airaria / DogBreedIdentification_RosettaStone

Using tensorflow and pytorch to classify dog breeds ( kaggle competition)

Dog Breed Identification

Fine-tune pretrained models in tensorflow and pytorch to identify dog breeds, mostly for pedagogical purpose.

Purpose of the repo

is to show how to

• load and fine-tune pretrained model in tensorflow and pytorch,
• read large dataset using the dataset paradigm in tensorflow and pytorch,
• do data augmentation on image data in tensorflow and pytorch,
• make a comparision between tensorflow and pytorch.

Prerequisites

• tensorflow >= 1.3
• pytorch >= 0.3
• numpy, pandas, sklearn, Pillow

Usage

1. Download the data from kaggle: www.kaggle.com/c/dog-breed-identification

2. Unzip labels.csv.zip and train.zip and put them under data/ folder, now the data/ folder contains labels.csv and a new folder train/ .

3. To train pytorch model, run

python pytorch_version/main.py

4. To train tensorflow model, download inceptionV3 model file from https://github.com/tensorflow/models/tree/master/research/slim , unzip, and put it under tf_version/, run:

python tf_version/main.py

Results

After 10 training epochs:

• tensorflow model (using InceptionV3): mean accuracy reaches 90% on val set.
• pytorch model (using Resnet50): mean accuracy reaches 87% on val set.

Acknowledgments

tf_version/preprocessing and tf_version/nets are borrowed from slim :https://github.com/tensorflow/models/tree/master/research/slim