tusharsingh62 / classifier-using-pretrained-autoencoder

Design and train a network that combines supervised and unsupervised architecture in one model to achieve a classification task

classifier-using-pretrained-autoencoder

Tested on docker container

• Build docker image from Dockerfile

docker build -t cifar .

• Create docker container based on above docker image

docker run --gpus 0 -it -v $(pwd):/mnt -p 8080:8080 cifar

• Enter docker container and follow the steps to reproduce the experiments results

docker exec -it {container_id} /bin/bash

• Current directory is mounted on /mnt path inside the docker container
• Go to /mnt directory inside the docker container

Download CIFAR-10 dataset

cd data/

sudo wget http://pjreddie.com/media/files/cifar.tgz

tar xzf cifar.tgz

rm cifar.tgz

Process CIFAR-10 dataset and prepare train, test dataset according to the cifar10_train_labels.txt file

cd ../ # Move to the root folder

python prepare_cifar10.py --images_dir data/cifar/train/ --out_dir data/processed_cifar/train/ --set_num_classes

python prepare_cifar10.py --images_dir data/cifar/test/ --out_dir data/processed_cifar/test/

Distribution of training dataset after processing the cifar-10

Training data config: {'ship': 716, 'airplane': 714, 'deer': 2500, 'bird': 2500, 'horse': 714, 'cat': 714, 'truck': 2500, 'automobile': 714, 'dog': 714, 'frog': 714}

Handling Imbalance dataset

• Data Augmentation of minority classes
• Setup class weights in the loss function

Data Augmentation and Train the autoencoder

python train_autoencoder.py

• Please check the default parameters for above autoencoder training script
• It will create another data/cifar10_aug cifar-10 data directory after augmentation
• Also it start training the autoencoder (unsupervised learning) on augmented cifar-10 dataset

Train Classifier

Experiment-1

• Optimizer--> SGD
• Xavier Initialization
• Data Augmentation
• Others are default parameters

python train_classifier.py --add_fc_layers --train_from_scratch

Experiment-2

• Optimizer--> SGD
• Autoencoder pretrained Initialization
• Data Augmentation
• Others are default parameters

python train_classifier.py --add_fc_layers

Experiment-3

• Optimizer--> SGD
• Autoencoder pretrained Initialization
• Weight balance for each classes in the loss function
• Others are default parameters

python train_classifier.py --add_fc_layers --balance_weights

Experiment-4

• Optimizer--> SGD
• Xavier Initialization
• Weight balance for each classes in the loss function
• Others are default parameters

python train_classifier.py --add_fc_layers --balance_weights --train_from_scratch

Evalution

python eval.py --model_path {} --data_dir {}

Experiment Results

AutoEncoder Reconstruction

AutoEncoder Training

All experiments results

Data Augmentation SGD with prerained auto encoder initialization