def backward(self, indices, grad_outputs):
xvar, Wvar = self.get_retained_inputs()
x = xvar.data
W = Wvar.data
gyvar, = grad_outputs
gy = gyvar.data
xp = backend.get_array_module(x, W)
stride_row, stride_col = self.sy, self.sx
output_row, output_col = W.shape[1], W.shape[2]
ret = []
if 0 in indices:
gx = xp.zeros_like(x)
for i in moves.range(output_row):
for j in moves.range(output_col):
slice_row = slice(i * stride_row,
i * stride_row + W.shape[4])
slice_col = slice(j * stride_col,
j * stride_col + W.shape[5])
# ochans * ichans * krows * kcols
W_slice = W[:, i, j, ...]
# nsamps * ochans
gy_slice = gy[..., i, j]
# -> nsamps * ichans * krows * kcols
gx[:, :, slice_row,
slice_col] += xp.tensordot(gy_slice,
W_slice,
axes=[(1, ), (0, )])
ret.append(
chainer.functions.cast(variable.as_variable(gx), x.dtype))
if 1 in indices:
gW = xp.empty_like(W)
for i in moves.range(output_row):
for j in moves.range(output_col):
slice_row = slice(i * stride_row,
i * stride_row + W.shape[4])
slice_col = slice(j * stride_col,
j * stride_col + W.shape[5])
# nsamps * inchans * krows * kcols
x_slice = x[:, :, slice_row, slice_col]
# nsamps * outchans * 1 * 1
gy_slice = gy[:, :, i, j]
gW[:, i, j, :, :, :] = xp.tensordot(gy_slice,
x_slice,
axes=[(0, ), (0, )])
ret.append(
chainer.functions.cast(variable.as_variable(gW), W.dtype))
if 2 in indices:
gb = chainer.functions.sum(gyvar, axis=0)
ret.append(gb)
return ret
def backward(self, indices, grad_outputs):
xvar, Wvar = self.get_retained_inputs()
x = xvar.data
W = Wvar.data
gyvar, = grad_outputs
gy = gyvar.data
xp = backend.get_array_module(x, W)
stride_row, stride_col = self.sy, self.sx
output_row, output_col = W.shape[1], W.shape[2]
ret = []
if 0 in indices:
gx = xp.zeros_like(x)
for i in moves.range(output_row):
for j in moves.range(output_col):
slice_row = slice(i * stride_row,
i * stride_row + W.shape[4])
slice_col = slice(j * stride_col,
j * stride_col + W.shape[5])
# ochans * ichans * krows * kcols
W_slice = W[:, i, j, ...]
# nsamps * ochans
gy_slice = gy[..., i, j]
# -> nsamps * ichans * krows * kcols
gx[:, :, slice_row, slice_col] += xp.tensordot(
gy_slice, W_slice, axes=[(1,), (0,)]
)
ret.append(chainer.functions.cast(variable.as_variable(gx),
x.dtype))
if 1 in indices:
gW = xp.empty_like(W)
for i in moves.range(output_row):
for j in moves.range(output_col):
slice_row = slice(i * stride_row,
i * stride_row + W.shape[4])
slice_col = slice(j * stride_col,
j * stride_col + W.shape[5])
# nsamps * inchans * krows * kcols
x_slice = x[:, :, slice_row, slice_col]
# nsamps * outchans * 1 * 1
gy_slice = gy[:, :, i, j]
gW[:, i, j, :, :, :] = xp.tensordot(
gy_slice, x_slice, axes=[(0,), (0,)]
)
ret.append(chainer.functions.cast(variable.as_variable(gW),
W.dtype))
if 2 in indices:
gb = chainer.functions.sum(gyvar, axis=0)
ret.append(gb)
return ret
def backward(self, indexes, grad_outputs):
if self.grad is None:
grad = (self.xp.ones(self.input_shape, self.input_dtype), )
else:
grad = self.grad
return tuple(None if g is None else variable.as_variable(g)
for g in grad)
def backward(self, inputs, grad_outputs):
grad = self.grad
return tuple(None if g is None else variable.as_variable(g)
for g in grad)
def backward(self, inputs, grad_outputs):
grad = self.grad
return tuple(
None if g is None else variable.as_variable(g)
for g in grad)
