LINCOLN DataScience Challenge #2

CIFAR-10 - Image Recongnition

Content of the directory

• convNet folder: it's a python package containing all the functions and class relative to the custom implementation of the CNN.
• MNIST and CIFAR-10 folders: contains the training datasets.
• ConvNet-MNIST.ipynb: jupyter notebook that train a CNN using the custom implementation on the MNIST dataset.
• ResNet-CIFAR-10.ipynb: jupyter notebook that train a Residual Network using Keras on the CIFAR-10 dataset.

Module requirements

For the custom implementation and the residual network

• numpy
• sklearn
• pickle
• matplotlib
• time
• scipy

For the residual network

• keras
• tensorflow

Importing datasets

MNIST

Place the 'train.csv' corresponding to the MNIST dataset in the 'MNIST' folder.

CIFAR-10

Place the 'train' folder and the 'trainLabels.csv' containing the training image and labels in the 'CIFAR-10' folder.

How to use the custom implementation of CNN

To train a neural network using the custom implementation, you first need to load the conv_net class present in the convNet.conv_net module.

from convNet.conv_net import conv_net

import convNet.conv_net

Then you must initiate your model

model = conv_net()

Then set-up your CNN architecture using the different implemented layers

# Convolutionnal layer
model.add_conv(padding, # number of pixel to add around the image (zero-padding),
                        # not sure it actually works... keep it to 0
               stride, 
               filter_size,
               output_size)

# Relu layer
model.add_relu()

# Batchnorm layer
model.add_batchnorm()

# Maxpool layer
model.add_maxpool(filter_size)

# Dropout layer
model.add_dropout(p)

# Fully connected layer
model.add_fully(n)

Finnaly compile your model and train it!

# X must be shapped like (N,WIDTH,HEIGHT,CHANNEL)
# Y must be a numeric list of class (shape: (N))

model.compile()
model.train(train=(X,Y),
            test=(X_test,Y_test),
            b_size=100, # batch size
            l_rate=0.001, # learning rate
            n_epoch=20 # Epoch
            )

You can save and reload your model using this functions (base on pickle module)

import convNet.imfunc as imf

# to make a backup
imf.save_object(model, 'model_MNIST.pkl')

# to reload a model
model = imf.load_object('model_MNIST.pkl')

A more complete exemple is given in ConvNet-MNIST.ipynb.