scripts/export_anigen_dense.py

from __future__ import annotations
import json, os, sys, time
from pathlib import Path
import torch

OUT = Path(os.environ.get("CF_ONNX_OUT", "/tmp/cf-onnx-out"))
MODEL_ROOT = Path(os.environ.get("ANIGEN_MODEL_ROOT", "/tmp/anigen-model"))
APP_ROOT = Path(os.environ.get("ANIGEN_APP_ROOT", "/home/user/app"))
OUT.mkdir(parents=True, exist_ok=True)
sys.path.insert(0, str(APP_ROOT))
sys.path.insert(0, str(APP_ROOT / "third_parties" / "dsine"))


def ensure_weights():
    from huggingface_hub import snapshot_download
    snapshot_download(
        repo_id="VAST-AI/AniGen", token=os.environ.get("HF_TOKEN"), local_dir=MODEL_ROOT,
        allow_patterns=["ckpts/dinov2/**", "ckpts/dsine/**"],
    )
    os.chdir(MODEL_ROOT)


class DinoExport(torch.nn.Module):
    def __init__(self, model):
        super().__init__(); self.model=model
    def forward(self, pixel_values):
        return self.model(pixel_values, is_training=True)["x_prenorm"]


class DsineExport(torch.nn.Module):
    def __init__(self, model):
        super().__init__(); self.model=model
    def forward(self, image, intrins):
        return self.model(image, intrins=intrins)[-1]


def export_one(module, args, path: Path, input_names, output_names):
    path.parent.mkdir(parents=True, exist_ok=True)
    module.eval()
    t=time.time()
    with torch.inference_mode():
        torch.onnx.export(
            module, args, str(path), input_names=input_names, output_names=output_names,
            opset_version=18, do_constant_folding=True, dynamo=False,
            dynamic_axes=None, external_data=True,
        )
    import onnx
    model=onnx.load(str(path), load_external_data=False)
    onnx.checker.check_model(model)
    print(f"EXPORTED {path} {time.time()-t:.2f}s", flush=True)


def main():
    ensure_weights()
    print("loading dinov2", flush=True)
    dino = torch.hub.load('./ckpts/dinov2', 'dinov2_vitl14_reg', pretrained=True, source='local').eval().cpu()
    x=torch.zeros((1,3,518,518), dtype=torch.float32)
    export_one(DinoExport(dino), (x,), OUT/'onnx/dinov2/model.onnx', ['pixel_values'], ['x_prenorm'])
    del dino; import gc; gc.collect()

    print("loading dsine", flush=True)
    from anigen.utils.image_utils import load_dsine
    dsine=load_dsine('cuda').eval()
    image=torch.zeros((1,3,544,544),dtype=torch.float32,device='cuda')
    intrins=torch.tensor([[[471.117,0,259.0],[0,471.117,259.0],[0,0,1.0]]],dtype=torch.float32,device='cuda')
    export_one(DsineExport(dsine),(image,intrins),OUT/'onnx/dsine/model.onnx',['image','intrins'],['normal'])

    meta={
      'torch': torch.__version__, 'opset':18,
      'dinov2': {'input':[1,3,518,518], 'output':'x_prenorm'},
      'dsine': {'image_input':[1,3,544,544], 'intrinsics_input':[1,3,3], 'output':'normal'},
    }
    (OUT/'export_meta.json').write_text(json.dumps(meta,indent=2))

if __name__=='__main__': main()