Kaggle Cats vs Dogs Redux

Hi, this is my solution for the dogs vs cats redux kaggle competition that achieve the 51st place on the oficial public leaderboard. The solution is quite simple because I made it three days until the end of the competition and cannot improve it.

You can find the 3rd position solution on the kaggle blog

Architecture

I decide to start with a pre-trained model for this competition and I fallback to the Inception V3 model. The model achieve 99.8% accuracy. The final submission score can be improved in several ways, like emsembling more models, xgboost to combine classifiers, preprocess the training images(flip, rotate, scale), use external data, among others.

Installation

git clone https://github.com/mauri870/kaggle-cats-vs-dogs-redux.git
cd kaggle-cats-vs-dogs-redux

Preprocessing

First you need to download the train and test data from kaggle. The test and train images must be inside a test and train folder respectively

Run the preprocess script to prepare the data. It'll fit each image in a 299x299 box and fill the blank space in black color.

go run preprocess.go utils.go

Since the inception model expects the train images to be organized into subfolders, let's do that:

mkdir -p images/{dogs,cats}
cp -v images/train/cat* images/cats
cp -v images/train/dog* images/dogs

Retrain Inception V3

Here's my instructions to build and retrain the inception model:

Let's download and configure tensorflow:

export TF_VERSION=v1.0.0
wget -qO- https://github.com/tensorflow/tensorflow/archive/${TF_VERSION}.tar.gz | tar zx
cd tensorflow-${TF_VERSION}
./configure

Now we will retrain the last fully connected layers of the inception model:

python tensorflow/examples/image_retraining/retrain.py --flip_left_right --image_dir=$OLDPWD/images

Next we need to optimize our model because some ops used to train the original model are now deprecated and in case of the Golang tensorflow bindings will result in a fatal error

bazel build tensorflow/python/tools/optimize_for_inference
bazel-bin/tensorflow/python/tools/optimize_for_inference --input=/tmp/output_graph.pb --output=/tmp/output_graph_optimized.pb  --frozen_graph=True --input_names=Mul --output_names=final_result

Submission

Now we are ready to create the submission file!

Note: Refer to the official page in order to install and configure tensorflow for go

go run submission.go utils.go