cifar-10

Requirements

These softwares must be installed on your local machine:

Docker (with nvidia-docker2)

Conda virtual environment

Installation

Clone this repository and change directory to it
Copy env.example as .env and edit environment variables
Build custom image: docker build -t tensorflow/tensorflow:custom-gpu-py3 .

Usage

To start tensorflow container:

nvidia-docker run -it -u $(id -u):$(id -g) -v "$(realpath ./)":/tf/ -p 6006:6006 --name tf tensorflow/tensorflow:custom-gpu-py3 bash

From the container bash you can execute the train script:

cd tf
python train.py
# All available hyperparameters will be prompted.

Available models are linear, mlp, cnn, resnet, lstm

Example

python train.py --model mlp --hidden-layers 2 6 8 --units 10 64 512 --activation relu sigmoid tanh --output-activation sigmoid softmax --lr 0.1 0.01 --epochs 10 50 100 --momentum 0.9 0.7 --batch-size 32 1024

It will train all model combinaison using Talos grid search. It will generate a csv file at the end.

Optional

You can execute tensorboard:

docker exec -d tf tensorboard --logdir=./logs

Browse http://localhost:6006

walkoss / cifar-10