Adjoint-Network

Setup

Setting up conda

1. wget https://repo.anaconda.com/archive/Anaconda3-2018.12-Linux-x86_64.sh
   
2. bash Anaconda3–2018.12-Linux-x86_64.sh
   
3. . ~/.bashrc

Installing libraries

1. conda create -n fastai python=3.7
   
2. conda activate fastai
   
3. pip install fastai
   
4. conda install pytorch torchvision cudatoolkit=10.2 -c pytorch
   
5. conda install -c anaconda ipython
   

Running

python train.py --dataset cifar100 --compression_factor 16 --masking_factor 0.5
Running the above command would use default value for batch_size, image_size, lr, c, epoch and is_sgd.
If you want to change anyone of these value then use --default_config False
eg. python train.py --dataset cifar100 --default_config False --compression_factor 16 --masking_factor 0.5 --lr 0.1

Arguments

Argument	Discription	Domain
lr	Learning Rate	Float
is_adjoint_training	True for adjoint training, false otherwise	True/False
is_sgd	We support sgd and adam both	True/False
classes	Denote number of classes in the dataset	Integer
compression_factor	Normally used compression factors are 4,8,16. Default value is 4	Integer
masking_factor	Used to denote fraction of weight converted to zero in each kernel. Default value is None	(0,1)
resnet	the resnet model to be used	{18,34,50,101,152}
dataset	Dataset supported are cifar100, cifar10, imagenet, imagewoof and Oxford-IIIT Pet	{cifar100,cifar10,imagenet,imagewoof,pets}
default_config	Setting it to true will be using parameter used in the currect experiemnt for each dataset. By default it's true	True/False
batch_size	Batch size	Integer
image_size	Image size	Integer
epoch	Total number of epochs	Integer