def _ExtractVolumePatchesGrad(op, grad):
batch_size, planes_in, rows_in, cols_in, channels = [
dim.value for dim in op.inputs[0].shape.dims
]
input_bphwc = array_ops.shape(op.inputs[0])
batch_size = input_bphwc[0]
channels = input_bphwc[4]
# Create indices matrix for input tensor.
# Note that 0 is preserved for padding location,
# so indices for input start from 1 to 1 + rows_in * cols_in.
input_indices_num = 1 + planes_in * rows_in * cols_in
input_idx = array_ops.reshape(
math_ops.range(1, input_indices_num, dtype=ops.dtypes.int64),
(1, planes_in, rows_in, cols_in, 1))
input_idx_patched = gen_array_ops.extract_volume_patches(
input_idx, op.get_attr("ksizes"), op.get_attr("strides"),
op.get_attr("padding"))
# Create indices matrix for output tensor.
_, planes_out, rows_out, cols_out, _ = [
dim.value for dim in op.outputs[0].shape.dims
]
_, ksize_p, ksize_r, ksize_c, _ = op.get_attr("ksizes")
# Indices for output start from 0.
prc_indices_num = planes_out * rows_out * cols_out
output_indices_num = prc_indices_num * ksize_p * ksize_r * ksize_c
output_idx = array_ops.reshape(
math_ops.range(output_indices_num, dtype=ops.dtypes.int64),
(1, planes_out, rows_out, cols_out, ksize_p * ksize_r * ksize_c))
# Construct mapping table for indices: (input -> output).
idx_matrix = array_ops.concat([
array_ops.expand_dims(input_idx_patched, axis=-1),
array_ops.expand_dims(output_idx, axis=-1)
],
axis=-1)
idx_map = array_ops.reshape(idx_matrix, (-1, 2))
sp_shape = (input_indices_num, output_indices_num)
sp_mat_full = sparse_tensor.SparseTensor(
idx_map, array_ops.ones([output_indices_num], dtype=grad.dtype), sp_shape)
# Remove all padding locations [0, :].
sp_mat = sparse_ops.sparse_slice(sp_mat_full, (1, 0),
(input_indices_num - 1, output_indices_num))
grad_expanded = array_ops.transpose(
array_ops.reshape(
_IndexedSlicesToTensorNoWarning(grad),
(batch_size, planes_out, rows_out, cols_out, ksize_p, ksize_r,
ksize_c, channels)), (1, 2, 3, 4, 5, 6, 0, 7))
grad_flat = array_ops.reshape(grad_expanded, (-1, batch_size * channels))
jac = sparse_ops.sparse_tensor_dense_matmul(sp_mat, grad_flat)
grad_out = array_ops.reshape(
jac, (planes_in, rows_in, cols_in, batch_size, channels))
grad_out = array_ops.transpose(grad_out, (3, 0, 1, 2, 4))
return [grad_out]
def _ExtractVolumePatchesGrad(op, grad):
batch_size, planes_in, rows_in, cols_in, channels = [
dim.value for dim in op.inputs[0].shape.dims
]
input_bphwc = array_ops.shape(op.inputs[0])
batch_size = input_bphwc[0]
channels = input_bphwc[4]
# Create indices matrix for input tensor.
# Note that 0 is preserved for padding location,
# so indices for input start from 1 to 1 + rows_in * cols_in.
input_indices_num = 1 + planes_in * rows_in * cols_in
input_idx = array_ops.reshape(
math_ops.range(1, input_indices_num, dtype=ops.dtypes.int64),
(1, planes_in, rows_in, cols_in, 1))
input_idx_patched = gen_array_ops.extract_volume_patches(
input_idx, op.get_attr("ksizes"), op.get_attr("strides"),
op.get_attr("padding"))
# Create indices matrix for output tensor.
_, planes_out, rows_out, cols_out, _ = [
dim.value for dim in op.outputs[0].shape.dims
]
_, ksize_p, ksize_r, ksize_c, _ = op.get_attr("ksizes")
# Indices for output start from 0.
prc_indices_num = planes_out * rows_out * cols_out
output_indices_num = prc_indices_num * ksize_p * ksize_r * ksize_c
output_idx = array_ops.reshape(
math_ops.range(output_indices_num, dtype=ops.dtypes.int64),
(1, planes_out, rows_out, cols_out, ksize_p * ksize_r * ksize_c))
# Construct mapping table for indices: (input -> output).
idx_matrix = array_ops.concat([
array_ops.expand_dims(input_idx_patched, axis=-1),
array_ops.expand_dims(output_idx, axis=-1)
],
axis=-1)
idx_map = array_ops.reshape(idx_matrix, (-1, 2))
sp_shape = (input_indices_num, output_indices_num)
sp_mat_full = sparse_tensor.SparseTensor(
idx_map, array_ops.ones([output_indices_num], dtype=grad.dtype), sp_shape)
# Remove all padding locations [0, :].
sp_mat = sparse_ops.sparse_slice(sp_mat_full, (1, 0),
(input_indices_num - 1, output_indices_num))
with warnings.catch_warnings():
warnings.filterwarnings(
"ignore",
message="Converting sparse IndexedSlices to a dense Tensor.*")
grad_expanded = array_ops.transpose(
array_ops.reshape(grad, (batch_size, planes_out, rows_out, cols_out,
ksize_p, ksize_r, ksize_c, channels)),
(1, 2, 3, 4, 5, 6, 0, 7))
grad_flat = array_ops.reshape(grad_expanded, (-1, batch_size * channels))
jac = sparse_ops.sparse_tensor_dense_matmul(sp_mat, grad_flat)
grad_out = array_ops.reshape(
jac, (planes_in, rows_in, cols_in, batch_size, channels))
grad_out = array_ops.transpose(grad_out, (3, 0, 1, 2, 4))
return [grad_out]
