def svhn_encoder(x, numHidden, labels, num_labels, mb_size, image_width):
in_width = image_width
layerLst = []
c = [3, 64, 128, 256, 256]
layerLst += [ConvPoolLayer(in_channels = c[0], out_channels = c[1], kernel_len = 5, batch_norm = False)]
layerLst += [ConvPoolLayer(in_channels = c[1], out_channels = c[1], kernel_len = 5, batch_norm = False)]
layerLst += [ConvPoolLayer(in_channels = c[1], out_channels = c[1], kernel_len = 5, stride=2, batch_norm = False)]
layerLst += [ConvPoolLayer(in_channels = c[1], out_channels = c[2], kernel_len = 5, batch_norm = False)]
layerLst += [ConvPoolLayer(in_channels = c[2], out_channels = c[2], kernel_len = 5, batch_norm = False)]
layerLst += [ConvPoolLayer(in_channels = c[2], out_channels = c[2], kernel_len = 5, stride=2, batch_norm = False)]
layerLst += [ConvPoolLayer(in_channels = c[2], out_channels = c[3], kernel_len = 5, batch_norm = False)]
layerLst += [ConvPoolLayer(in_channels = c[3], out_channels = c[3], kernel_len = 5, batch_norm = False)]
layerLst += [ConvPoolLayer(in_channels = c[3], out_channels = c[4], kernel_len = 5, stride=2, batch_norm = False)]
layerLst += [HiddenLayer(num_in = 4 * 4 * c[4], num_out = numHidden, flatten_input = True, batch_norm = False)]
layerLst += [HiddenLayer(num_in = numHidden, num_out = numHidden, batch_norm = True)]
outputs = [normalize(x.transpose(0,3,1,2))]
for i in range(0, len(layerLst)):
outputs += [layerLst[i].output(outputs[-1])]
h1 = HiddenLayer(num_in = numHidden + num_labels, num_out = numHidden, batch_norm = True)
h2 = HiddenLayer(num_in = numHidden, num_out = numHidden, batch_norm = True)
h1_out = h1.output(T.concatenate([outputs[-1], labels], axis = 1))
h2_out = h2.output(h1_out)
return {'layers' : layerLst + [h1,h2], 'extra_params' : [], 'output' : h2_out}
params += [layerParams[paramKey]]
l2_loss += T.mean(T.sqr(params[-1]))
l2_loss = 0.0001 * T.sqrt(l2_loss)
params += encoder_extra_params + decoder_extra_params
for param in params:
print param.get_value().shape
variational_loss = config['vae_weight'] * 0.5 * T.sum(z_mean**2 + z_var - T.log(z_var) - 1.0)
#smoothness_penalty = 0.001 * (total_denoising_variation_penalty(x_reconstructed.transpose(0,3,1,2)[:,0:1,:,:]) + total_denoising_variation_penalty(x_reconstructed.transpose(0,3,1,2)[:,1:2,:,:]) + total_denoising_variation_penalty(x_reconstructed.transpose(0,3,1,2)[:,2:3,:,:]))
square_loss = config['square_loss_weight'] * 1.0 * T.sum(T.sqr(normalize(x) - normalize(x_reconstructed)))
loss = 0.0
loss += l2_loss
loss += square_loss
netDist = NetDist(x, x_reconstructed, config)
if config['style_weight'] > 0.0:
style_loss, style_out_1, style_out_2 = netDist.get_dist_style()
style_loss *= config['style_weight']
else:
style_loss = style_out_1 = style_out_2 = theano.shared(np.asarray(0.0).astype('float32'))
params += [layerParams[paramKey]]
l2_loss += T.mean(T.sqr(params[-1]))
l2_loss = 0.0001 * T.sqrt(l2_loss)
params += encoder_extra_params + decoder_extra_params
for param in params:
print param.get_value().shape
variational_loss = config['vae_weight'] * 0.5 * T.sum(z_mean**2 + z_var - T.log(z_var) - 1.0)
smoothness_penalty = 0.0 * (total_denoising_variation_penalty(x_reconstructed.transpose(0,3,1,2)[:,0:1,:,:]) + total_denoising_variation_penalty(x_reconstructed.transpose(0,3,1,2)[:,1:2,:,:]) + total_denoising_variation_penalty(x_reconstructed.transpose(0,3,1,2)[:,2:3,:,:]))
raw_square_loss = T.sum(T.sqr(normalize(x) - normalize(x_reconstructed)))
square_loss = raw_square_loss * config['square_loss_weight']
loss = 0.0
loss += l2_loss
loss += square_loss
netDist = NetDist(x, x_reconstructed, config)
if config['style_weight'] > 0.0:
style_loss, style_out_1, style_out_2 = netDist.get_dist_style()
style_loss *= config['style_weight']
else:
