Test to show Lasagne/Theano automagically doing one-hot upsampling to reverse maxpooling

maxpool_undo.py

import numpy as np
from numpy import float32

import theano
import theano.tensor as T
import lasagne


######################################################
# Test of reversing maxpooling

# create a network that applies transformation -> maxpool -> untransformation -> unmaxpool
# (the transformations are to satisfy myself that it's not happening by accident)

input_var = T.tensor4('X')
datashape = (2,3,4,5)
network = lasagne.layers.InputLayer(datashape, input_var)
network = lasagne.layers.ExpressionLayer(network, lambda x: (x - 1) * 0.5)
network = lasagne.layers.MaxPool2DLayer(network, pool_size=3, stride=2)
maxpool_layer = network # store a pointer to this one
network = lasagne.layers.ExpressionLayer(network, lambda x: (x * 2) + 1)
network = lasagne.layers.InverseLayer(network, maxpool_layer)


# create a dataset
input = np.zeros(datashape, dtype=float32)
datashape_flat = np.product(datashape)
for _ in range(10):
	input.flat[np.random.randint(datashape_flat)] = np.random.randint(5) * 2 + 1  # odd numbers, we can be sure will be floating-point identical despite our transformations

# process the dataset, storing the latents and the output
output_fn  = theano.function([input_var], lasagne.layers.get_output(network))
latents_fn = theano.function([input_var], lasagne.layers.get_output(maxpool_layer))
output = output_fn(input)
latents = latents_fn(input)

# print the latents, the output, and whether they're identical
print("input:")
print input
print("latents (shape %s)" % (str(latents.shape)))
#print latents
print("output:")
print output
#print("error (input-output):")
#print input-output