def train_ae(ae, patch_shape=(20, 20), n_hidden=400, n_samples=50000,
batch_size=500, epsilon=0.1, show_input=False, show_filters=False,
sparsity=0.01, beta=0.0, weight_decay=0.0, corruption_rate=0.3,
img_dir=None, norm_axis=0, n_epochs=5, n_reconstruct=10):
""" train an autoencoder """
img_data = get_input(n_samples=n_samples, patch_shape=patch_shape, epsilon=epsilon, norm_axis=norm_axis)
data = theano.shared(img_data, borrow=True)
if show_input:
plot_images(img_data, shape=patch_shape)
n_vis = patch_shape[0]*patch_shape[1]
trained = ae.train(data, n_visible=n_vis, n_hidden=n_hidden, batch_size=batch_size, sparsity=sparsity,
weight_decay=weight_decay, beta=beta, corruption_rate=corruption_rate, n_epochs=n_epochs)
if img_dir is not None:
if ae.__name__ == "dae":
opts = "_{}".format(corruption_rate)
elif ae.__name__ == "sae":
opts = "_{}_{}_{}".format(weight_decay, beta, sparsity)
fp = os.path.join(img_dir, "ff_{}_{}_{}_{}_{}_{}_{}{}.png"
.format(ae.__name__, patch_shape, n_hidden,
n_samples, batch_size, epsilon, norm_axis, opts))
else:
fp = None
render_filters(trained.w1.get_value(borrow=True), patch_shape, show=show_filters, image_file=fp)
reconstruct(trained, img_data, patch_shape, n_reconstruct)
return trained
def train_ae(ae,
patch_shape=(20, 20),
n_hidden=400,
n_samples=50000,
batch_size=500,
epsilon=0.1,
show_input=False,
show_filters=False,
sparsity=0.01,
beta=0.0,
weight_decay=0.0,
corruption_rate=0.3,
img_dir=None,
norm_axis=0,
n_epochs=5,
n_reconstruct=10):
""" train an autoencoder """
img_data = get_input(n_samples=n_samples,
patch_shape=patch_shape,
epsilon=epsilon,
norm_axis=norm_axis)
data = theano.shared(img_data, borrow=True)
if show_input:
plot_images(img_data, shape=patch_shape)
n_vis = patch_shape[0] * patch_shape[1]
trained = ae.train(data,
n_visible=n_vis,
n_hidden=n_hidden,
batch_size=batch_size,
sparsity=sparsity,
weight_decay=weight_decay,
beta=beta,
corruption_rate=corruption_rate,
n_epochs=n_epochs)
if img_dir is not None:
if ae.__name__ == "dae":
opts = "_{}".format(corruption_rate)
elif ae.__name__ == "sae":
opts = "_{}_{}_{}".format(weight_decay, beta, sparsity)
fp = os.path.join(
img_dir,
"ff_{}_{}_{}_{}_{}_{}_{}{}.png".format(ae.__name__, patch_shape,
n_hidden, n_samples,
batch_size, epsilon,
norm_axis, opts))
else:
fp = None
render_filters(trained.w1.get_value(borrow=True),
patch_shape,
show=show_filters,
image_file=fp)
reconstruct(trained, img_data, patch_shape, n_reconstruct)
return trained
def main():
"""
Train a denoising autoencoder for mnist digits using various corruption rates.
Display and save an image of the learned filters.
"""
datasets = load_data()
train_set_x = np.asarray(datasets[0][0], dtype=theano.config.floatX)
test_set_x = np.asarray(datasets[2][0], dtype=theano.config.floatX)
rates = [0.3]
n_reconstruct = 20
for rate in rates:
ae = train_dae(train_set_x, corruption_rate=rate)
fn = "img/mnist_filters_{}.png".format(rate)
render_filters(ae.w1.get_value(borrow=True), tile_shape, show=True, image_file=fn)
reconstruct(ae, test_set_x, tile_shape, n_reconstruct)
