Sparse training

UP supports sparse training; Codes

Configs

It contains the illustration of common configs.

Repos

Dataset related modules

1. Dataset types:

• imagenet

• custom_cls

• coco

2. The type of datasets can be chosen by setting ‘type’ in Dataset. The config is as followed.

dataset:
  type: cls * or coco
  kwargs:
    ...

3. The following written should refer to the chosen dataset.

Sparse training setting

runtime:
  runner:
    type: sparse

sparsity:
  mask_generator:
    type: NormalMaskGenerator
  fake_sparse:
    type: FakeSparse
  scheduler:
    type: AmbaLevelPruneScheduler
    kwargs:
      total_iters: None
      sparsity_table: [30,40,50,60]
      no_prune_keyword: ''
      no_prune_layer: ''
      prun_algo: 1
      prun_algo_tuning: 0.5
      dw_no_prune: False
      do_sparse_analysis: False
      output_dir: path_to/amba/faster_rcnn_r50_fpn_improve_amba_sparse_30_to_90/sparse_analysis
      save_dir: path_to/amba/faster_rcnn_r50_fpn_improve_amba_sparse_30_to_90/sparse_ckpts
  leaf_module: [Space2Depth, FrozenBatchNorm2d]