model = NeuralNetwork(
n_inputs=32 * 32 * 3,
input_shape=(64, 3, 32, 32),
layers=[
(ConvolutionLayer, {
'image_shape': (3, 32, 32),
'filter_shape': (3, 5, 5),
'n_kernels': 20,
'non_linearity': activations.rectify
}),
(
ConvolutionLayer,
{
'filter_shape': (
20, 3, 3
), # Change this to only (3, 3), the first arg is the number of kernels in the previous layer
'n_kernels': 40,
'non_linearity': activations.rectify
}),
(FlattenLayer, {}),
(HiddenLayer, {
'n_units': 1024,
'non_linearity': activations.rectify
}),
(DropoutLayer, {
'probability': 0.5
}),
(HiddenLayer, {
'n_units': 1024,
'non_linearity': activations.rectify
}),
(DropoutLayer, {
'probability': 0.5
}),
(HiddenLayer, {
'n_units': 10,
'non_linearity': activations.softmax
})
],
trainer=(
SGDTrainer,
{
'batch_size': 64,
'learning_rate': 0.1,
'n_epochs': 400,
#'global_L2_regularization': 0.0001,
'dynamic_learning_rate': (ExponentialDecay, {
'decay': 0.99
}),
}))
model = NeuralNetwork(
n_inputs=28*28,
layers = [
(HiddenLayer,
{
'n_units': 150,
'non_linearity': activations.rectify
}),
(HiddenLayer,
{
'n_units': 80,
'non_linearity': activations.rectify
}),
(HiddenLayer,
{
'n_units': 10,
'non_linearity': activations.softmax
})
],
trainer=(SGDTrainer,
{
'batch_size': 20,
'learning_rate': 0.1,
'n_epochs': 400,
'global_L2_regularization': 0.0001,
'dynamic_learning_rate': (ExponentialDecay, {'decay': 0.99}),
}
)
)
return rval
# Load MNIST
datasets = load_data("mnist.pkl.gz")
model = NeuralNetwork(
n_inputs=28*28,
batch_size=20,
layers = [
(ConvolutionLayer,
{
'image_shape': (1, 28, 28),
'filter_shape': (1, 5, 5),
'n_kernels': 40,
'non_linearity': activations.rectify
}),
(FlattenLayer, {}),
(HiddenLayer,
{
'n_units': 80,
'non_linearity': activations.rectify
}),
(HiddenLayer,
{
'n_units': 10,
'non_linearity': activations.softmax
})
]
)
model.train(datasets[0], datasets[1])
model = NeuralNetwork(
n_inputs=28*28,
layers = [
(DropoutLayer, {'probability': 0.2}),
(HiddenLayer,
{
'n_units': 800,
'non_linearity': activations.rectify
}),
(DropoutLayer, {'probability': 0.5}),
(HiddenLayer,
{
'n_units': 800,
'non_linearity': activations.rectify
}),
(DropoutLayer, {'probability': 0.5}),
(HiddenLayer,
{
'n_units': 10,
'non_linearity': activations.softmax
})
],
trainer=(SGDTrainer,
{
'batch_size': 100,
'learning_rate': 0.1,
'n_epochs': 400,
#'global_L2_regularization': 0.0001,
'dynamic_learning_rate': (ExponentialDecay, {'decay': 0.99}),
}
)
)
import sys
sys.path.append("../")
sys.path.append("../neuralmind")
import gzip
import cPickle
import numpy as np
import theano
import theano.tensor as T
from neuralmind import NeuralNetwork
from layers import HiddenLayer
import activations
import datasets
# Load MNIST
datasets = datasets.load_mnist("mnist.pkl.gz")
model = NeuralNetwork(n_inputs=28 * 28,
layers=[(HiddenLayer, {
'n_units': 10,
'non_linearity': activations.softmax
})])
model.train(datasets[0], datasets[1])