Text2Human/configs/index_pred_net.yml

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

# dataset configs
batch_size: 4
num_workers: 4
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

model_type: VQGANTextureAwareSpatialHierarchyInferenceModel
# network configs
embed_dim: 256
n_embed: 1024
codebook_spatial_size: 2

# bottom level vqvae
bot_n_embed: 512
bot_double_z: false
bot_z_channels: 256
bot_resolution: 512
bot_in_channels: 3
bot_out_ch: 3
bot_ch: 128
bot_ch_mult: [1, 1, 2, 4]
bot_num_res_blocks: 2
bot_attn_resolutions: [64]
bot_dropout: 0.0
bot_vae_path: ./pretrained_models/vqvae_bottom.pth

# top level vqgan
top_double_z: false
top_z_channels: 256
top_resolution: 512
top_in_channels: 3
top_out_ch: 3
top_ch: 128
top_ch_mult: [1, 1, 2, 2, 4]
top_num_res_blocks: 2
top_attn_resolutions: [32]
top_dropout: 0.0
top_vae_path: ./pretrained_models/vqvae_top.pth

# unet configs
encoder_in_channels: 256
fc_in_channels: 64
fc_in_index: 4
fc_channels: 64
fc_num_convs: 1
fc_concat_input: False
fc_dropout_ratio: 0.1
fc_num_classes: 512
fc_align_corners: False

disc_layers: 3
disc_weight_max: 1
disc_start_step: 30001
n_channels: 3
ndf: 64
nf: 128
perceptual_weight: 1.0

num_segm_classes: 24

# training configs
val_freq: 5
print_freq: 100
weight_decay: 0
manual_seed: 2021
num_epochs: 100
lr: !!float 1.0e-04
lr_decay: step
gamma: 1.0
step: 50
optimizer: Adam
loss_function: cross_entropy