Tiny-Imagenet-200

This repository is my personal research code for exploration of Convolutional Neural Networks, specifically on the Tiny-Imagenet-200 dataset. I plan to start small with subsets of 10 classes to benchmark against CIFAR-10, then eventually expand to larger and larger subsets, making my way up to all 200 classes to compare against Stanford's CS231N results.

Table of Contents

• Getting Started
• Setting up your Environment
  • Anaconda
  • Pip
• Creating Class Sets
• Training the Network
• Evaluating the Network
• Tweakable Parameters

---

Getting Started

To use this code you will first need to download the dataset from it's website: http://cs231n.stanford.edu/tiny-imagenet-200.zip

Alternatively, you can run the following command in your terminal if you have wget installed to download it to your current directory:

wget http://cs231n.stanford.edu/tiny-imagenet-200.zip

Setting Up Your Environment

Anaconda

Using Anaconda is a GREAT way to keep everything consistent regardless of what machine you run your code on.

The following command will create an Anaconda virtual environment with the modules listed in requirements.txt installed. This is very useful on clusters where you don't necessarily have root priveliges. To be specific, I used Anaconda/4.0.0 in these cases. Locally, any version should be fine.

conda create --name py3 python=3.5.2 --file requirements.txt

In the case of an error involving new versions of libraries not being backwards compatible, I saved the conda environment with all of the versions that were used to create the first 2000 models seen in utils/2000networks.csv in utils/working_conda_env.yml. To recreate the environment with these exact library verions, run this command:

conda env create -f utils/working_conda_env.yml

NOTE: For some reason, this conda environment doe not work properly on my current configuration of Arch Linux. However, when running it the same way on Ubuntu 16.04, it works perfectly fine. This needs to be revisited.

Pip

Then you will need to install all required libraries by running the command:

pip install -r requirements.txt

Creating class sets

To create sets of classes, just run python produce_files.py and enter your desired number of sets, and number of classes per set.

Currently, I have number of classes set to 200 as default and image size set to 64x64x3 as default.

To train this network, make whatever sets of classes you need using:

python utils/produce_files.py

and by following the prompts.

Training Network

Currently, everything is meant to be run from the highest level directory of this repository. Paths could be incorrect if you run code from the directory it is contained in, and this code would lose it's current generality.

To benchmark results against CIFAR-10, run the following command to resize images from 64x64x3 to 32x32x3 and train the network on a 10-class subset on the CIFAR LeNet architecture:

python networks/train_tiny_lenet.py --resize=True --num_classes=10

To choose which set of classes you train the network on when executing the command, you can use the wnids optional argument to pass the relative path to your chosen set of classes:

python networks/train_tiny_lenet.py --resize=True --num_classes=10 --wnids='random/0'

Otherwise, you will be prompted to input the path to a set of classes when simply executing and default values for most parameters will be set:

python networks/train_tiny_lenet.py

Evaluating Trained Network

To evaluate the accuracy of a network that's already been trained, you can use the --load optional argument as demonstrated below

python networks/train_tiny_lenet.py --resize=True --num_classes=10 --wnids='random/0' --load='work/training/tiny_imagenet/sets/random/0/best_weights_val_acc.hdf5'

Tweakable Parameters

# String: Choice of whether to use 'cpu', 'gpu', '2gpu', Default='cpu'
--hardware

# Int: How many images to pass through the network at once, Default=100
--batch_size

# Int: How many times to run all of the data through the network, Default=25
--num_epochs

# Int: Number of classes the network is being trained on
--num_classes

# Float: Adjustable hyperparameter, Default=0.001
--learning_rate

# Float: Adjustable hyperparameter, Default=0.00
--weight_decay

# String: "True" or "False", Whether to preprocess data in certain ways, Default="False"
--data_augmentation

# String: Choice of 'train_acc', 'train_loss', 'val_acc', 'val_loss' to monitor
# for saving model checkpoints, Default='val_acc'
--best_criterion

# String: Path to set of classes to train on, Default=User_Input
--wnids

# String: "True"=32x32, "False"=64x64, Default="False"
--resize

# String: Path to saved model to evaluate accuracy of
--load

# String: "True"=Normalize images by dividing each color channel by 255, Default="False"
--normalize

Notes To Self

TODO:

• Check if int values of images are correct as opposed to floats like in Matlab.
• Check if resized images look correct with imshow
• Run and save CIFAR-10 results to compare on CIFAR-LeNet

IDEAS:

• Train a network to learn best classes to put together in a set?
• Without using a network, recursively train a network, get the best classes and train on those, and repeat
• Can a Generative Adversarial Network be used to increase the number of training/validation images per class by generating realistic images?

ISSUES: