def unwind(self, v, ktarget):
if is_eq(self, ktarget):
ktarget.resume(v)
else:
self.k.unwind(v, ktarget)
def block_lookup(self, n, k, v):
if is_eq(n, self.name):
return k.unwind(v, self.cont)
else:
return self.others.block_lookup(n, k, v)
# forward pass with transposed kernel
n_filters = cd.convweights.shape[0]
inp_reshaped = inp.detach().numpy().transpose(1, 2, 3,
0).reshape(n_filters, -1)
W_reshape = dilated.reshape(n_filters, -1)
out_col = W_reshape.T @ inp_reshaped
# image is recovered through col2im
out_image = col2im_indices(out_col,
cd.inp.shape,
dilated_size,
dilated_size,
padding=padding)
is_eq(output, out_image)
################################################################################################
### Backward Pass
dout_col = im2col_indices(cd.convlossTrans.detach().numpy(),
dilated_size,
dilated_size,
padding=padding)
#dout_reshaped = cd.convlossTrans.detach().numpy().transpose(1, 2, 3, 0).reshape(n_filter, -1)
dX_col = W_reshape @ dout_col
c, _, h, w = inp.shape
dX = dX_col.reshape(n_filters, h, w, c).transpose(3, 0, 1, 2)
is_eq(dX, inp.grad)
dW = dout_col @ inp_reshaped.T
dW = dW.reshape(dilated.shape)
