options/option_vqvae.py

import argparse

def get_args_parser():
    parser = argparse.ArgumentParser(description='Optimal Transport AutoEncoder training for Amass',
                                     add_help=True,
                                     formatter_class=argparse.ArgumentDefaultsHelpFormatter)

    ## dataloader
    parser.add_argument('--dataname', type=str, default='t2m', help='dataset directory')
    parser.add_argument('--out_dir', type=str, default='./out/', help='output directory')
    parser.add_argument('--resume_pth', type=str, help='path to saved vqvae model')
    parser.add_argument('--window_size', type=int, default=64, help='training motion length')

    ## train
    parser.add_argument('--batch_size', type=int, default=256, help='batch size')
    parser.add_argument('--learning_rate', type=float, default=2e-4, help='learning rate')
    parser.add_argument('--weight_decay', type=float, default=0.0, help='weight decay')
    parser.add_argument('--warmup_steps', type=int, default=1000, help='number of total iterations for warmup')
    parser.add_argument('--total_iter', default=300000, type=int, help='number of total iterations to run')
    parser.add_argument('--lr', default=2e-4, type=float, help='max learning rate')
    parser.add_argument('--lr_scheduler', default=[200000], nargs="+", type=int, help="learning rate schedule (iterations)")
    parser.add_argument('--gamma', default=0.05, type=float, help="learning rate decay")
    parser.add_argument("--commit", type=float, default=0.02, help="hyper-parameter for the commitment loss")
    parser.add_argument('--loss_vel', type=float, default=0.5, help='hyper-parameter for the velocity loss')
    parser.add_argument('--recons_loss', type=str, default='l1_smooth', help='reconstruction loss')
    parser.add_argument('--print_iter', default=200, type=int, help='print frequency')
    parser.add_argument('--eval_iter', default=1000, type=int, help='evaluation frequency')
    parser.add_argument('--seed', default=123, type=int, help='seed for initializing training.')

    ## model
    parser.add_argument("--code_dim", type=int, default=512, help="embedding dimension")
    parser.add_argument("--nb_code", type=int, default=512, help="nb of embedding")
    parser.add_argument("--mu", type=float, default=0.99, help="exponential moving average to update the codebook")
    parser.add_argument("--down_t", type=int, default=2, help="downsampling rate")
    parser.add_argument("--stride_t", type=int, default=2, help="stride size")
    parser.add_argument("--width", type=int, default=512, help="width of the network")
    parser.add_argument("--depth", type=int, default=3, help="depth of the network")
    parser.add_argument("--dilation_growth_rate", type=int, default=3, help="dilation growth rate")
    parser.add_argument("--output_emb_width", type=int, default=512, help="output embedding width")
    parser.add_argument('--vq_act', type=str, default='relu', choices = ['relu', 'silu', 'gelu'], help='dataset directory')
    parser.add_argument('--vq_norm', type=str, default=None, help='dataset directory')

    ## quantizer
    parser.add_argument("--quantizer", type=str, default='ema_reset', choices = ['ema', 'orig', 'ema_reset', 'reset'], help="eps for optimal transport")
    parser.add_argument('--beta', type=float, default=1.0, help='commitment loss in standard VQ')

    return parser.parse_args()