1-Keras
2-Chainer
3-Tensorflow
4-Pytorch
5-Theano
6-Mxnet
1- MNIST
2- CIFAR10
3- CIFAR100
4- SVHN
1- PTB
2- Manythings
3- IMDB
1- VOC12
There are two experiments one of them uses CPU and the other uses GPU.
* The CPU experiments are performed on a single machine running on Centos release 7.5.1804 with 32 core Intel Xeon Processor (Skylake, IBRS) @ 2.00GHz;64 GB DIMM memory; and 240 GB SSD hard drive.
* For Keras, version 2.2.4 is used on Tensorflow 1.11.0.
* For Chainer, version 4.5.0 is used.
* For Tensorflow, version 1.11 is used.
* For Pytorch, version 0.4.1 is used
* For Theano, version 1.0.2 is used.
* For MXNet, version 1.3.0 is used.
* It contain the CPU source code
* The Generated graphs
* The logs of the experiment
* The GPU experiments are performed on a single machine running on Debian GNU/Linux 9 (stretch) with 8 core Intel(R) Xeon(R) CPU @ 2.00GHz; NVIDIA Tesla P4;36 GB DIMM memory; and 300 GB SSD hard drive.
* For Keras, version 2.2.4 is used and run on Tensorflow version 1.11.0.
* For Chainer, version 4.5.0 is used.
* For Tensorflow, version 1.11.0 is used.
* For Pytorch, version 0.4.1 is used, For Theano, version 1.0.2 is used.
* For MXNet, version 1.3.0 is used.
* For Chainer, MXNet and Pytorch, we used CUDA 9.2 and cuDNN 7.2.1, and for Tensorflow, Keras, and Theano, we used CUDA 10.0 and cuDNN 7.3
* It contain the GPU experiment
* The Generated graphs
* The logs of the experiment
* It contain file for each framework with the commands needed for installation
* It's recommended to create different environmnet for each framework using conda or virtualenvs
There exist 3 files for logging the resources during the experiment: CPU log, GPU Log, memory Log.
1- Install the environment for each framework using the installation guide
2- Clone the project
3- For running the experiment over MNIST datset for Keras framework for example, you will find in CPU folder the source code, There is a file for each framework.
4- There is a method in the main function that is called runModel, this methods holds the name of the dataset and the number of epochs needed for this run.
There are many other optional parameters with the following default values, these paramters including:
1- Learning Rate=0.01
2- momentum=0.5
3- Weight Decay=1e-6
4- batch size = 128
create a new environment with conda $ conda create -n [my-env-name] activate the environment you created $ source activate [my-env-name] install pip in the virtual environment $ conda install pip