TRAINING.md

Training

Below, we provide the training commands for training Segmenter + ALGM on the following datasets:

• ADE20K
• Cityscapes
• Pascal-Context

Different backbone possibilities:

• ViT-Ti: vit_tiny_patch16_384
• ViT-S: vit_small_patch16_384
• ViT-B: vit_base_patch16_384
• ViT-L: vit_large_patch16_384

For more configuration options, see segm/config.yml and segm/train.py.

ADE20K

Segmenter + ALGM, and ViT-Ti backbone:

python -m segm.train    --log-dir runs/vit_tiny_layers_1_5_T_0.9 \
                        --dataset ade20k \
                        --backbone vit_tiny_patch16_384 \
                        --decoder mask_transformer \
                        --patch-type algm \
                        --selected-layers 1 5 \
                        --threshold 0.9

Segmenter + ALGM, and ViT-S backbone:

python -m segm.train  --log-dir runs/vit_small_layers_1_5_T_0.88/ \
                      --dataset ade20k \
                      --backbone vit_small_patch16_384 \
                      --decoder mask_transformer \
                      --patch-type algm \
                      --selected-layers 1 5 \
                      --merging-window-size 2 2 \
                      --threshold 0.88 

Segmenter + ALGM, and ViT-B backbone:

python -m segm.train    --log-dir runs/vit_base_layers_1_5_T_0.94 \
                        --dataset ade20k \
                        --backbone vit_base_patch16_384 \
                        --decoder mask_transformer \
                        --patch-type algm \
                        --selected-layers 1 5 \
                        --threshold 0.94

Segmenter + ALGM, and ViT-L backbone:

python -m segm.train    --log-dir runs/vit_large_layers_1_7_T_0.95 \
                        --dataset ade20k_large \
                        --backbone vit_large_patch16_384 \
                        --decoder mask_transformer \
                        --patch-type algm \
                        --selected-layers 1 7 \
                        --threshold 0.95

Cityscapes

Segmenter + ALGM, and ViT-S backbone:

python -m segm.train  --log-dir runs/vit_small_layers_1_5_T_0./ \
                      --dataset Cityscapes \
                      --backbone vit_small_patch16_384 \
                      --decoder mask_transformer \
                      --patch-type algm \
                      --selected-layers 1 5 \
                      --merging-window-size 2 2 \
                      --threshold  0.955

Pascal-Context

Segmenter + ALGM, and ViT-S backbone:

python -m segm.train  --log-dir runs/vit_small_layers_1_5_T_0./ \
                      --dataset  pascal_context \
                      --backbone vit_small_patch16_384 \
                      --decoder mask_transformer \
                      --patch-type algm \
                      --selected-layers 1 5 \
                      --merging-window-size 2 2 \
                      --threshold 0.88