datasets.py

import os
import numpy as np
from skimage import io

from nolearn.lasagne import NeuralNet
from lasagne.layers import InputLayer
from lasagne.layers import Conv2DLayer
from lasagne.layers import MaxPool2DLayer
from lasagne.layers import DropoutLayer
from lasagne.layers import DenseLayer
from lasagne.updates import nesterov_momentum
from lasagne.nonlinearities import softmax

def load_data():

	labels = {}
	f = open('trainLabels.csv','r')
	lines = f.readline()
	lines = f.readlines()

	for l in lines:
		l = l.strip()
		l = l.split(',')
		labels[l[0]] = int(l[1])

	img_list = os.listdir(r'../data/train')

	n = len(img_list)
	x = np.zeros((n,1,256,256))
	y = np.zeros((n,))
	
	i = 0
	for img in img_list:
		if img.endswith('.jpeg'):
			ii = io.imread(os.path.join(r'../data/train',img))
			x[i,:,:] = ii[:,:,1]
			y[i] = labels[img.replace('.jpeg','')]
			i += 1
			print i

	x /= 255.0
	y = y.astype('int32')
	return x,y

def get_test_data():

	img_list = os.listdir(r'../data/test')
	n = len(img_list)
	x = np.zeros((n,1,256,256))
	
	i = 0
	lst = []
	for img in img_list:
		if img.endswith('.jpeg'):
			ii = io.imread(os.path.join(r'../data/train',img))
			x[i,:,:] = ii[:,:,1]
			lst.append(img.replace('.jpeg',''))
			i += 1
			print i
	
	return x,lst
	
def build_model():

	l = [('input', InputLayer), 
		('conv1', Conv2DLayer), ('pool1', MaxPool2DLayer), ('drop1',DropoutLayer), 
		('conv2', Conv2DLayer), ('pool2', MaxPool2DLayer), ('drop2',DropoutLayer),
		('conv3', Conv2DLayer), ('pool3', MaxPool2DLayer),('drop3',DropoutLayer),
		('output',DenseLayer)]
		
	net = NeuralNet(layers=l, input_shape=(None, 1, 256, 256), 
			conv1_num_filters=10, conv1_filter_size=(3,3),
			pool1_ds = (3,3), drop1_p=0.5,
			conv2_num_filters=10, conv2_filter_size=(3,3),
			pool2_ds = (3,3), drop2_p=0.5,
			conv3_num_filters=10, conv3_filter_size=(3,3),
			pool3_ds = (3,3), drop3_p=0.5,
			
			output_num_units = 5,
			output_nonlinearity = softmax,
			
			update=nesterov_momentum,
			update_learning_rate=0.01,
			update_momentum=0.9,
			eval_size=0.2, verbose=1, max_epochs=200)

	return net

if __name__ == '__main__':

	x,y = load_data()
	
	net = build_model()
	net = net.fit(x,y)
	
	xtest,lst = get_test_data()
	yhat = net.predict(xtest)
	yhat = yhat.astype('int32')
	
	f = open('submission.csv','w')
	
	f.write('image,level\n')
	i = 0
	for l in lst:
		f.write(l+','+str(yhat[i])+'\n')
		i += 1
		
	f.close()