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

# dataset configs
batch_size: 4
num_workers: 4
pose_dir: ./datasets/densepose
texture_ann_file: ./datasets/texture_ann/test
shape_ann_path: ./datasets/shape_ann/test_ann_file.txt
downsample_factor: 2

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

# bottom level vqgan
bot_n_embed: 512
bot_codebook_spatial_size: 2
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
index_pred_encoder_in_channels: 256
index_pred_fc_in_channels: 64
index_pred_fc_in_index: 4
index_pred_fc_channels: 64
index_pred_fc_num_convs: 1
index_pred_fc_concat_input: False
index_pred_fc_dropout_ratio: 0.1
index_pred_fc_num_classes: 512
index_pred_fc_align_corners: False
pretrained_index_network: ./pretrained_models/index_pred_net.pth

# 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
segm_token_path: ./pretrained_models/parsing_token.pth

# 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
pretrained_sampler: ./pretrained_models/sampler.pth

# shape network configs
shape_embedder_dim: 8
shape_embedder_out_dim: 128
shape_attr_class_num: [2, 4, 6, 5, 4, 3, 5, 5, 3, 2, 2, 2, 2, 2, 2]
shape_encoder_in_channels: 1
shape_fc_in_channels: 64
shape_fc_in_index: 4
shape_fc_channels: 64
shape_fc_num_convs: 1
shape_fc_concat_input: False
shape_fc_dropout_ratio: 0.1
shape_fc_num_classes: 24
shape_fc_align_corners: False
pretrained_parsing_gen: ./pretrained_models/parsing_gen.pth

manual_seed: 2021
sample_steps: 256