- Data Efficient Stagewise Knowledge Distillation
This repository presents the code implementation for Stagewise Knowledge Distillation, a technique for improving knowledge transfer between a teacher model and student model.
- Install the dependencies using
conda
with therequirements.yml
fileconda env create -f environment.yml
- Setup the
stagewise-knowledge-distillation
package itselfpip install -e .
- Apart from the above mentioned dependencies, it is recommended to have an Nvidia GPU (CUDA compatible) with at least 8 GB of video memory (most of the experiments will work with 6 GB also). However, the code works with CPU only machines as well.
In this work, ResNet architectures are used. Particularly, we used ResNet10, 14, 18, 20 and 26 as student networks and ResNet34 as the teacher network. The datasets used are CIFAR10, Imagenette and Imagewoof. Note that Imagenette and Imagewoof are subsets of ImageNet.
-
Before any experiments, you need to download the data and saved weights of teacher model to appropriate locations.
-
The following script
- downloads the datasets
- saves 10%, 20%, 30% and 40% splits of each dataset separately
- downloads teacher model weights for all 3 datasets
# assuming you are in the root folder of the repository cd image_classification/scripts bash setup.sh
For detailed information on the various experiments, refer to the paper. In all the image classification experiments, the following common training arguments are listed with the possible values they can take:
- dataset (
-d
) : imagenette, imagewoof, cifar10 - model (
-m
) : resnet10, resnet14, resnet18, resnet20, resnet26, resnet34 - number of epochs (
-e
) : Integer is required - percentage of dataset (
-p
) : 10, 20, 30, 40 (don't use this argument at all for full dataset experiments) - random seed (
-s
) : Give any random seed (for reproducibility purposes) - gpu (
-g
) : Don't use unless training on CPU (in which case, use-g 'cpu'
as the argument). In case of multi-GPU systems, runCUDA_VISIBLE_DEVICES=id
in the terminal before the experiment, whereid
is the ID of your GPU according tonvidia-smi
output. - Comet ML API key (
-a
) (optional) : If you want to use Comet ML for tracking your experiments, then either put your API key as the argument or make it the default argument in thearguments.py
file. Otherwise, no need of using this argument. - Comet ML workspace (
-w
) (optional) : If you want to use Comet ML for tracking your experiments, then either put your workspace name as the argument or make it the default argument in thearguments.py
file. Otherwise, no need of using this argument.
In the following subsections, example commands for training are given for one experiment each.
Full Imagenette dataset, ResNet10
python3 no_teacher.py -d imagenette -m resnet10 -e 100 -s 0
Traditional KD (FitNets)
20% Imagewoof dataset, ResNet18
python3 traditional_kd.py -d imagewoof -m resnet18 -p 20 -e 100 -s 0
30% CIFAR10 dataset, ResNet14
python3 fsp_kd.py -d cifar10 -m resnet14 -p 30 -e 100 -s 0
10% Imagewoof dataset, ResNet26
python3 attention_transfer_kd.py -d imagewoof -m resnet26 -p 10 -e 100 -s 0
Full CIFAR10 dataset, ResNet14
python3 hinton_kd.py -d cifar10 -m resnet14 -e 100 -s 0
40% Imagenette dataset, ResNet20
python3 simultaneous_kd.py -d imagenette -m resnet20 -p 40 -e 100 -s 0
Full CIFAR10 dataset, ResNet10
python3 stagewise_kd.py -d cifar10 -m resnet10 -e 100 -s 0
In this work, ResNet backbones are used to construct symmetric U-Nets for semantic segmentation. Particularly, we used ResNet10, 14, 18, 20 and 26 as the backbones for student networks and ResNet34 as the backbone for the teacher network. The dataset used is the Cambridge-driving Labeled Video Database (CamVid).
- The following script
- downloads the data (and shifts it to appropriate folder)
- saves 10%, 20%, 30% and 40% splits of each dataset separately
- downloads the pretrained teacher weights in appropriate folder
# assuming you are in the root folder of the repository cd semantic_segmentation/scripts bash setup.sh
For detailed information on the various experiments, refer to the paper. In all the semantic segmentation experiments, the following common training arguments are listed with the possible values they can take:
- dataset (
-d
) : camvid - model (
-m
) : resnet10, resnet14, resnet18, resnet20, resnet26, resnet34 - number of epochs (
-e
) : Integer is required - percentage of dataset (
-p
) : 10, 20, 30, 40 (don't use this argument at all for full dataset experiments) - random seed (
-s
) : Give any random seed (for reproducibility purposes) - gpu (
-g
) : Don't use unless training on CPU (in which case, use-g 'cpu'
as the argument). In case of multi-GPU systems, runCUDA_VISIBLE_DEVICES=id
in the terminal before the experiment, whereid
is the ID of your GPU according tonvidia-smi
output. - Comet ML API key (
-a
) (optional) : If you want to use Comet ML for tracking your experiments, then either put your API key as the argument or make it the default argument in thearguments.py
file. Otherwise, no need of using this argument. - Comet ML workspace (
-w
) (optional) : If you want to use Comet ML for tracking your experiments, then either put your workspace name as the argument or make it the default argument in thearguments.py
file. Otherwise, no need of using this argument.
Note: Currently, there are no plans for adding Attention Transfer KD and FSP KD experiments for semantic segmentation.
In the following subsections, example commands for training are given for one experiment each.
Full CamVid dataset, ResNet10
python3 pretrain.py -d camvid -m resnet10 -e 100 -s 0
Traditional KD (FitNets)
20% CamVid dataset, ResNet18
python3 traditional_kd.py -d camvid -m resnet18 -p 20 -e 100 -s 0
40% CamVid dataset, ResNet20
python3 simultaneous_kd.py -d camvid -m resnet20 -p 40 -e 100 -s 0
10 % CamVid dataset, ResNet10
python3 stagewise_kd.py -d camvid -m resnet10 -p 10 -e 100 -s 0
If you use this code or method in your work, please cite using
@misc{kulkarni2020data,
title={Data Efficient Stagewise Knowledge Distillation},
author={Akshay Kulkarni and Navid Panchi and Sharath Chandra Raparthy and Shital Chiddarwar},
year={2020},
eprint={1911.06786},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
Built by Akshay Kulkarni, Navid Panchi and Sharath Chandra Raparthy.