import io import random import torch import torch.nn.functional as F from contextlib import asynccontextmanager from fastapi import FastAPI, HTTPException from fastapi.responses import Response from torchvision.utils import save_image from train import FlowMatchingUNet, sample_images, DEVICE model = FlowMatchingUNet() def load_model(): ckpt = torch.load("weights/checkpoint_latest.pth", map_location="cpu", weights_only=True) model.load_state_dict(ckpt["ema"]) model.to(DEVICE) model.eval() def generate_png(steps: int = 20) -> bytes: """Generate one image and return it as PNG bytes.""" with torch.no_grad(): samples = sample_images(model, num_samples=1, steps=steps) big = F.interpolate(samples, scale_factor=2, mode="bicubic", align_corners=False) big = big.clamp(0, 1) buffer = io.BytesIO() save_image(big.cpu(), buffer, format="png", nrow=1) return buffer.getvalue() @asynccontextmanager async def lifespan(app: FastAPI): load_model() yield app = FastAPI(lifespan=lifespan) @app.get("/", response_class=Response) async def get_image(): try: png = generate_png() except Exception as e: raise HTTPException(status_code=500, detail=f"Generation failed: {e}") return Response(content=png, media_type="image/png") if __name__ == "__main__": import uvicorn uvicorn.run(app, host="0.0.0.0", port=8000)