diff --git a/examples/pruning_filters_for_efficient_convnets/vgg19.schedule_filter_rank.yaml b/examples/pruning_filters_for_efficient_convnets/vgg19.schedule_filter_rank.yaml new file mode 100755 index 0000000000000000000000000000000000000000..6d1b86796a3b3359b0795c18a00cd0f4d89a770e --- /dev/null +++ b/examples/pruning_filters_for_efficient_convnets/vgg19.schedule_filter_rank.yaml @@ -0,0 +1,50 @@ +# +# This schedule performs 3D (filter-wise) regularization of some of the convolution layers, together with +# element-wise pruning using sensitivity-pruning. +# +# time python3 compress_classifier.py -a=vgg19 -p=50 ../../../data.imagenet --epochs=10 --lr=0.00001 --compress=../pruning_filters_for_efficient_convnets/vgg19.schedule_filter_rank.yaml --pretrained +# + +version: 1 +pruners: + vgg_manual: + class: 'L1RankedStructureParameterPruner' + reg_regims: + # 'features.module.0.weight': [0.1, '3D'] + # 'features.module.2.weight': [0.1, '3D'] + # 'features.module.5.weight': [0.1, '3D'] + 'features.module.7.weight': [0.1, '3D'] + 'features.module.10.weight': [0.1, '3D'] + 'features.module.12.weight': [0.1, '3D'] + 'features.module.14.weight': [0.1, '3D'] + 'features.module.16.weight': [0.1, '3D'] + 'features.module.19.weight': [0.1, '3D'] + 'features.module.21.weight': [0.1, '3D'] + 'features.module.23.weight': [0.1, '3D'] + 'features.module.25.weight': [0.1, '3D'] + 'features.module.28.weight': [0.1, '3D'] + 'features.module.30.weight': [0.1, '3D'] + 'features.module.32.weight': [0.1, '3D'] + 'features.module.34.weight': [0.1, '3D'] + +extensions: + net_thinner: + class: 'FilterRemover' + thinning_func_str: vgg_remove_filters + +lr_schedulers: + # Learning rate decay scheduler + pruning_lr: + class: StepLR + step_size: 50 + gamma: 0.10 + + +policies: + - pruner: + instance_name: vgg_manual + epochs: [0] + + - extension: + instance_name: net_thinner + epochs: [0]