In [ ]:

    

# imports and logger!
from opendeep.log import config_root_logger
from opendeep.models import GSN
from opendeep.optimization import SGD
from opendeep.data import MNIST
from opendeep.utils.misc import closest_to_square_factors
from PIL import Image as pil_img
from opendeep.utils.image import tile_raster_images
import numpy

config_root_logger()


    

In [ ]:

    

# A denoising autoencoder (DAE) is a special case of a generative stochastic network (GSN).
# GSN's can have multiple denoising layers that interact with each other both above and below.
dae = GSN(
    input_size=28*28,
    hidden_size=1000,
    visible_activation='sigmoid',
    hidden_activation='tanh',
    layers=1,
    walkbacks=3,
    input_noise='salt_and_pepper',
    input_noise_level=0.3
)


    

In [ ]:

    

# Create the MNIST data object
mnist = MNIST(concat_train_valid=True)

# Create the optimizer object
optimizer = SGD(dataset=mnist,
                epochs=40, 
                batch_size=100, 
                learning_rate=.25,
                lr_decay='exponential',
                lr_decay_factor=.995,
                momentum=.5,
                nesterov_momentum=False)


    

In [ ]:

    

# Train the model with the optimizer on the mnist dataset!
dae.train(optimizer)


    

In [ ]:

    

# Run some numbers to see the output
n_examples = 100
xs_test = mnist.test_inputs[:n_examples]
noisy_xs_test = dae.f_noise(xs_test)
reconstructed = dae.run(noisy_xs_test)
# Concatenate stuff
stacked = numpy.vstack(
    [numpy.vstack([xs_test[i * 10: (i + 1) * 10],
                   noisy_xs_test[i * 10: (i + 1) * 10],
                   reconstructed[i * 10: (i + 1) * 10]])
     for i in range(10)])
number_reconstruction = pil_img.fromarray(
    tile_raster_images(stacked, (dae.image_height, dae.image_width), (10, 30), (1, 1))
)

number_reconstruction.save("dae_opendeep_test.png")

# Construct image from the weight matrix
image = pil_img.fromarray(
    tile_raster_images(
        X=dae.weights_list[0].get_value(borrow=True).T,
        img_shape=(28, 28),
        tile_shape=closest_to_square_factors(dae.layer_sizes[1]),
        tile_spacing=(1, 1)
    )
)
image.save("dae_opendeep_filters.png")
print "Done!"


    

In [ ]: