|
from transformers import PretrainedConfig |
|
|
|
class in2INConfig(PretrainedConfig): |
|
def __init__(self, |
|
num_layers=8, |
|
num_heads=8, |
|
dropout=0.1, |
|
input_dim=262, |
|
latent_dim=1024, |
|
ff_size=2048, |
|
activation="gelu", |
|
diffusion_steps=1000, |
|
beta_scheduler="cosine", |
|
sampler="uniform", |
|
motion_rep="global", |
|
finetune=False, |
|
text_encoder="clip", |
|
t_bar=700, |
|
control="text", |
|
strategy="ddim50", |
|
cfg_weight=3, |
|
cfg_weight_interaction=3, |
|
cfg_weight_individual=1, |
|
mode="interaction", |
|
**kwargs): |
|
|
|
self.NUM_LAYERS = num_layers |
|
self.NUM_HEADS = num_heads |
|
self.DROPOUT = dropout |
|
self.INPUT_DIM = input_dim |
|
self.LATENT_DIM = latent_dim |
|
self.FF_SIZE = ff_size |
|
self.ACTIVATION = activation |
|
self.DIFFUSION_STEPS = diffusion_steps |
|
self.BETA_SCHEDULER = beta_scheduler |
|
self.SAMPLER = sampler |
|
self.MOTION_REP = motion_rep |
|
self.FINETUNE = finetune |
|
self.TEXT_ENCODER = text_encoder |
|
self.T_BAR = t_bar |
|
self.CONTROL = control |
|
self.STRATEGY = strategy |
|
self.CFG_WEIGHT = cfg_weight |
|
self.CFG_WEIGHT_INTERACTION = cfg_weight_interaction |
|
self.CFG_WEIGHT_INDIVIDUAL = cfg_weight_individual |
|
self.MODE = mode |
|
|
|
super().__init__(**kwargs) |
|
|