A convolutional autoencoder...for those times when you need pre-training.

You need the following dependencies to run the network:

• [Numpy] (http://www.numpy.org)
• [Skimage] (http://scikit-image.org)
• [Theano (Bleeding-edge)] (http://deeplearning.net/software/theano/install.html#install-bleeding-edge)

To initialize the network, pass a list containing each of the encoding layers (Convolutional and Pooling) arranged hierarchically - topmost layer at the beginning of list.

layers = [
			PoolLayer((2, 2), 'max'),
			ConvLayer(16, 1, (3, 3))
		]

The network will automatically create the corresponding decoding layers.

Pre-training is done by calling

train(data, test, params)

Arguments:

• data: A no_imgs x img_length x img_width x no_channels array representing images.
• test: A no_imgs x img_length x img_width x no_channels array representing images.
• params: A dictionary of training parameters.

Here is a quick description of the training parameters:

• epochs: Integer repr. number of training epochs.
• batch_size: Integer repr. size of each training batch.
• view_kernels: Boolean indicating if kernels should be displayed.
• view_recon: Boolean indicationg if reconstructed images should be displayed.
• no_images: Integer repr. number of reconstructed images to display.
• eps_w: Integer repr. learning rate for weights/kernels.
• eps_b: Integer repr. learning rate for bias.
• eps_decay: Float repr. Decay constant for learning rate.
• eps_intvl: Integer repr. no. intervals to decay learning rate.
• eps_satr: Integer repr. no. iterations after which the learning rate stops decaying. If 'inf', learning rate keeps decaying throughout training.
• mu: Float repr. Momentum constant.
• l2: Float repr. l2 weight decay constant.
• RMSProp: Boolean option to use RMS prop.
• RMSProp_decay: Float repr. decay constant for RMS prop.
• minsq_RMSProp: Floar repr. constant for RMS prop denominator.

An example of parameters is as follows

{
	'epochs': 50,
	'batch_size': 500,
	'view_kernels': False,
	'view_recon': True,
	'no_images': 12,
	'eps_w': 0.005,
	'eps_b': 0.005,
	'eps_decay': 9,
	'eps_intvl': 10,
	'eps_satr': 'inf',
	'mu': 0.7,
	'l2': 0.95,
	'RMSProp': True,
	'RMSProp_decay': 0.9,
	'minsq_RMSProp': 0.01,
}

Below shows both the reconstructed and actual images gotten on the Toronto Faces Dataset using the above parameters

Currently to perform greedy-layer wise training, you'll have to manually train each of the layers separately - using the output of the trained layers below as input.

You can save and load a trained model by calling saveModel(filename) and loadModel(filename) respectively.

1. Support greedy-layer wise pre-training.
2. Investigate better error metric.
3. Implement the addition of noise during training.