def col2im_indices(cols, x_shape, filter=(2, 2), padding=0, stride=(1, 1)):
""" An implementation of col2im based on fancy indexing and np.add.at
Args:
cols: [B*OH*OW, kH*kW*Cin]
x_shape: Shape of initial input (B, H, W, Cin)
filter: Shape of filter (kH, kW)
Returns: [B, H, W, Cin]
"""
N, H, W, C = x_shape
H_padded, W_padded = H + 2 * padding, W + 2 * padding
x_padded = np.zeros((N, C, H_padded, W_padded), dtype=cols.dtype)
k, i, j = get_im2col_indices((N, C, H, W), filter[0], filter[1], padding,
stride)
# cols => [B*OH*OW, kH*kW*Cin]
# cols_reshaped => [B, OH*OW, kH*kW*Cin]
cols_reshaped = cols.reshape(N, -1, C * filter[0] * filter[1])
# [B, C * kH * kW, OH*OW]
cols_reshaped = cols_reshaped.transpose(0, 2, 1)
idx = np.array(range(x_padded.shape[0])).reshape((x_padded.shape[0], 1, 1))
np.scatter_add(x_padded.astype(np.float32),
(idx.astype(np.uint64), k.astype(
np.uint64), i.astype(np.uint64), j.astype(np.uint64)),
cols_reshaped.astype(np.float32))
# np.add.at(x_padded, (slice(None), k, i, j), cols_reshaped)
if padding == 0:
return x_padded.transpose(0, 2, 3, 1)
return x_padded[:, :, padding:-padding,
padding:-padding].transpose(0, 2, 3, 1)
def test_scatter_add(self):
a = cupy.zeros((3, ), dtype=numpy.float32)
i = cupy.array([1, 1], numpy.int32)
v = cupy.array([2., 1.], dtype=numpy.float32)
with testing.assert_warns(DeprecationWarning):
cupy.scatter_add(a, i, v)
testing.assert_array_equal(a, cupy.array([0, 3, 0]))
def col2im_indices(self, cols):
field_height, field_width, padding, stride = self.k, self.k, self.p, self.s
N, C, H, W = self.x_shape
H_padded, W_padded = H + 2 * padding, W + 2 * padding
x_padded = cp.zeros((N, C, H_padded, W_padded), dtype=cols.dtype)
k, i, j = self.get_im2col_indices()
cols_reshaped = cols.reshape(C * field_height * field_width, -1, N)
cols_reshaped = cols_reshaped.transpose(2, 0, 1).astype(cp.float32)
cp.scatter_add(x_padded, (slice(None), k, i, j), cols_reshaped)
if padding == 0:
return x_padded
return x_padded[:, :, padding:-padding, padding:-padding]
def column_to_image(cols,
images_shape,
filter_shape,
stride,
output_shape='same'):
batch_size, channels, height, width = images_shape
pad_h, pad_w = determine_padding(filter_shape, output_shape)
height_padded = height + pad_h[0] + pad_h[1]
width_padded = width + pad_w[0] + pad_w[1]
images_padded = cp.zeros(
(batch_size, channels, height_padded, width_padded))
k, i, j = get_im2col_indices(images_shape, filter_shape, (pad_h, pad_w),
stride)
cols = cp.reshape(
cols, (channels * filter_shape[0] * filter_shape[1], -1, batch_size))
cols = cp.transpose(cols, (2, 0, 1))
cp.scatter_add(images_padded, (slice(None), k, i, j), cols)
return images_padded[:, :, pad_h[0]:height + pad_h[0],
pad_w[0]:width + pad_w[0]]
def test_scatter_add(self):
a = cupy.zeros((3,), dtype=numpy.float32)
i = cupy.array([1, 1], numpy.int32)
v = cupy.array([2., 1.], dtype=numpy.float32)
cupy.scatter_add(a, i, v)
testing.assert_array_equal(a, cupy.array([0, 3, 0]))