def test_convolutional_feature_map_layer_grad():
layer = MultiConvolutionalFeatureMapLayer((5, 5), (10, 10), 4, [[0, 1], [0, 2], [1, 2], [0, 1, 2]])
layer.initialize()
raw_input = random.normal(size=10 * 10 * 4)
inputs = [raw_input[(100 * index) : (100 * (index + 1))].reshape((10, 10)) for index in xrange(4)]
def func(inp):
layer.params[...] = inp
return np.sum(out.sum() for out in layer.fprop(inputs))
params = random.normal(size=len(layer.params))
layer.params[...] = params
layer.grad([np.ones((6, 6)) for index in xrange(4)], inputs)
approx = fd_grad(func, params)
assert_array_almost_equal(layer._grad, approx)
def test_convolutional_feature_map_layer_grad():
layer = MultiConvolutionalFeatureMapLayer(
(5, 5), (10, 10), 4, [[0, 1], [0, 2], [1, 2], [0, 1, 2]])
layer.initialize()
raw_input = random.normal(size=10 * 10 * 4)
inputs = [
raw_input[(100 * index):(100 * (index + 1))].reshape((10, 10))
for index in xrange(4)
]
def func(inp):
layer.params[...] = inp
return np.sum(out.sum() for out in layer.fprop(inputs))
params = random.normal(size=len(layer.params))
layer.params[...] = params
layer.grad([np.ones((6, 6)) for index in xrange(4)], inputs)
approx = fd_grad(func, params)
assert_array_almost_equal(layer._grad, approx)
