# Network Architecture
patch_size = (28,28)
visible_size = patch_size[0] * patch_size[1]
hidden_size = 196
#hidden_size = 10
# Training params
weight_decay, sparsity_param, beta = 3e-3, 0.1, 3
max_iter = 400 # Maximum number of iterations of L-BFGS to run
# Get the data
num_samples = 100
num_samples = 100000
patches,_ = sample_images.get_mnist_data('../data/mnist.pkl.gz',
lambda l: l >= 5,
train=True,
num_samples=num_samples)
# set up L-BFGS args
theta = sparse_autoencoder.initialize_params(hidden_size, visible_size)
sae_cost = partial(sparse_autoencoder.cost,
visible_size=visible_size,
hidden_size=hidden_size,
weight_decay=weight_decay,
beta=beta,
sparsity_param=sparsity_param,
data=patches)
# Train!
trained, cost, d = scipy.optimize.lbfgsb.fmin_l_bfgs_b(sae_cost, theta,
maxfun=max_iter,
def softmax_predict(softmax_model, data):
"""Accepts model and data.
Model shape is (num_labels, input_size),
data shape is (input_size, num_examples).
Returns array of predicted labels from 0..num_labels .
"""
# note: do not need to use exponential or sigmoid since
#it's monotonic
return np.argmax(np.dot(softmax_model, data), axis=0)
if __name__=='__main__':
num_examples = 60000
data, labels = sample_images.get_mnist_data('../data/mnist.pkl.gz',
train=True,
num_examples=num_examples)
print len(labels)
input_size = 28 * 28 # Size of input vector (MNIST images are 28x28)
num_classes = 10 # Number of classes (MNIST images fall into 10 classes)
weight_decay = 1e-4 # Weight decay parameter
max_iter = 400
theta = 0.005 * np.random.randn(num_classes * input_size)
trained = softmax_train(input_size,
num_classes,
weight_decay,
data,
