Allow passing SMPL/HumanML3D initialization parameters

[UuuNyaa]

I have made it possible to initialize SMPL/HumanML3D by specifying the model path.
Now, you can load a model of any path with the following code:

class dotdict(dict):
    """dot.notation access to dictionary attributes"""
    __getattr__ = dict.get
    __setattr__ = dict.__setitem__
    __delattr__ = dict.__delitem__

SMPL_DATA_PATH = os.path.join(DATA_PATH, 'body_models', 'smpl')
SMPL_MODEL_PATH = os.path.join(SMPL_DATA_PATH, 'SMPL_NEUTRAL.pkl')
JOINT_REGRESSOR_TRAIN_EXTRA_PATH = os.path.join(SMPL_DATA_PATH, 'J_regressor_extra.npy')

data_loader = data.DataLoader(
    HumanML3D(mode='train', base_path=DATA_PATH, split='test', num_frames=max_frames),
    batch_size=batch_size, shuffle=True,
    num_workers=8, drop_last=True, collate_fn=t2m_collate
)

logging.info("Creating model and diffusion...")
model = model_util.MDM(
    smpl_model_path=SMPL_MODEL_PATH,
    joint_regressor_train_extra_path=JOINT_REGRESSOR_TRAIN_EXTRA_PATH,
    **model_util.get_model_args(dotdict({
        'dataset': dataset,
        'latent_dim': 512,
        'layers': 8,
        'cond_mask_prob': 0.1,
        'arch': 'trans_enc',
        'emb_trans_dec': False,
    }), data_loader)
)

diffusion = model_util.create_gaussian_diffusion(dotdict({
    'diffusion_steps': 1000,
    'diffusion_sampling_steps': 100,
    'noise_schedule': 'cosine',
    'sigma_small': True,
    'lambda_vel': 0.0,
    'lambda_rcxyz': 0.0,
    'lambda_fc': 0.0,
}))

[UuuNyaa]

Now we can specify the sampling timesteps for ddim.

[UuuNyaa]

I did motion editing (In-Betweening) on my CPU with ddim sampling.

$ time PYTHONPATH=mdm python3 motion_edit.py
TRANS_ENC init
EMBED TEXT
Loading CLIP...
100%|████████████| 100/100 [00:19<00:00,  5.19it/s]

real    0m23.509s
user    3m57.046s
sys     0m3.641s

CPU model and Settings

• CPU: AMD Ryzen 9 3900XT 12-Core Processor

    seed = 1
    max_frames = 196
    guidance_param = 2.5
    device = 0
    text_condition = "a person jumps"
    num_samples = 1
    batch_size = 1
    dataset = 'humanml'
    prefix_end = 0.25
    suffix_start = 0.75
    diffusion_steps = 1000
    diffusion_sampling_steps = 100

    input_motion = np.load(os.path.join(DATA_PATH, 'dataset', 'joints', '000021.npy'))[:, :22]