on {DEVICE}...")
model = get_model()

if os.path.exists(CHECKPOINT_PATH):
    print(f"Loading weights from {CHECKPOINT_PATH}...")
    # Load the state dictionary
    state_dict = torch.load(CHECKPOINT_PATH, map_location=DEVICE)
    model.load_state_dict(state_dict)
    model.to(DEVICE)
    model.eval() # Important: Turn off Dropout for inference
    print("✅ Model loaded successfully!")
else:
    print(f"⚠️ Warning: Checkpoint not found at {CHECKPOINT_PATH}")
    print("Inference will use random weights (garbage output).")

# 3. Define Transforms (Must match training)
transform = T.Compose([
    T.Resize((200, 200)),
    T.ToTensor(),
    T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

@app.get("/")
def home():
    return {"status": "Age Predictor API is running"}

@app.post("/predict")
async def predict_age(file: UploadFile = File(...)):
    """
    Accepts an image file and returns predicted age.
    """
    try:
        # 1. Read and Process Image
        content = await file.read()
        image = Image.open(io.BytesIO(content)).convert("RGB")
        
        # 2. Transform
        input_tensor = transform(image).unsqueeze(0).to(DEVICE)
        
        # 3. Inference
        with torch.no_grad():
            prediction = model(input_tensor).squeeze()
            predicted_age = prediction.item()
            
        # 4. Return Result
        return {
            "filename": file.filename,
            "predicted_age": round(predicted_age, 1) # Round to 1 decimal
        }
        
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

if __name__ == "__main__":
    # Host 0.0.0.0 is needed for cloud environments/Docker
    uvicorn.run(app, host="0.0.0.0", port=8000)