Binary_Classifier

A simple Cat-Dog Classifier built on ResNet-50 using Keras 2.1.5 with Tensorflow (1.4) backend on the Kaggle Dataset (link).

Prerequisites

1. Make sure the dataset from this link is extracted inside the project folder.

2. Python 3+

3. Install Tensorflow:

   $ pip install tensorflow
   

4. Install Keras (2.1.5+):

   $ pip install keras
   

5. Install h5py:

   $ pip install h5py
   

6. Install PIL (Pillow):

   $ pip install Pillow
   

Detailed guides to install these libraries can be found here for MacOS and Ubuntu.

Testing

To test a single image for prediction open terminal and run ( Ubuntu/macOS recomended ):

$ python predict.py

By default the above command will classify the first image in test_images directory.

To pass custom images for testing:

$ python predict.py --image <image name>

Fill the with the custom image name. Even custom model can be passed as flag for testing:

$ python predict.py --image <image name> --model <model name>

All these flags are given default value so it will execute the script directly without specifying them explicitly.

Training

The dataset in the mentioned link will download the folder kagglecatsanddogs_3367a with a subfolder named PetImages.

Copy the subfolders of PetImages into Data\Cat and Data\Dog accordingly.

The obtained dataset has too many corrupt images. We first need to delete them. This is done by simply running the script ( works only on Ubuntu/macOS ):

$ python clean.py

We can then start the training:

$ python train.py

Parameters like resolution, batch_size etc can be changed in the script.

Model Compression (Optional)

In order to compress the model with the tradeoff of accuracy, the following steps can be done (High RAM, Good CPU (8 core) and Good GPU recommened):

1.$ git clone https://github.com/DwangoMediaVillage/keras_compressor.git 2. $ cd ./keras_compressor 3. pip install .

This tool is used to compress the model to reduce disk usage.

Once training is done the script would generate a file named resnet50_best.h5, we can compress this file using terminal command:

To control the accuracy reduction flag can be passed like this:

$ keras-compressor.py --error 0.001 resnet50_best.h5 compressed.h5

Which would generate a new compressed file name compressed.h5.

The compression takes a very long time since it looks at every layer one by one and compresses it.

Tensorboard

To visualise the network graph and the training loss, run in terminal:

$ tensorboard --logdir=<full path to log dir>

replace <full path to log dir> with your full path.

This will generate an IP Address, navigate to this IP Address using a web browser.

Results

These results were obtained after deleting the corrupt images using a batch size of 16 and split ratio of 0.3 (training and testing split). It was trained for 12 epochs but the loss stopped improving at the 4th epoch.

Category	%
Training Loss	8.5%
Training Accuracy	96.6%
Validation Loss	9.3%
Validation Accuracy	96.55%
Test Weights Accuracy	95.97%