This repository contains code for the following paper:

Deepak Mittal, Shweta Bhardwaj, Balaraman Ravindran, Mitesh M. Khapra. Recovering from Random Pruning: On the Plasticity of Deep Convolutional Neural Networks. IEEE Conference on Winter Applications in Computer Vision 2018 [https://arxiv.org/abs/1801.10447].

• Requirements
• Dataset
• PreTrained Models
• Code organization

Requirements

• tensorflow: 1.3.0
• skimage: 0.14.5
• Python: 2.7.18 (supports Python3)
• tqdm: 4.59.0
• cPickle: 1.71
• argparse: 1.1

Dataset

• ImageNet-1000: 2017 version
• To convert raw images to tf-records, you can refer to this code: https://github.com/shwetabhardwaj44/ImageNet_images_to_TFRecords

PreTrained Models

Pretrained full VGG-16 model trained on ImageNet is uploaded here: https://drive.google.com/file/d/103FkgQqjClsBjx9PHVKNdV5RnSywI2qy/view?usp=sharing. Save this checkpoint in model_baseline folder under models folder.

Code Organization

1. Shell Scripts:

• run_TrainAndPrune.sh: Stage-I
• run_Finetune.sh: Stage-II

2. Main Code Files:

• train_and_prune.py: Binary to prune layer-by-layer and train the model for one epoch.
• finetune_PrunedModel.py: Binary to fine-tune (re-train) the final pruned model for around 20-25 epochs.

3. Pruning Masks:

• generate_RandomMask.py: generates mask_ratio_random.save according to the defined "prune_factor"
• generate_EntropyMask.py:
• generate_ScaledEntropyMask.py:
• generate_L1normMask.py:

4. Configuration Files:

• config.py:
• vgg_preprocessing.py: