export_onnx.py

"""Export the trained EMA U-Net to ONNX for browser inference.

Run:  python diffusion/export_onnx.py
Writes diffusion/web_static/unet.onnx and verifies it against PyTorch.
"""
import os

import numpy as np
import onnxruntime as ort
import torch

from train_diffusion import UNet, Diffusion, T

HERE = os.path.dirname(os.path.abspath(__file__))
CKPT = os.path.join(HERE, "out_big", "checkpoint.pt")
OUT = os.path.join(HERE, "web_static", "unet.onnx")

os.makedirs(os.path.dirname(OUT), exist_ok=True)

model = UNet(base=128).eval()
model.load_state_dict(torch.load(CKPT, map_location="cpu", weights_only=True)["ema"])

x = torch.randn(1, 3, 32, 32)
t = torch.tensor([500], dtype=torch.int64)

torch.onnx.export(
    model, (x, t), OUT,
    input_names=["x", "t"], output_names=["eps"],
    opset_version=17, dynamo=False,
)
print(f"exported: {os.path.getsize(OUT)/1e6:.1f} MB")

sess = ort.InferenceSession(OUT, providers=["CPUExecutionProvider"])
with torch.no_grad():
    ref = model(x, t).numpy()
out = sess.run(None, {"x": x.numpy(), "t": t.numpy()})[0]
err = np.abs(ref - out).max()
print(f"max abs diff vs pytorch: {err:.2e}")
assert err < 1e-3, "ONNX output mismatch"
print("verified OK")