| # Use this file to finetune from a pretrained 256x256 model | |
| dataset_params: | |
| root_dir: ./video-preprocessing/vox2-768 | |
| frame_shape: 768,768,3 | |
| id_sampling: True | |
| augmentation_params: | |
| flip_param: | |
| horizontal_flip: True | |
| time_flip: True | |
| jitter_param: | |
| brightness: 0.1 | |
| contrast: 0.1 | |
| saturation: 0.1 | |
| hue: 0.1 | |
| model_params: | |
| common_params: | |
| num_tps: 10 | |
| num_channels: 3 | |
| bg: True | |
| multi_mask: True | |
| generator_params: | |
| block_expansion: 64 | |
| max_features: 512 | |
| num_down_blocks: 3 | |
| dense_motion_params: | |
| block_expansion: 64 | |
| max_features: 1024 | |
| num_blocks: 5 | |
| scale_factor: 0.25 | |
| avd_network_params: | |
| id_bottle_size: 128 | |
| pose_bottle_size: 128 | |
| train_params: | |
| visualize_model: False | |
| num_epochs: 80 | |
| num_repeats: 10 | |
| # Higher LR seems to bring problems when finetuning | |
| lr_generator: 3.0e-5 | |
| batch_size: 2 | |
| scales: [1, 0.5, 0.25, 0.125, 0.0625] | |
| dataloader_workers: 8 | |
| checkpoint_freq: 2 | |
| dropout_epoch: 0 | |
| dropout_maxp: 0.3 | |
| dropout_startp: 0.1 | |
| dropout_inc_epoch: 10 | |
| bg_start: 81 | |
| freeze_kp_detector: True | |
| freeze_bg_predictor: True | |
| transform_params: | |
| sigma_affine: 0.05 | |
| sigma_tps: 0.005 | |
| points_tps: 5 | |
| loss_weights: | |
| perceptual: [10, 10, 10, 10, 10] | |
| equivariance_value: 10 | |
| warp_loss: 10 | |
| bg: 10 | |
| optimizer: 'adamw' | |
| optimizer_params: | |
| betas: [ 0.9, 0.999 ] | |
| weight_decay: 0.1 | |
| train_avd_params: | |
| num_epochs: 200 | |
| num_repeats: 1 | |
| batch_size: 1 | |
| dataloader_workers: 6 | |
| checkpoint_freq: 1 | |
| epoch_milestones: [140, 180] | |
| lr: 1.0e-3 | |
| lambda_shift: 1 | |
| random_scale: 0.25 | |
| visualizer_params: | |
| kp_size: 5 | |
| draw_border: True | |
| colormap: 'gist_rainbow' | |