def forward(self, X: Variable):
"""
:param X: X is a 4d tensor, [batch, channel, row, col]
# TODO add channel in different places
:return:
"""
# print("starting convolution.")
def idx_three2one(idx, shape):
new_idx = idx[0] * np.prod(shape[1:]) + idx[1] * shape[2] + idx[2]
return new_idx
# Notice that we only need to calculate mapping once for all epochs
if self.initialize:
self.n, self.in_channel, self.x, self.y = X.shape
# We first calculate the new matrix size.
self.x_new = int((self.x - self.kernel_size[0] + 2 * self.padding[0]) / self.stride[0] + 1)
self.y_new = int((self.y - self.kernel_size[1] + 2 * self.padding[1]) / self.stride[1] + 1)
self.old_length = self.in_channel * self.x * self.y
self.new_length = self.out_channel * self.x_new * self.y_new
# The thing about mapping is that we have to calculate the mapping during each iteration,
# Because w has changed within each iteration
# On the other hand, we have to keep w changing at the same time.
# We only need to initialize b once, with the knowledge of xnew and ynew
# Initialize the kernel
self.w = Variable(np.random.normal(0, 0.01, (self.out_channel, self.in_channel,
self.kernel_size[0], self.kernel_size[1])),
trainable=self.trainable)
self.b = Variable(np.random.normal(0, 0.01, (1, self.out_channel, self.x_new, self.y_new)),
trainable=self.trainable, param_share=True)
'''
Now we create a w2mapping, the mapping itself we only need it once for all.
After we know the mapping, we can easily do the back-prop and forward-prop each time.
'''
self.w2mapping = []
# Logic 1, without sorting
for filter_idx in range(self.out_channel):
for i in range(self.x_new):
for j in range(self.y_new):
# Index for new matrix
mapping_new = idx_three2one((filter_idx, i, j),
(self.out_channel, self.x_new, self.y_new))
x_start = int(i * self.stride[0])
y_start = int(j * self.stride[1])
for ix in range(self.kernel_size[0]):
for jx in range(self.kernel_size[1]):
for channel_idx in range(self.in_channel):
# Index for old matrix
mapping_old = idx_three2one((channel_idx, x_start + ix, y_start + jx),
(self.in_channel, self.x, self.y))
# We have to record, which one in the mapping matrix is from which w
self.w2mapping.append([(filter_idx, channel_idx, ix, jx),
(mapping_old, mapping_new)])
self.initialize = False
# End Initialize
input_image_flattened = X.reshape((self.n, self.old_length))
new_image_flattened = input_image_flattened.sparse_dot_with_mapping(self.w, self.w2mapping,
self.old_length,
self.new_length)
output = new_image_flattened.reshape((self.n, self.out_channel,
self.x_new, self.y_new))
# Add bias if necessary
if self.bias:
output1 = output + self.b
return output1
return output
def forward(self, X: Variable):
self.n, self.c, self.x, self.y = X.shape
output = X.reshape((self.n, self.c * self.x * self.y))
return output