Upload 3 files

options/get_eval_option.py

@@ -0,0 +1,83 @@
+from argparse import Namespace
+import re
+from os.path import join as pjoin
+
+
+def is_float(numStr):
+    flag = False
+    numStr = str(numStr).strip().lstrip('-').lstrip('+')
+    try:
+        reg = re.compile(r'^[-+]?[0-9]+\.[0-9]+$')
+        res = reg.match(str(numStr))
+        if res:
+            flag = True
+    except Exception as ex:
+        print("is_float() - error: " + str(ex))
+    return flag
+
+
+def is_number(numStr):
+    flag = False
+    numStr = str(numStr).strip().lstrip('-').lstrip('+')
+    if str(numStr).isdigit():
+        flag = True
+    return flag
+
+
+def get_opt(opt_path, device):
+    opt = Namespace()
+    opt_dict = vars(opt)
+
+    skip = ('-------------- End ----------------',
+            '------------ Options -------------',
+            '\n')
+    print('Reading', opt_path)
+    with open(opt_path) as f:
+        for line in f:
+            if line.strip() not in skip:
+                # print(line.strip())
+                key, value = line.strip().split(': ')
+                if value in ('True', 'False'):
+                    opt_dict[key] = (value == 'True')
+                #     print(key, value)
+                elif is_float(value):
+                    opt_dict[key] = float(value)
+                elif is_number(value):
+                    opt_dict[key] = int(value)
+                else:
+                    opt_dict[key] = str(value)
+
+    # print(opt)
+    opt_dict['which_epoch'] = 'finest'
+    opt.save_root = pjoin(opt.checkpoints_dir, opt.dataset_name, opt.name)
+    opt.model_dir = pjoin(opt.save_root, 'model')
+    opt.meta_dir = pjoin(opt.save_root, 'meta')
+
+    if opt.dataset_name == 't2m':
+        opt.data_root = './dataset/Sample1/'
+        opt.motion_dir = pjoin(opt.data_root, 'new_joint_vecs')
+        opt.text_dir = pjoin(opt.data_root, 'texts')
+        opt.joints_num = 22
+        opt.dim_pose = 263
+        opt.max_motion_length = 196
+        opt.max_motion_frame = 196
+        opt.max_motion_token = 55
+    elif opt.dataset_name == 'kit':
+        opt.data_root = './dataset/KIT-ML/'
+        opt.motion_dir = pjoin(opt.data_root, 'new_joint_vecs')
+        opt.text_dir = pjoin(opt.data_root, 'texts')
+        opt.joints_num = 21
+        opt.dim_pose = 251
+        opt.max_motion_length = 196
+        opt.max_motion_frame = 196
+        opt.max_motion_token = 55
+    else:
+        raise KeyError('Dataset not recognized')
+
+    opt.dim_word = 300
+    opt.num_classes = 200 // opt.unit_length
+    opt.is_train = False
+    opt.is_continue = False
+    opt.device = device
+
+    return opt

options/option_transformer.py

@@ -0,0 +1,68 @@
+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='kit', help='dataset directory')
+    parser.add_argument('--batch-size', default=128, type=int, help='batch size')
+    parser.add_argument('--fps', default=[20], nargs="+", type=int, help='frames per second')
+    parser.add_argument('--seq-len', type=int, default=64, help='training motion length')
+    
+    ## optimization
+    parser.add_argument('--total-iter', default=100000, type=int, help='number of total iterations to run')
+    parser.add_argument('--warm-up-iter', default=1000, type=int, help='number of total iterations for warmup')
+    parser.add_argument('--lr', default=2e-4, type=float, help='max learning rate')
+    parser.add_argument('--lr-scheduler', default=[60000], nargs="+", type=int, help="learning rate schedule (iterations)")
+    parser.add_argument('--gamma', default=0.05, type=float, help="learning rate decay")
+    
+    parser.add_argument('--weight-decay', default=1e-6, type=float, help='weight decay') 
+    parser.add_argument('--decay-option',default='all', type=str, choices=['all', 'noVQ'], help='disable weight decay on codebook')
+    parser.add_argument('--optimizer',default='adamw', type=str, choices=['adam', 'adamw'], help='disable weight decay on codebook')
+    
+    ## vqvae arch
+    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=3, 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')
+
+    ## gpt arch
+    parser.add_argument("--block-size", type=int, default=25, help="seq len")
+    parser.add_argument("--embed-dim-gpt", type=int, default=512, help="embedding dimension")
+    parser.add_argument("--clip-dim", type=int, default=512, help="latent dimension in the clip feature")
+    parser.add_argument("--num-layers", type=int, default=2, help="nb of transformer layers")
+    parser.add_argument("--n-head-gpt", type=int, default=8, help="nb of heads")
+    parser.add_argument("--ff-rate", type=int, default=4, help="feedforward size")
+    parser.add_argument("--drop-out-rate", type=float, default=0.1, help="dropout ratio in the pos encoding")
+    
+    ## quantizer
+    parser.add_argument("--quantizer", type=str, default='ema_reset', choices = ['ema', 'orig', 'ema_reset', 'reset'], help="eps for optimal transport")
+    parser.add_argument('--quantbeta', type=float, default=1.0, help='dataset directory')
+
+    ## resume
+    parser.add_argument("--resume-pth", type=str, default=None, help='resume vq pth')
+    parser.add_argument("--resume-trans", type=str, default=None, help='resume gpt pth')
+    
+    
+    ## output directory 
+    parser.add_argument('--out-dir', type=str, default='output_GPT_Final/', help='output directory')
+    parser.add_argument('--exp-name', type=str, default='exp_debug', help='name of the experiment, will create a file inside out-dir')
+    parser.add_argument('--vq-name', type=str, default='exp_debug', help='name of the generated dataset .npy, will create a file inside out-dir')
+    ## other
+    parser.add_argument('--print-iter', default=200, type=int, help='print frequency')
+    parser.add_argument('--eval-iter', default=5000, type=int, help='evaluation frequency')
+    parser.add_argument('--seed', default=123, type=int, help='seed for initializing training. ')
+    parser.add_argument("--if-maxtest", action='store_true', help="test in max")
+    parser.add_argument('--pkeep', type=float, default=1.0, help='keep rate for gpt training')
+    
+    
+    return parser.parse_args()

options/option_vq.py

@@ -0,0 +1,61 @@
+import argparse
+
+def get_args_parser():
+    parser = argparse.ArgumentParser(description='Optimal Transport AutoEncoder training for AIST',
+                                     add_help=True,
+                                     formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+
+    ## dataloader  
+    parser.add_argument('--dataname', type=str, default='kit', help='dataset directory')
+    parser.add_argument('--batch-size', default=128, type=int, help='batch size')
+    parser.add_argument('--window-size', type=int, default=64, help='training motion length')
+
+    ## optimization
+    parser.add_argument('--total-iter', default=200000, type=int, help='number of total iterations to run')
+    parser.add_argument('--warm-up-iter', default=1000, type=int, help='number of total iterations for warmup')
+    parser.add_argument('--lr', default=2e-4, type=float, help='max learning rate')
+    parser.add_argument('--lr-scheduler', default=[50000, 400000], nargs="+", type=int, help="learning rate schedule (iterations)")
+    parser.add_argument('--gamma', default=0.05, type=float, help="learning rate decay")
+
+    parser.add_argument('--weight-decay', default=0.0, type=float, help='weight 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=0.1, help='hyper-parameter for the velocity loss')
+    parser.add_argument('--recons-loss', type=str, default='l2', help='reconstruction loss')
+    
+    ## vqvae arch
+    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')
+
+    ## resume
+    parser.add_argument("--resume-pth", type=str, default=None, help='resume pth for VQ')
+    parser.add_argument("--resume-gpt", type=str, default=None, help='resume pth for GPT')
+    
+    
+    ## output directory 
+    parser.add_argument('--out-dir', type=str, default='output_vqfinal/', help='output directory')
+    parser.add_argument('--results-dir', type=str, default='visual_results/', help='output directory')
+    parser.add_argument('--visual-name', type=str, default='baseline', help='output directory')
+    parser.add_argument('--exp-name', type=str, default='exp_debug', help='name of the experiment, will create a file inside out-dir')
+    ## other
+    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.')
+    
+    parser.add_argument('--vis-gt', action='store_true', help='whether visualize GT motions')
+    parser.add_argument('--nb-vis', default=20, type=int, help='nb of visualizations')
+    
+    
+    return parser.parse_args()