Skip to content

Latest commit

 

History

History
63 lines (63 loc) · 2.18 KB

README.md

File metadata and controls

63 lines (63 loc) · 2.18 KB
Raw

Configuration Document

This document explains each component of configuration file. You can find example configuration files in config_examples folder.

Train configuration

max_iterate

Maximum iteration number. Default is 100000.

batch_size

Batch size. Default is 16.

dataset

Target dataset. Currently only 'car', 'chair', 'kitti' and 'synthia' is allowed.

dataset_format

(Not used) Only npy format is allowed.

is_pose_matrix

If this is true, use pose matrix as pose input. Default is false. Only used for scene data.

lr

Learning rate. Default is 5e-5.

export_image_per

How frequently export image during training. Default is max_iterate / 10.

available_gpu_ids

If you have multiple gpus, use multiple gpu ids, e.g. [0,1,2,3]. If you have only single gpu, set this value to [0].

multiprocess_max

Maximum number of multiprocessing. This is not needed be same with available_gpu_ids. Even you have only single gpu, you can train multiple models at once. But be careful not to cause overflow.

image_size

Image size. Default is 256.

parent_folder

Export models to this folder.

model_list

A list of models to train.

model_type

Only two types are allowed.

  • t for Tatarchenko15 (Pixel Generation)
  • z for Zhou16 (Appearance Flow)

attention_strategy

Attention/skip connection strategy. Followings are allowed.

  • no : Vanilla
  • u_net : U-Net without attention.
  • u_attn : Attention U-Net.
  • h_attn or h : Flow based hard attention.
  • cr_attn or cr : Cross attention.
  • mixed : You can use different strategy for each layer by setting following attention_strategy_details.

attention_strategy_details

A dictionary of (layer_size, strategy). If attention_strategy is mixed, this is used.

Test configuration

Because many parameters are overlapped with training, we only explain the others. Test is done in exhaustive way.

parent_folder

Load models from this folder.

result_export_folder

Export results to this folder.

target_scene_infos

Scene number to export. It should be in form of

  • objects : [model id, input azimuth, input elevation, target azimuth, target elevation].
  • scenes : [scene id, input frame, target frame]