def post_epoch(*args):
save_params(*args)
# Save visualization weights.
W_norm = utils.rescale(args[0].W)
img = Image.fromarray(np.uint8(utils.tile(W_norm, channel_count=3) * 255))
img.save(os.path.join(output_dir, ('w%i.png' % args[1])))
# Estimate sparsity from subset of data.
h_mean = grbm.sample_h_noisy_relu(args[0], inputs[0:5000], True)[1]
mean_activation = np.mean(h_mean > 0)
print 'approx mean activation: %f' % mean_activation
def post_epoch(*args):
save_params(*args)
# Save visualization weights.
W_norm = utils.rescale(args[0].W)
img = Image.fromarray(
np.uint8(utils.tile(W_norm, channel_count=3) * 255))
img.save(os.path.join(output_dir, ('w%i.png' % args[1])))
# Estimate sparsity from subset of data.
h_mean = grbm.sample_h_noisy_relu(args[0], inputs[0:5000], True)[1]
mean_activation = np.mean(h_mean > 0)
print 'approx mean activation: %f' % mean_activation
def post_epoch(*args):
W_norm = utils.rescale(args[0].W)
utils.save_image(utils.tile(W_norm),
os.path.join(output_dir, ('w%i.png' % args[1])))
# Estimate sparsity from subset of data.
h_mean = grbm.sample_h_noisy_relu(args[0], inputs[0:5000], True)[1]
mean_activation = np.mean(h_mean > 0)
print 'approx mean activation: %f' % mean_activation
# The callback from optimize.sgd needs modifying so that it
# passes the reconstrcution error as an argument to make this
# work. (This was used when I did the original experiments.)
# error_history.append(args[2])
sparsity_history.append(mean_activation)
save_params(args[0], args[1])