The process of calculating the generation metrics is extremely slow

[Dai-Wenxun]

The bottleneck is in the SMPLX_Util.get_body_vertices_sequence, since it will load the smplx pretrained weights repeatedly. For example, there are 1319 examples of action walk, and the k == 10, then the number of times of the loading process will be 13190, making the IO time extremely long. My suggestion is: simply instance 1 smplx model with batch_size=max_motion_len, and select the unmasked smpl parameters after the inference of smplx model.
Here is my code to test the time cost of three modes:

import torch
import smplx
import mmengine as me

test_mode = 'cuda'  # cpu, cuda, cuda_static

device = 'cpu'
if test_mode in ['cuda', 'cuda_static']:
    device = 'cuda'

seq_len = 60
torch_param = dict()
torch_param['body_pose'] = torch.randn(seq_len, 63).to(device)
torch_param['betas'] = torch.randn(seq_len, 10).to(device)
torch_param['transl'] = torch.randn(seq_len, 3).to(device)
torch_param['global_orient'] = torch.randn(seq_len, 3).to(device)
torch_param['left_hand_pose'] = torch.randn(seq_len, 45).to(device)
torch_param['right_hand_pose'] = torch.randn(seq_len, 45).to(device)

static_model = smplx.create(model_path='data/models_smplx_v1_1/models',
                            model_type='smplx',
                            gender='neutral',
                            num_betas=10,
                            use_pca=False,
                            batch_size=seq_len,
                            ext='npz')

static_model = static_model.to(device)

for i in me.track_iter_progress(range(100)):
    if test_mode in ['cpu', 'cuda']:
        model = smplx.create(model_path='data/models_smplx_v1_1/models',
                             model_type='smplx',
                             gender='neutral',
                             num_betas=10,
                             use_pca=False,
                             batch_size=seq_len,
                             ext='npz').to(device)
        output = model(return_verts=True, **torch_param)
    elif test_mode == 'cuda_static':
        output = static_model(return_verts=True, **torch_param)

When test_mode = 'cpu': [>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 100/100, 1.7 task/s, elapsed: 58s, ETA: 0s When test_mode = 'cuda': [>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 100/100, 1.7 task/s, elapsed: 60s, ETA: 0s When test_mode = 'cuda_static': [>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 100/100, 40.2 task/s, elapsed: 2s, ETA: 0s 40x faster when the seq_len=60, half of the max_motion_len.

[Silverster98]

Thanks for your suggestion. At the beginning of our implementation, we considered how to use SMPL-X more conveniently, while considering efficiency less. In fact, we decide to implement a smpl-kit to use SMPL more easily and efficiently. If you are interested in this project, we can contribute to this repo together.