bimanual-pose-vae / example.py
WhynotGPT's picture
Release joint_mix40 VAE weights, training and GT neighborhood sampling code
34988af verified
Raw History Blame Contribute Delete
2.01 kB
"""CPU/GPU smoke example. Synthetic inputs only; not an accuracy evaluation."""
import argparse
import json
from pathlib import Path
import torch
from bimanual_vae import load_pair, pack
from hand_vae import aa_to_matrix
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--device", default="cpu")
args = parser.parse_args()
torch.manual_seed(0)
vae, _ = load_pair(str(Path(__file__).resolve().parent / "joint_mix40.pt"), args.device)
# Synthetic full MANO local axis angles, left hand first, right hand second.
local_aa = torch.zeros(2, 2, 15, 3, device=args.device, requires_grad=True)
root_aa = torch.zeros(2, 2, 3, device=args.device, requires_grad=True)
wrist = torch.tensor([[[0., 0., .5], [.2, 0., .5]],
[[0., 0., .6], [.2, .02, .6]]],
device=args.device, requires_grad=True)
local_rot, root_rot = aa_to_matrix(local_aa), aa_to_matrix(root_aa)
energy = vae.energy(local_rot, root_rot, wrist)
loss = 0.1 * energy.mean()
loss.backward()
for x in (local_aa, root_aa, wrist):
assert x.grad is not None and torch.isfinite(x.grad).all()
assert torch.isfinite(energy).all()
assert all(not p.requires_grad and p.grad is None for p in vae.parameters())
with torch.no_grad():
rot, pos = pack(local_rot, root_rot, wrist)
rot_rec, pos_rec, mu, logvar = vae(rot, pos)
assert rot_rec.shape == (2, 31, 3, 3)
assert pos_rec.shape == (2, 3)
assert mu.shape == logvar.shape == (2, 40)
print(json.dumps({"device": args.device, "torch": torch.__version__,
"synthetic_inputs": True, "energy": energy.detach().cpu().tolist(),
"rotation_shape": list(rot_rec.shape),
"position_shape": list(pos_rec.shape),
"latent_shape": list(mu.shape),
"finite_input_gradients": True, "frozen_vae": True}, indent=2))
if __name__ == "__main__":
main()