-
Notifications
You must be signed in to change notification settings - Fork 54
/
Copy pathconfig_mld_humanml3d.yaml
116 lines (106 loc) · 3.36 KB
/
config_mld_humanml3d.yaml
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
####################################
# The following are general settings
####################################
# Experiment name, more details in Section 'Experiment Name Explanation'
NAME: 1222_PELearn_Diff_Latent1_MEncDec49_MdiffEnc49_bs64_clip_uncond75_01
# Debug mode. Set to True will enter the debug mode, then the program will
# 1. use a tiny dataset for trianing and evaluation
# 2. validate more intensively
# 3. will not use `wandb logger`
DEBUG: False
# Devices. Optional: “cpu”, “gpu”
ACCELERATOR: 'gpu'
# Index of GPUs eg. [0] or [0,1,2,3]
DEVICE: [0]
#####################################
# The following are training settings
#####################################
TRAIN:
# Model stage. Optional: "vae", "diffusion"
STAGE: diffusion
# Training dataset name
DATASETS: ['humanml3d']
# Number of dataloader workers
NUM_WORKERS: 8
# Size of batches
BATCH_SIZE: 64
# Total epochs for training
END_EPOCH: 2000
RESUME: '' # Resume training from this path
PRETRAINED_VAE: '' # vae model path
OPTIM:
TYPE: AdamW # Optimizer type
LR: 1e-4 # Learning rate
# Ablation study configurations.
ABLATION:
SKIP_CONNECT: True
PE_TYPE: mld
DIFF_PE_TYPE: mld
#####################################
# The following are validation settings
#####################################
EVAL:
DATASETS: ['humanml3d'] # Evaluating datasets
BATCH_SIZE: 32 # Evaluating Batch size
SPLIT: test
#####################################
# The following are testing settings
#####################################
TEST:
CHECKPOINTS: ./models/mld_humanml3d_checkpoint/1222_mld_humanml3d_FID041.ckpt # Pretrained model path
DATASETS: ['humanml3d'] # training datasets
SPLIT: test
BATCH_SIZE: 1 # training Batch size
MEAN: False
NUM_SAMPLES: 1
FACT: 1
#####################################
# The following are basic datasets settings
#####################################
DATASET:
JOINT_TYPE: 'humanml3d' # join type
#####################################
# The following are metric settings
#####################################
METRIC:
TYPE: ['TemosMetric', 'TM2TMetrics']
#####################################
# The following are training losses settings
#####################################
LOSS:
TYPE: mld # Losses type
LAMBDA_LATENT: 1.0e-5 # Lambda for latent Losses
LAMBDA_KL: 1.0e-4 # Lambda for kl Losses
LAMBDA_REC: 1.0 # Lambda for reconstruction Losses
LAMBDA_GEN: 1.0 # Lambda for text-motion generation losses
LAMBDA_CROSS: 1.0 # Lambda for reconstruction Losses
LAMBDA_CYCLE: 0.0 # Lambda for cycle Losses
LAMBDA_PRIOR: 0.0
DIST_SYNC_ON_STEP: False # Sync Losses on step when distributed trained
#####################################
# The following are basic model settings
#####################################
model:
vae: true # whether vae model
model_type: mld # model type
condition: 'text'
latent_dim: [1, 256] # latent dimension
ff_size: 1024 #
num_layers: 9 # number of layers
num_head: 4 # number of head layers
droupout: 0.1 # dropout rate
activation: gelu # activation type
guidance_scale: 7.5 #
guidance_uncondp: 0.1 # 0.1 0.25
#####################################
# The following are loggers settings
#####################################
LOGGER:
SACE_CHECKPOINT_EPOCH: 200
LOG_EVERY_STEPS: 1
VAL_EVERY_STEPS: 200
TENSORBOARD: True
WANDB:
PROJECT: null
OFFLINE: False
RESUME_ID: null