def link(self, input):
"""
Convolve input feature maps with filters.
Input: Feature map of dimension (batch_size, stack_size, nb_rows, nb_cols)
Output: Feature map of dimension (batch_size, nb_filters, output_rows, output_cols)
"""
if self.k_max is None:
raise Exception("k_max has not been defined in the layer %s!" %
self.name)
self.input = input
# 2D convolutional layer
self.conv2d_layer.link(self.input)
self.conv_out = self.conv2d_layer.conv_out
# k-max pooling
k_max_layer = pooling.KMaxPoolingLayer1(self.k_max)
self.pooled_out = k_max_layer.link(self.conv_out)
# bias + squash function
self.linear_output = self.pooled_out + self.conv2d_layer.bias.dimshuffle(
'x', 0, 'x', 'x')
self.output = T.tanh(self.linear_output)
return self.output
def test_kmax_pooling_layer_1():
"""
Doesn't seem to work if k-max is bigger than the third dimension.
"""
print "Testing k-max Pooling Layer 1..."
input = T.tensor4('input_test')
k_max = T.iscalar('')
kmax_pooling_layer_old = pooling.KMaxPoolingLayer1(k_max)
output = kmax_pooling_layer_old.link(input)
for i in xrange(1000):
if i % 50 == 0:
print "%i" % i,
# random parameters
input_value = np.random.rand(
np.random.randint(1, 20),
np.random.randint(1, 20),
np.random.randint(10, 20),
np.random.randint(1, 100)
).astype(floatX)
k_max_value = np.random.randint(1, 10)
expected_value = kmax_pooling(input_value, k_max_value)
# print k_max_value, input_value.shape, expected_value.shape
# print output.eval({input:input_value, k_max:k_max_value}).shape
assert expected_value.shape[:2] + expected_value.shape[3:] == input_value.shape[:2] + input_value.shape[3:]
assert expected_value.shape[2] in [k_max_value, input_value.shape[2]]
np.testing.assert_array_almost_equal(output.eval({input: input_value, k_max: k_max_value}), expected_value)
print "OK"
print "All tests ran successfully for k-max Pooling Layer 1."