def testGradientsExplicit(self):
sp_input = self._SparseTensor_4x6()
# SparseSliceGrad does not currently have a GPU kernel.
with test_util.force_cpu():
start, size = [0, 0], [4, 1]
sp_output = sparse_ops.sparse_slice(sp_input, start, size)
input_grad_vals = sparse_ops.sparse_slice_grad(
sp_output.values, sp_input.indices, start, sp_output.indices)
self.assertAllEqual(input_grad_vals,
[0, 0, 0, 0, 0, 0, 0, 20, 0, 0, 30, 0, 0, 0])
start, size = [0, 1], [4, 1]
sp_output = sparse_ops.sparse_slice(sp_input, start, size)
input_grad_vals = sparse_ops.sparse_slice_grad(
sp_output.values, sp_input.indices, start, sp_output.indices)
self.assertAllEqual(input_grad_vals,
[0, 0, 0, 0, 11, 0, 0, 0, 0, 0, 0, 0, 0, 0])
start, size = [1, 3], [3, 1]
sp_output = sparse_ops.sparse_slice(sp_input, start, size)
input_grad_vals = sparse_ops.sparse_slice_grad(
sp_output.values, sp_input.indices, start, sp_output.indices)
self.assertAllEqual(input_grad_vals,
[0, 0, 0, 0, 0, 13, 0, 0, 23, 0, 0, 0, 33, 0])
sp_input = self._SparseTensor_4x6_empty()
start, size = [0, 0], [4, 1]
sp_output = sparse_ops.sparse_slice(sp_input, start, size)
input_grad_vals = sparse_ops.sparse_slice_grad(
sp_output.values, sp_input.indices, start, sp_output.indices)
self.assertAllEqual(input_grad_vals, [])
def testSliceMatrixUnevenRows(self):
with self.session(use_gpu=False):
sp_input = self._SparseTensor_5x7()
sp_tensor0 = sparse_ops.sparse_slice(sp_input, [0, 0], [3, 7])
sp_tensor1 = sparse_ops.sparse_slice(sp_input, [3, 0], [3, 7])
self.assertAllEqual(sp_tensor0.indices.eval(),
[[0, 0], [0, 2], [0, 4], [0, 5], [1, 1], [1, 3],
[1, 4], [1, 6], [2, 0], [2, 3], [2, 5]])
self.assertAllEqual(sp_tensor0.values.eval(),
[0, 2, 4, 5, 11, 13, 14, 16, 20, 23, 25])
self.assertAllEqual(sp_tensor0.dense_shape.eval(), [3, 7])
self.assertAllEqual(
sp_tensor1.indices.eval(),
[[0, 0], [0, 2], [0, 3], [0, 5], [1, 1], [1, 4], [1, 6]])
self.assertAllEqual(sp_tensor1.values.eval(),
[30, 32, 33, 35, 41, 44, 46])
self.assertAllEqual(sp_tensor1.dense_shape.eval(), [2, 7])
sp_tensor0 = sparse_ops.sparse_slice(sp_input, [0, 0], [2, 7])
sp_tensor1 = sparse_ops.sparse_slice(sp_input, [2, 0], [2, 7])
sp_tensor2 = sparse_ops.sparse_slice(sp_input, [4, 0], [2, 7])
self.assertAllEqual(
sp_tensor0.indices.eval(),
[[0, 0], [0, 2], [0, 4], [0, 5], [1, 1], [1, 3], [1, 4], [1, 6]])
self.assertAllEqual(sp_tensor0.values.eval(),
[0, 2, 4, 5, 11, 13, 14, 16])
self.assertAllEqual(sp_tensor0.dense_shape.eval(), [2, 7])
self.assertAllEqual(sp_tensor1.values.eval(),
[20, 23, 25, 30, 32, 33, 35])
self.assertAllEqual(sp_tensor1.dense_shape.eval(), [2, 7])
self.assertAllEqual(sp_tensor2.indices.eval(), [[0, 1], [0, 4], [0, 6]])
self.assertAllEqual(sp_tensor2.values.eval(), [41, 44, 46])
self.assertAllEqual(sp_tensor2.dense_shape.eval(), [1, 7])
return
def testSliceMatrixUnevenRows(self):
with self.session(use_gpu=False):
sp_input = self._SparseTensor_5x7()
sp_tensor0 = sparse_ops.sparse_slice(sp_input, [0, 0], [3, 7])
sp_tensor1 = sparse_ops.sparse_slice(sp_input, [3, 0], [3, 7])
self.assertAllEqual(sp_tensor0.indices,
[[0, 0], [0, 2], [0, 4], [0, 5], [1, 1], [1, 3],
[1, 4], [1, 6], [2, 0], [2, 3], [2, 5]])
self.assertAllEqual(sp_tensor0.values,
[0, 2, 4, 5, 11, 13, 14, 16, 20, 23, 25])
self.assertAllEqual(sp_tensor0.dense_shape, [3, 7])
self.assertAllEqual(
sp_tensor1.indices,
[[0, 0], [0, 2], [0, 3], [0, 5], [1, 1], [1, 4], [1, 6]])
self.assertAllEqual(sp_tensor1.values, [30, 32, 33, 35, 41, 44, 46])
self.assertAllEqual(sp_tensor1.dense_shape, [2, 7])
sp_tensor0 = sparse_ops.sparse_slice(sp_input, [0, 0], [2, 7])
sp_tensor1 = sparse_ops.sparse_slice(sp_input, [2, 0], [2, 7])
sp_tensor2 = sparse_ops.sparse_slice(sp_input, [4, 0], [2, 7])
self.assertAllEqual(
sp_tensor0.indices,
[[0, 0], [0, 2], [0, 4], [0, 5], [1, 1], [1, 3], [1, 4], [1, 6]])
self.assertAllEqual(sp_tensor0.values, [0, 2, 4, 5, 11, 13, 14, 16])
self.assertAllEqual(sp_tensor0.dense_shape, [2, 7])
self.assertAllEqual(sp_tensor1.values, [20, 23, 25, 30, 32, 33, 35])
self.assertAllEqual(sp_tensor1.dense_shape, [2, 7])
self.assertAllEqual(sp_tensor2.indices, [[0, 1], [0, 4], [0, 6]])
self.assertAllEqual(sp_tensor2.values, [41, 44, 46])
self.assertAllEqual(sp_tensor2.dense_shape, [1, 7])
return
def testGradientsExplicit(self):
sp_input = self._SparseTensor_4x6()
start, size = [0, 0], [4, 1]
sp_output = sparse_ops.sparse_slice(sp_input, start, size)
input_grad_vals = sparse_ops.sparse_slice_grad(sp_output.values,
sp_input.indices, start,
sp_output.indices)
# pyformat: disable
self.assertAllEqual(input_grad_vals,
[0, 0, 0, 0, 0, 0, 0, 20, 0, 0, 30, 0, 0, 0])
# pyformat: enable
start, size = [0, 1], [4, 1]
sp_output = sparse_ops.sparse_slice(sp_input, start, size)
input_grad_vals = sparse_ops.sparse_slice_grad(sp_output.values,
sp_input.indices, start,
sp_output.indices)
# pyformat: disable
self.assertAllEqual(input_grad_vals,
[0, 0, 0, 0, 11, 0, 0, 0, 0, 0, 0, 0, 0, 0])
# pyformat: enable
start, size = [1, 3], [3, 1]
sp_output = sparse_ops.sparse_slice(sp_input, start, size)
input_grad_vals = sparse_ops.sparse_slice_grad(sp_output.values,
sp_input.indices, start,
sp_output.indices)
# pyformat: disable
self.assertAllEqual(input_grad_vals,
[0, 0, 0, 0, 0, 13, 0, 0, 23, 0, 0, 0, 33, 0])
# pyformat: enable
# Test empty slice of non-empty input.
start, size = [2, 1], [2, 1]
sp_output = sparse_ops.sparse_slice(sp_input, start, size)
input_grad_vals = sparse_ops.sparse_slice_grad(sp_output.values,
sp_input.indices, start,
sp_output.indices)
# pyformat: disable
self.assertAllEqual(input_grad_vals,
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0])
# pyformat: enable
sp_input = self._SparseTensor_4x6_empty()
start, size = [0, 0], [4, 1]
sp_output = sparse_ops.sparse_slice(sp_input, start, size)
input_grad_vals = sparse_ops.sparse_slice_grad(sp_output.values,
sp_input.indices, start,
sp_output.indices)
self.assertAllEqual(input_grad_vals, [])
def testSliceMatrixRows(self):
with self.session(use_gpu=False):
sp_input = self._SparseTensor_4x6()
sp_tensor0 = sparse_ops.sparse_slice(sp_input, [0, 0], [2, 6])
sp_tensor1 = sparse_ops.sparse_slice(sp_input, [2, 0], [3, 7])
self.assertAllEqual(
sp_tensor0.indices,
[[0, 0], [0, 2], [0, 4], [0, 5], [1, 1], [1, 3], [1, 4]])
self.assertAllEqual(sp_tensor0.values, [0, 2, 4, 5, 11, 13, 14])
self.assertAllEqual(sp_tensor0.dense_shape, [2, 6])
self.assertAllEqual(
sp_tensor1.indices,
[[0, 0], [0, 3], [0, 5], [1, 0], [1, 2], [1, 3], [1, 5]])
self.assertAllEqual(sp_tensor1.values, [20, 23, 25, 30, 32, 33, 35])
self.assertAllEqual(sp_tensor1.dense_shape, [2, 6])
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 testSliceMatrixRows(self):
with self.session(use_gpu=False):
sp_input = self._SparseTensor_4x6()
sp_tensor0 = sparse_ops.sparse_slice(sp_input, [0, 0], [2, 6])
sp_tensor1 = sparse_ops.sparse_slice(sp_input, [2, 0], [3, 7])
self.assertAllEqual(
sp_tensor0.indices.eval(),
[[0, 0], [0, 2], [0, 4], [0, 5], [1, 1], [1, 3], [1, 4]])
self.assertAllEqual(sp_tensor0.values.eval(), [0, 2, 4, 5, 11, 13, 14])
self.assertAllEqual(sp_tensor0.dense_shape.eval(), [2, 6])
self.assertAllEqual(
sp_tensor1.indices.eval(),
[[0, 0], [0, 3], [0, 5], [1, 0], [1, 2], [1, 3], [1, 5]])
self.assertAllEqual(sp_tensor1.values.eval(),
[20, 23, 25, 30, 32, 33, 35])
self.assertAllEqual(sp_tensor1.dense_shape.eval(), [2, 6])
def slice_fn(t=t, ndims=ndims, new_shape=new_shape):
"""Slices the tensor."""
if isinstance(t, sparse_tensor.SparseTensor):
return sparse_ops.sparse_slice(t, [0] * ndims,
math_ops.to_int64(new_shape))
else:
return array_ops.slice(t, [0] * ndims, new_shape)
def _ExtractImagePatchesGrad(op, grad):
batch_size, rows_in, cols_in, channels = [
dim.value for dim in op.inputs[0].shape.dims
]
input_bhwc = array_ops.shape(op.inputs[0])
batch_size = input_bhwc[0]
channels = input_bhwc[3]
# 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 + rows_in * cols_in
input_idx = array_ops.reshape(math_ops.range(1, input_indices_num,
dtype=ops.dtypes.int64),
(1, rows_in, cols_in, 1))
input_idx_patched = gen_array_ops.extract_image_patches(
input_idx,
op.get_attr("ksizes"),
op.get_attr("strides"),
op.get_attr("rates"),
op.get_attr("padding"))
# Create indices matrix for output tensor.
_, rows_out, cols_out, _ = [dim.value for dim in op.outputs[0].shape.dims]
_, ksize_r, ksize_c, _ = op.get_attr("ksizes")
# Indices for output start from 0.
output_indices_num = rows_out * cols_out * ksize_r * ksize_c
output_idx = array_ops.reshape(math_ops.range(output_indices_num,
dtype=ops.dtypes.int64),
(1, rows_out, cols_out, 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(
grad, (batch_size, rows_out, cols_out, ksize_r, ksize_c, channels)),
(1, 2, 3, 4, 0, 5))
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, (rows_in, cols_in, batch_size, channels))
grad_out = array_ops.transpose(grad_out, (2, 0, 1, 3))
return [grad_out]
def _ExtractImagePatchesGrad(op, grad):
input_bhwc = array_ops.shape(op.inputs[0], out_type=dtypes.int64)
batch_size, rows_in, cols_in, channels = input_bhwc[0], input_bhwc[1], \
input_bhwc[2], input_bhwc[3]
# 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 + rows_in * cols_in
input_idx = array_ops.reshape(
math_ops.range(1, input_indices_num, dtype=ops.dtypes.int64),
(1, rows_in, cols_in, 1))
input_idx_patched = gen_array_ops.extract_image_patches(
input_idx, op.get_attr("ksizes"), op.get_attr("strides"),
op.get_attr("rates"), op.get_attr("padding"))
# Create indices matrix for output tensor.
output_bhwc = array_ops.shape(op.outputs[0], out_type=dtypes.int64)
rows_out, cols_out = output_bhwc[1], output_bhwc[2]
_, ksize_r, ksize_c, _ = op.get_attr("ksizes")
# Indices for output start from 0.
output_indices_num = rows_out * cols_out * ksize_r * ksize_c
output_idx = array_ops.reshape(
math_ops.range(output_indices_num, dtype=ops.dtypes.int64),
(1, rows_out, cols_out, 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, rows_out, cols_out, ksize_r, ksize_c, channels)),
(1, 2, 3, 4, 0, 5))
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, (rows_in, cols_in, batch_size, channels))
grad_out = array_ops.transpose(grad_out, (2, 0, 1, 3))
return [grad_out]
def testSliceEmpty(self):
with test_util.use_gpu():
sp_empty = self._SparseTensor_4x6_empty()
sp_input = self._SparseTensor_4x6()
sparse_tensor0 = sparse_ops.sparse_slice(sp_empty, [0, 0], [4, 1])
sparse_tensor1 = sparse_ops.sparse_slice(sp_input, [1, 1], [0, 0])
sparse_tensor2 = sparse_ops.sparse_slice(sp_input, [2, 1], [2, 1])
empty_inds = np.empty(shape=(0, 2), dtype=np.int64)
self.assertAllEqual(sparse_tensor0.indices, empty_inds)
self.assertAllEqual(sparse_tensor0.values, [])
self.assertAllEqual(sparse_tensor0.dense_shape, [4, 1])
self.assertAllEqual(sparse_tensor1.indices, empty_inds)
self.assertAllEqual(sparse_tensor1.values, [])
self.assertAllEqual(sparse_tensor1.dense_shape, [0, 0])
self.assertAllEqual(sparse_tensor2.indices, empty_inds)
self.assertAllEqual(sparse_tensor2.values, [])
self.assertAllEqual(sparse_tensor2.dense_shape, [2, 1])
def testSliceEmpty(self):
# SparseSlice does not currently have a GPU kernel.
with test_util.force_cpu():
sp_empty = self._SparseTensor_4x6_empty()
sp_input = self._SparseTensor_4x6()
sparse_tensor0 = sparse_ops.sparse_slice(sp_empty, [0, 0], [4, 1])
sparse_tensor1 = sparse_ops.sparse_slice(sp_input, [1, 1], [0, 0])
sparse_tensor2 = sparse_ops.sparse_slice(sp_input, [2, 1], [2, 1])
empty_inds = np.empty(shape=(0, 2), dtype=np.int64)
self.assertAllEqual(sparse_tensor0.indices, empty_inds)
self.assertAllEqual(sparse_tensor0.values, [])
self.assertAllEqual(sparse_tensor0.dense_shape, [4, 1])
self.assertAllEqual(sparse_tensor1.indices, empty_inds)
self.assertAllEqual(sparse_tensor1.values, [])
self.assertAllEqual(sparse_tensor1.dense_shape, [0, 0])
self.assertAllEqual(sparse_tensor2.indices, empty_inds)
self.assertAllEqual(sparse_tensor2.values, [])
self.assertAllEqual(sparse_tensor2.dense_shape, [2, 1])
def testSliceColumns(self):
with self.session(use_gpu=False):
sp_input = self._SparseTensor_4x6()
sparse_tensor0 = sparse_ops.sparse_slice(sp_input, [0, 0], [4, 2])
sparse_tensor1 = sparse_ops.sparse_slice(sp_input, [0, 2], [5, 2])
sparse_tensor2 = sparse_ops.sparse_slice(sp_input, [0, 4], [5, 3])
self.assertAllEqual(sparse_tensor0.indices,
[[0, 0], [1, 1], [2, 0], [3, 0]])
self.assertAllEqual(sparse_tensor0.values, [0, 11, 20, 30])
self.assertAllEqual(sparse_tensor0.dense_shape, [4, 2])
self.assertAllEqual(sparse_tensor1.indices,
[[0, 0], [1, 1], [2, 1], [3, 0], [3, 1]])
self.assertAllEqual(sparse_tensor1.values, [2, 13, 23, 32, 33])
self.assertAllEqual(sparse_tensor1.dense_shape, [4, 2])
self.assertAllEqual(sparse_tensor2.indices,
[[0, 0], [0, 1], [1, 0], [2, 1], [3, 1]])
self.assertAllEqual(sparse_tensor2.values, [4, 5, 14, 25, 35])
self.assertAllEqual(sparse_tensor2.dense_shape, [4, 2])
def testSliceColumns(self):
with self.session(use_gpu=False):
sp_input = self._SparseTensor_4x6()
sparse_tensor0 = sparse_ops.sparse_slice(sp_input, [0, 0], [4, 2])
sparse_tensor1 = sparse_ops.sparse_slice(sp_input, [0, 2], [5, 2])
sparse_tensor2 = sparse_ops.sparse_slice(sp_input, [0, 4], [5, 3])
self.assertAllEqual(sparse_tensor0.indices.eval(),
[[0, 0], [1, 1], [2, 0], [3, 0]])
self.assertAllEqual(sparse_tensor0.values.eval(), [0, 11, 20, 30])
self.assertAllEqual(sparse_tensor0.dense_shape.eval(), [4, 2])
self.assertAllEqual(sparse_tensor1.indices.eval(),
[[0, 0], [1, 1], [2, 1], [3, 0], [3, 1]])
self.assertAllEqual(sparse_tensor1.values.eval(), [2, 13, 23, 32, 33])
self.assertAllEqual(sparse_tensor1.dense_shape.eval(), [4, 2])
self.assertAllEqual(sparse_tensor2.indices.eval(),
[[0, 0], [0, 1], [1, 0], [2, 1], [3, 1]])
self.assertAllEqual(sparse_tensor2.values.eval(), [4, 5, 14, 25, 35])
self.assertAllEqual(sparse_tensor2.dense_shape.eval(), [4, 2])
def testSliceAllRows(self):
with self.session():
sp_input = self._SparseTensor_4x6()
sp_tensor0 = sparse_ops.sparse_slice(sp_input, [0, 0], [1, 6])
sp_tensor1 = sparse_ops.sparse_slice(sp_input, [1, 0], [1, 6])
sp_tensor2 = sparse_ops.sparse_slice(sp_input, [2, 0], [1, 7])
sp_tensor3 = sparse_ops.sparse_slice(sp_input, [3, 0], [2, 7])
self.assertAllEqual(sp_tensor0.indices, [[0, 0], [0, 2], [0, 4], [0, 5]])
self.assertAllEqual(sp_tensor0.values, [0, 2, 4, 5])
self.assertAllEqual(sp_tensor0.dense_shape, [1, 6])
self.assertAllEqual(sp_tensor1.indices, [[0, 1], [0, 3], [0, 4]])
self.assertAllEqual(sp_tensor1.values, [11, 13, 14])
self.assertAllEqual(sp_tensor1.dense_shape, [1, 6])
self.assertAllEqual(sp_tensor2.indices, [[0, 0], [0, 3], [0, 5]])
self.assertAllEqual(sp_tensor2.values, [20, 23, 25])
self.assertAllEqual(sp_tensor2.dense_shape, [1, 6])
self.assertAllEqual(sp_tensor3.indices, [[0, 0], [0, 2], [0, 3], [0, 5]])
self.assertAllEqual(sp_tensor3.values, [30, 32, 33, 35])
self.assertAllEqual(sp_tensor3.dense_shape, [1, 6])
def testSliceAllColumns(self):
with self.session(use_gpu=False):
sp_input = self._SparseTensor_4x6()
sparse_tensor0 = sparse_ops.sparse_slice(sp_input, [0, 0], [4, 1])
sparse_tensor1 = sparse_ops.sparse_slice(sp_input, [0, 1], [4, 1])
sparse_tensor2 = sparse_ops.sparse_slice(sp_input, [0, 2], [4, 1])
sparse_tensor3 = sparse_ops.sparse_slice(sp_input, [0, 3], [4, 1])
sparse_tensor4 = sparse_ops.sparse_slice(sp_input, [0, 4], [5, 1])
sparse_tensor5 = sparse_ops.sparse_slice(sp_input, [0, 5], [6, 3])
self.assertAllEqual(sparse_tensor0.indices, [[0, 0], [2, 0], [3, 0]])
self.assertAllEqual(sparse_tensor0.values, [0, 20, 30])
self.assertAllEqual(sparse_tensor0.dense_shape, [4, 1])
self.assertAllEqual(sparse_tensor1.indices, [[1, 0]])
self.assertAllEqual(sparse_tensor1.values, [11])
self.assertAllEqual(sparse_tensor1.dense_shape, [4, 1])
self.assertAllEqual(sparse_tensor2.indices, [[0, 0], [3, 0]])
self.assertAllEqual(sparse_tensor2.values, [2, 32])
self.assertAllEqual(sparse_tensor2.dense_shape, [4, 1])
self.assertAllEqual(sparse_tensor3.indices, [[1, 0], [2, 0], [3, 0]])
self.assertAllEqual(sparse_tensor3.dense_shape, [4, 1])
self.assertAllEqual(sparse_tensor3.values, [13, 23, 33])
self.assertAllEqual(sparse_tensor4.indices, [[0, 0], [1, 0]])
self.assertAllEqual(sparse_tensor4.values, [4, 14])
self.assertAllEqual(sparse_tensor4.dense_shape, [4, 1])
self.assertAllEqual(sparse_tensor5.indices, [[0, 0], [2, 0], [3, 0]])
self.assertAllEqual(sparse_tensor5.values, [5, 25, 35])
self.assertAllEqual(sparse_tensor5.dense_shape, [4, 1])
def testSliceAllColumns(self):
with self.test_session(use_gpu=False):
sp_input = self._SparseTensor_4x6()
sparse_tensor0 = sparse_ops.sparse_slice(sp_input, [0, 0], [4, 1])
sparse_tensor1 = sparse_ops.sparse_slice(sp_input, [0, 1], [4, 1])
sparse_tensor2 = sparse_ops.sparse_slice(sp_input, [0, 2], [4, 1])
sparse_tensor3 = sparse_ops.sparse_slice(sp_input, [0, 3], [4, 1])
sparse_tensor4 = sparse_ops.sparse_slice(sp_input, [0, 4], [5, 1])
sparse_tensor5 = sparse_ops.sparse_slice(sp_input, [0, 5], [6, 3])
self.assertAllEqual(sparse_tensor0.indices.eval(),
[[0, 0], [2, 0], [3, 0]])
self.assertAllEqual(sparse_tensor0.values.eval(), [0, 20, 30])
self.assertAllEqual(sparse_tensor0.dense_shape.eval(), [4, 1])
self.assertAllEqual(sparse_tensor1.indices.eval(), [[1, 0]])
self.assertAllEqual(sparse_tensor1.values.eval(), [11])
self.assertAllEqual(sparse_tensor1.dense_shape.eval(), [4, 1])
self.assertAllEqual(sparse_tensor2.indices.eval(), [[0, 0], [3, 0]])
self.assertAllEqual(sparse_tensor2.values.eval(), [2, 32])
self.assertAllEqual(sparse_tensor2.dense_shape.eval(), [4, 1])
self.assertAllEqual(sparse_tensor3.indices.eval(),
[[1, 0], [2, 0], [3, 0]])
self.assertAllEqual(sparse_tensor3.dense_shape.eval(), [4, 1])
self.assertAllEqual(sparse_tensor3.values.eval(), [13, 23, 33])
self.assertAllEqual(sparse_tensor4.indices.eval(), [[0, 0], [1, 0]])
self.assertAllEqual(sparse_tensor4.values.eval(), [4, 14])
self.assertAllEqual(sparse_tensor4.dense_shape.eval(), [4, 1])
self.assertAllEqual(sparse_tensor5.indices.eval(),
[[0, 0], [2, 0], [3, 0]])
self.assertAllEqual(sparse_tensor5.values.eval(), [5, 25, 35])
self.assertAllEqual(sparse_tensor5.dense_shape.eval(), [4, 1])
def testGradients(self):
sp_input = self._SparseTensor_4x6(val_dtype=np.float32)
start_and_size = [([0, 0], [4, 2]), ([0, 2], [5, 2]), ([0, 4], [5, 3])]
with self.session(use_gpu=False):
for start, size in start_and_size:
sp_output = sparse_ops.sparse_slice(sp_input, start, size)
nnz_in = len(self.evaluate(sp_input.values))
nnz_out = len(self.evaluate(sp_output.values))
err = gradient_checker.compute_gradient_error(
[sp_input.values], [(nnz_in, )], sp_output.values,
(nnz_out, ))
self.assertLess(err, 1e-3)
def testGradients(self):
sp_input = self._SparseTensor_4x6(val_dtype=np.float32)
start_and_size = [([0, 0], [4, 2]),
([0, 2], [5, 2]),
([0, 4], [5, 3])]
with self.session(use_gpu=False):
for start, size in start_and_size:
sp_output = sparse_ops.sparse_slice(sp_input, start, size)
nnz_in = len(sp_input.values.eval())
nnz_out = len(sp_output.values.eval())
err = gradient_checker.compute_gradient_error(
[sp_input.values], [(nnz_in,)], sp_output.values, (nnz_out,))
self.assertLess(err, 1e-3)
def _embedding_lookup_for_sparse_tensor(
inp: sparse_tensor.SparseTensor,
weight: Optional[sparse_tensor.SparseTensor],
table: tf_variables.Variable,
feature: tpu_embedding_v2_utils.FeatureConfig) -> ops.Tensor:
"""Embedding lookup for sparse tensor based on its feature config.
Args:
inp: a single SparseTensor input.
weight: None or SparseTensor which has the same shape of the input.
table: a table variable.
feature: a feature config.
Returns:
Embedding lookup result.
"""
if not feature.output_shape and feature.max_sequence_length > 0:
batch_size = math_ops.cast(array_ops.shape(inp)[0], dtype=dtypes.int64)
sparse_shape = array_ops.stack(
[batch_size, feature.max_sequence_length], axis=0)
# TPU Embedding truncates sequences to max_sequence_length, and if we
# don't truncate, scatter_nd will error out if the index was out of
# bounds.
truncated_inp = sparse_ops.sparse_slice(inp,
start=[0, 0],
size=sparse_shape)
dense_output_shape = array_ops.stack(
[batch_size, feature.max_sequence_length, feature.table.dim],
axis=0)
return array_ops.scatter_nd(
truncated_inp.indices,
array_ops.gather(table.read_value(), truncated_inp.values),
dense_output_shape)
else:
inp_rank = inp.dense_shape.get_shape()[0]
if (not feature.validate_weights_and_indices and inp_rank is not None
and inp_rank <= 2):
return embedding_ops.embedding_lookup_sparse_v2(
table, inp, sp_weights=weight, combiner=feature.table.combiner)
else:
return embedding_ops.safe_embedding_lookup_sparse_v2(
table,
inp,
sparse_weights=weight,
combiner=feature.table.combiner)
def _transform_feature(self, inputs):
"""Returns dense `Tensor` representing feature.
Args:
inputs: A `_LazyBuilder` object to access inputs.
Returns:
Transformed feature `Tensor`.
Raises:
ValueError: if input rank is not known at graph building time.
"""
id_weight_pair = self.categorical_column._get_sparse_tensors(inputs) # pylint: disable=protected-access
id_tensor = id_weight_pair.id_tensor
weight_tensor = id_weight_pair.weight_tensor
# If the underlying column is weighted, return the input as a dense tensor.
if weight_tensor is not None:
weighted_column = sparse_ops.sparse_merge(
sp_ids=id_tensor,
sp_values=weight_tensor,
vocab_size=int(self._variable_shape[-1]))
# Remove (?, -1) index
weighted_column = sparse_ops.sparse_slice(weighted_column, [0, 0],
weighted_column.dense_shape)
#return sparse_ops.sparse_tensor_to_dense(weighted_column)
return array_ops.scatter_nd(weighted_column.indices,
weighted_column.values,
weighted_column.dense_shape)
dense_id_tensor = sparse_ops.sparse_tensor_to_dense(
id_tensor, default_value=-1)
# One hot must be float for tf.concat reasons since all other inputs to
# input_layer are float32.
one_hot_id_tensor = array_ops.one_hot(
dense_id_tensor,
depth=self._variable_shape[-1],
on_value=1.0,
off_value=0.0)
# Reduce to get a multi-hot per example.
return math_ops.reduce_sum(one_hot_id_tensor, axis=[-2])
def testSliceMatrixUnevenCols(self):
with self.session(use_gpu=False):
sp_input = self._SparseTensor_5x7()
sp_tensor0 = sparse_ops.sparse_slice(sp_input, [0, 0], [5, 3])
sp_tensor1 = sparse_ops.sparse_slice(sp_input, [0, 3], [5, 2])
sp_tensor2 = sparse_ops.sparse_slice(sp_input, [0, 5], [5, 2])
self.assertAllEqual(
sp_tensor0.indices.eval(),
[[0, 0], [0, 2], [1, 1], [2, 0], [3, 0], [3, 2], [4, 1]])
self.assertAllEqual(sp_tensor0.values.eval(),
[0, 2, 11, 20, 30, 32, 41])
self.assertAllEqual(sp_tensor0.dense_shape.eval(), [5, 3])
self.assertAllEqual(
sp_tensor1.indices.eval(),
[[0, 1], [1, 0], [1, 1], [2, 0], [3, 0], [4, 1]])
self.assertAllEqual(sp_tensor1.values.eval(),
[4, 13, 14, 23, 33, 44])
self.assertAllEqual(sp_tensor1.dense_shape.eval(), [5, 2])
self.assertAllEqual(sp_tensor2.indices.eval(),
[[0, 0], [1, 1], [2, 0], [3, 0], [4, 1]])
self.assertAllEqual(sp_tensor2.values.eval(), [5, 16, 25, 35, 46])
self.assertAllEqual(sp_tensor2.dense_shape.eval(), [5, 2])
sp_tensor0 = sparse_ops.sparse_slice(sp_input, [0, 0], [5, 2])
sp_tensor1 = sparse_ops.sparse_slice(sp_input, [0, 2], [5, 2])
sp_tensor2 = sparse_ops.sparse_slice(sp_input, [0, 4], [5, 2])
sp_tensor3 = sparse_ops.sparse_slice(sp_input, [0, 6], [5, 2])
self.assertAllEqual(sp_tensor0.indices.eval(),
[[0, 0], [1, 1], [2, 0], [3, 0], [4, 1]])
self.assertAllEqual(sp_tensor0.values.eval(), [0, 11, 20, 30, 41])
self.assertAllEqual(sp_tensor0.dense_shape.eval(), [5, 2])
self.assertAllEqual(sp_tensor1.indices.eval(),
[[0, 0], [1, 1], [2, 1], [3, 0], [3, 1]])
self.assertAllEqual(sp_tensor1.values.eval(), [2, 13, 23, 32, 33])
self.assertAllEqual(sp_tensor1.dense_shape.eval(), [5, 2])
self.assertAllEqual(
sp_tensor2.indices.eval(),
[[0, 0], [0, 1], [1, 0], [2, 1], [3, 1], [4, 0]])
self.assertAllEqual(sp_tensor2.values.eval(),
[4, 5, 14, 25, 35, 44])
self.assertAllEqual(sp_tensor2.dense_shape.eval(), [5, 2])
self.assertAllEqual(sp_tensor3.indices.eval(), [[1, 0], [4, 0]])
self.assertAllEqual(sp_tensor3.values.eval(), [16, 46])
self.assertAllEqual(sp_tensor3.dense_shape.eval(), [5, 1])
def testSliceMatrixUnevenCols(self):
with self.test_session(use_gpu=False):
sp_input=self._SparseTensor_5x7()
sp_tensor0 = sparse_ops.sparse_slice(sp_input, [0, 0], [5, 3])
sp_tensor1 = sparse_ops.sparse_slice(sp_input, [0, 3], [5, 2])
sp_tensor2 = sparse_ops.sparse_slice(sp_input, [0, 5], [5, 2])
self.assertAllEqual(sp_tensor0.indices.eval(),
[[0, 0], [0, 2], [1, 1], [2, 0], [3, 0], [3, 2],
[4, 1]])
self.assertAllEqual(sp_tensor0.values.eval(),
[0, 2, 11, 20, 30, 32, 41])
self.assertAllEqual(sp_tensor0.dense_shape.eval(), [5, 3])
self.assertAllEqual(sp_tensor1.indices.eval(),
[[0, 1], [1, 0], [1, 1], [2, 0], [3, 0], [4, 1]])
self.assertAllEqual(sp_tensor1.values.eval(),
[4, 13, 14, 23, 33, 44])
self.assertAllEqual(sp_tensor1.dense_shape.eval(), [5, 2])
self.assertAllEqual(sp_tensor2.indices.eval(),
[[0, 0], [1, 1], [2, 0], [3, 0], [4, 1]])
self.assertAllEqual(sp_tensor2.values.eval(), [5, 16, 25, 35, 46])
self.assertAllEqual(sp_tensor2.dense_shape.eval(), [5, 2])
sp_tensor0 = sparse_ops.sparse_slice(sp_input, [0, 0], [5, 2])
sp_tensor1 = sparse_ops.sparse_slice(sp_input, [0, 2], [5, 2])
sp_tensor2 = sparse_ops.sparse_slice(sp_input, [0, 4], [5, 2])
sp_tensor3 = sparse_ops.sparse_slice(sp_input, [0, 6], [5, 2])
self.assertAllEqual(sp_tensor0.indices.eval(),
[[0, 0], [1, 1], [2, 0], [3, 0], [4, 1]])
self.assertAllEqual(sp_tensor0.values.eval(), [0, 11, 20, 30, 41])
self.assertAllEqual(sp_tensor0.dense_shape.eval(), [5, 2])
self.assertAllEqual(sp_tensor1.indices.eval(),
[[0, 0], [1, 1], [2, 1], [3, 0], [3, 1]])
self.assertAllEqual(sp_tensor1.values.eval(), [2, 13, 23, 32, 33])
self.assertAllEqual(sp_tensor1.dense_shape.eval(), [5, 2])
self.assertAllEqual(sp_tensor2.indices.eval(),
[[0, 0], [0, 1], [1, 0], [2, 1], [3, 1], [4, 0]])
self.assertAllEqual(sp_tensor2.values.eval(), [4, 5, 14, 25, 35, 44])
self.assertAllEqual(sp_tensor2.dense_shape.eval(), [5, 2])
self.assertAllEqual(sp_tensor3.indices.eval(), [[1, 0], [4, 0]])
self.assertAllEqual(sp_tensor3.values.eval(), [16, 46])
self.assertAllEqual(sp_tensor3.dense_shape.eval(), [5, 1])
def pad_sparse_embedding_lookup_indices(sparse_indices, padded_size):
"""Creates statically-sized Tensors containing indices and weights.
From third_party/cloud_tpu/models/movielens/tpu_embedding.py
Also computes sparse_indices.values % embedding_table_size, for equivalent
functionality to sparse_column_with_integerized_feature. The returned
padded weight Tensor also doubles as a mask indicating which values in
the returned padded indices Tensor are indices versus padded zeros.
Args:
sparse_indices: SparseTensor of embedding lookup indices.
padded_size: Number of columns of the returned Tensors. Indices which fall
out of bounds will be truncated to the padded size.
Returns:
(sparse_indices.values padded to the specified size,
a mask the same size as the returned padded values in which 0s
indicate padded locations and 1s (or values from sparse_weights)
indicate actual values)
"""
batch_size = sparse_indices.dense_shape[0]
sparse_indices = sparse_ops.sparse_slice(sparse_indices, [0, 0],
[batch_size, padded_size])
indices, values = sparse_indices.indices, sparse_indices.values
padded_values = array_ops.scatter_nd(indices,
math_ops.cast(values, dtypes.int32),
shape=(batch_size, padded_size))
weights = array_ops.ones_like(values, dtype=dtypes.float32)
padded_mask = array_ops.scatter_nd(indices,
weights,
shape=(batch_size, padded_size))
return padded_values, padded_mask
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]
