This document explains each component of configuration file.
You can find example configuration files in config_examples
folder.
Maximum iteration number. Default is 100000.
Batch size. Default is 16.
Target dataset. Currently only 'car', 'chair', 'kitti' and 'synthia' is allowed.
(Not used) Only npy
format is allowed.
If this is true, use pose matrix as pose input. Default is false. Only used for scene data.
Learning rate. Default is 5e-5.
How frequently export image during training. Default is max_iterate / 10.
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].
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. Default is 256.
Export models to this folder.
A list of models to train.
Only two types are allowed.
t
for Tatarchenko15 (Pixel Generation)z
for Zhou16 (Appearance Flow)
Attention/skip connection strategy. Followings are allowed.
no
: Vanillau_net
: U-Net without attention.u_attn
: Attention U-Net.h_attn
orh
: Flow based hard attention.cr_attn
orcr
: Cross attention.mixed
: You can use different strategy for each layer by setting followingattention_strategy_details
.
A dictionary of (layer_size, strategy).
If attention_strategy
is mixed
, this is used.
Because many parameters are overlapped with training, we only explain the others. Test is done in exhaustive way.
Load models from this folder.
Export results to this folder.
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]