Text2Human/configs/sampler.yml

name: sampler
use_tb_logger: true
set_CUDA_VISIBLE_DEVICES: ~
gpu_ids: [3]

# dataset configs
batch_size: 4
num_workers: 1
train_img_dir: ./datasets/train_images
test_img_dir: ./datasets/test_images
segm_dir: ./datasets/segm
pose_dir: ./datasets/densepose
train_ann_file: ./datasets/texture_ann/train
val_ann_file: ./datasets/texture_ann/val
test_ann_file: ./datasets/texture_ann/test
downsample_factor: 2

# pretrained models
img_ae_path: ./pretrained_models/vqvae_top.pth
segm_ae_path: ./pretrained_models/parsing_token.pth

model_type: TransformerTextureAwareModel
# network configs

# image autoencoder
img_embed_dim: 256
img_n_embed: 1024
img_double_z: false
img_z_channels: 256
img_resolution: 512
img_in_channels: 3
img_out_ch: 3
img_ch: 128
img_ch_mult: [1, 1, 2, 2, 4]
img_num_res_blocks: 2
img_attn_resolutions: [32]
img_dropout: 0.0

# segmentation tokenization
segm_double_z: false
segm_z_channels: 32
segm_resolution: 512
segm_in_channels: 24
segm_out_ch: 24
segm_ch: 64
segm_ch_mult: [1, 1, 2, 2, 4]
segm_num_res_blocks: 1
segm_attn_resolutions: [16]
segm_dropout: 0.0
segm_num_segm_classes: 24
segm_n_embed: 1024
segm_embed_dim: 32

# sampler configs
codebook_size: 18432
segm_codebook_size: 1024
texture_codebook_size: 18
bert_n_emb: 512
bert_n_layers: 24
bert_n_head: 8
block_size: 512 # 32 x 16
latent_shape: [32, 16]
embd_pdrop: 0.0
resid_pdrop: 0.0
attn_pdrop: 0.0
num_head: 18

# loss configs
loss_type: reweighted_elbo
mask_schedule: random

sample_steps: 256

# training configs
val_freq: 5
print_freq: 100
weight_decay: 0
manual_seed: 2021
num_epochs: 100
lr: !!float 1e-4
lr_decay: step
gamma: 1.0
step: 50