- Datasets are at Google Drive: /tmp/[DATASET_NAME], DATASET_NAME: {flowers102, cub200, cars, dogs})
- Datasets are in tfrecord format.
- Well-trained models and finetuned checkpoints are stored at Google Drive: ./tmp/[DATASET_NAME]-models/[MODEL_NAME]-glb_only (MODEL_NAME: {inception_v2, inception_v3, resnet_v1_50, resnet_v1_101, vgg_16})
- Finetune code: finetune_checkpoint.py
- Example script:
./finetune_scripts/run_finetune_checkpoint_[MODEL_NAME].sh
In this first experiment, we conduct a head-to-head com- parison to a recent incremental training method for CNN pruning.
- Code: block_sensitivity.py
- Example script:
./basemix_scripts/run_block_sensitivity.sh
- Code: basemix_greedy.py
- Example script:
./basemix_scripts/run_basemix_greedy.sh
- Stage 1: p1_train_manual_configs.py;
- Example script:
./p1_train_scripts/run_p1_train_manual_configs.sh - Stage 2: p2_finetune_manual_configs.py
- Example script:
./p2_train_scripts/run_p2_finetune_manual_configs.sh - configurations are in ./configs_greedy/
In this experiment, we compare GTSN-based training with the enumerative method. The objective of the pruning is to find the smallest network that meets a predefined accuracy target.
- Code: basemix_[MODEL_NAME].py
- Example script:
./basemix_scripts/run_basemix_[MODEL_NAME].sh - configurations are in ./configs_enum
- Code: basemix_with_distillation_[MODEL_NAME].py
- Example script:
./basemix_scripts/run_basemix_with_distillation_[MODEL_NAME].sh
- Stage 1: p1_train_[MODEL_NAME].py;
- Example script:
./p1_train_scripts/run_p1_train_[MODEL_NAME].sh - Stage 2: p2_finetune_manual_configs.py
- Example script:
./p2_train_scripts/run_p2_finetune_[MODEL_NAME].sh
- Tested on tensorflow version 1.3.0, 1.7.0
- Required packages: tensorflow, mpi4py