def run_net_3_conv_fully_connected_softmax():
""" Runs net 3 with with layers = [convolutional, fully connected, softmax] """
# 20 * 12 * 12 is the size of the pool layer, going into the fully connected layer
fcn_n_inp, fcn_n_out = 20 * 12 * 12, 100
sl_n_inp, sl_n_out = 100, 10
epochs, mini_batch_size, eta = 60, 10, 0.1
training_data, validation_data, test_data = cnn.load_data_shared()
filter = (20, 1, 5, 5)
image = (mini_batch_size, 1, 28, 28)
# Conv layer default to pool of 2 x 2
# Note: the convolution layer will perform the convolution and the pool layer operations using theano
layers = [
cnn.ConvPoolLayer(filter, image),
cnn.FullyConnectedLayer(fcn_n_inp, fcn_n_out), # fcn_n_out = sl_n_inp
cnn.SoftmaxLayer(sl_n_inp, sl_n_out)
]
net = network_test_3.NetworkTest3(layers, mini_batch_size)
net.stochastic_gradient_descent(training_data,
epochs,
mini_batch_size,
eta,
validation_data,
test_data,
print_mini_batch_iteration=False)
def run_net_3_conv_conv_fully_connected_softmax():
""" Runs net 3 with layers = [convolutional, convolutional, fully connected, softmax] """
training_data, validation_data, test_data = cnn.load_data_shared()
# there are 40 feature maps from the last pooling layer of 4 x 4 images
fcn_n_inp, fcn_n_out = 40 * 4 * 4, 100
sl_n_inp, sl_n_out = 100, 10
epochs, mini_batch_size, eta = 60, 10, 0.1
# original input image is a 1 dimensional 28 x 28 pixels
conv_1_image = (mini_batch_size, 1, 28, 28)
# this layer will have 20 feature maps, 1 input feature, and will use a 5 x 5 kernel
conv_1_filter = (20, 1, 5, 5)
# 20 12 x 12 feature maps coming from the previous pool layer
conv_2_image = (mini_batch_size, 20, 12, 12)
# 40 feature maps as input, 20 input features, and a 5 x 5 kernel
conv_2_filter = (40, 20, 5, 5)
layers = [
cnn.ConvPoolLayer(conv_1_filter, conv_1_image),
cnn.ConvPoolLayer(conv_2_filter, conv_2_image),
cnn.FullyConnectedLayer(fcn_n_inp, fcn_n_out),
cnn.SoftmaxLayer(sl_n_inp, sl_n_out),
]
net = network_test_3.NetworkTest3(layers, mini_batch_size)
net.stochastic_gradient_descent(training_data,
epochs,
mini_batch_size,
eta,
validation_data,
test_data,
print_mini_batch_iteration=False)
def run_net_3_with_dropout():
""" Runs net 3 with layers = [conv, conv, full, full, softmax] with ReLU activation and dropout """
functs = functions.Functions()
training_data, validation_data, test_data = cnn.load_data_shared()
full_conn_1_input, full_conn_1_output = 40 * 4 * 4, 100
full_conn_2_input, full_conn_2_output = 100, 100
# drop half of the neurons
p_dropout = 0.5
sl_n_inp, sl_n_out = 100, 10
epochs, mini_batch_size, eta, lmbda = 30, 10, 0.03, 0.1
conv_1_image = (mini_batch_size, 1, 28, 28)
conv_1_filter = (20, 1, 5, 5)
conv_2_image = (mini_batch_size, 20, 12, 12)
conv_2_filter = (40, 20, 5, 5)
layers = [
cnn.ConvPoolLayer(conv_1_filter,
conv_1_image,
activation_fn=functs.relu),
cnn.ConvPoolLayer(conv_2_filter,
conv_2_image,
activation_fn=functs.relu),
cnn.FullyConnectedLayer(full_conn_1_input,
full_conn_1_output,
activation_fn=functs.relu,
p_dropout=p_dropout),
cnn.FullyConnectedLayer(full_conn_2_input,
full_conn_2_output,
activation_fn=functs.relu,
p_dropout=p_dropout),
cnn.SoftmaxLayer(sl_n_inp, sl_n_out),
]
net = network_test_3.NetworkTest3(layers, mini_batch_size)
net.stochastic_gradient_descent(training_data,
epochs,
mini_batch_size,
eta,
validation_data,
test_data,
lmbda=lmbda,
print_mini_batch_iteration=False)
def run_net_3_fully_connected_softmax():
""" Runs net 3 with layers = [fully connected, softmax]"""
fcn_n_inp, fcn_n_out = 784, 100
sl_n_inp, sl_n_out = 100, 10
epochs, mini_batch_size, eta = 60, 10, 0.1
training_data, validation_data, test_data = cnn.load_data_shared()
layers = [
cnn.FullyConnectedLayer(fcn_n_inp, fcn_n_out),
cnn.SoftmaxLayer(sl_n_inp, sl_n_out)
]
net = network_test_3.NetworkTest3(layers, mini_batch_size)
net.stochastic_gradient_descent(training_data,
epochs,
mini_batch_size,
eta,
validation_data,
test_data,
print_mini_batch_iteration=False)
def run_net_3_conv_conv_fully_connected_softmax_with_relu():
""" Runs net 3 with layers = [convolutional, convolutional, fully connected, softmax] and ReLU activation """
functs = functions.Functions()
training_data, validation_data, test_data = cnn.load_data_shared()
fcn_n_inp, fcn_n_out = 40 * 4 * 4, 100
sl_n_inp, sl_n_out = 100, 10
epochs, mini_batch_size, eta, lmbda = 60, 10, 0.03, 0.1
conv_1_image = (mini_batch_size, 1, 28, 28)
conv_1_filter = (20, 1, 5, 5)
conv_2_image = (mini_batch_size, 20, 12, 12)
conv_2_filter = (40, 20, 5, 5)
layers = [
cnn.ConvPoolLayer(conv_1_filter,
conv_1_image,
activation_fn=functs.relu),
cnn.ConvPoolLayer(conv_2_filter,
conv_2_image,
activation_fn=functs.relu),
cnn.FullyConnectedLayer(fcn_n_inp,
fcn_n_out,
activation_fn=functs.relu),
cnn.SoftmaxLayer(sl_n_inp, sl_n_out),
]
net = network_test_3.NetworkTest3(layers, mini_batch_size)
net.stochastic_gradient_descent(training_data,
epochs,
mini_batch_size,
eta,
validation_data,
test_data,
lmbda=lmbda,
print_mini_batch_iteration=False)