there seems no great improvement with SMIL

[Charlulote]

testing with cpu, i got this.

    model = SMIL(sparse=sparse).to(device, dtype, non_blocking=True)
    for i in range(3):
        pose = torch.from_numpy((np.random.rand(batch_size, pose_size) - 0.5) * 0.4) \
            .type(dtype).to(device)
        betas = torch.from_numpy((np.random.rand(batch_size, beta_size) - 0.5) * 0.06) \
            .type(dtype).to(device)
        trans = torch.from_numpy(np.zeros((batch_size, 3))).type(dtype).to(device)
        s = time()
        result, joints = model(betas, pose, trans)
        print(time() - s)
-----------------------
0.009973764419555664
0.008975744247436523
0.007978677749633789

And do the same with smpl.

0.014949560165405273
0.008976459503173828
0.00797891616821289

Is there anything wrong or misunderstood?

[CalciferZh]

Transferring data to and from gpu may take some time. Having batched input may also make some difference.