def test_dygraph(self):
with fluid.dygraph.guard(fluid.CPUPlace()):
index_data = np.array([[1, 1], [0, 1], [1, 3]]).astype(np.int64)
index = fluid.dygraph.to_variable(index_data)
updates = paddle.rand(shape=[3, 9, 10], dtype='float32')
shape = [3, 5, 9, 10]
output = paddle.scatter_nd(index, updates, shape)
def mm_backwardcorrect_sparse_embed(sparseX: FloatTensor,
denseY: paddle.fluid.dygraph.core.VarBase):
update_inds = sparseX.indices[0]
updates = paddle.nn.functional.embedding(
sparseX.indices[1], denseY, sparse=False) * sparseX.values.view(-1, 1)
ret_Mat2 = paddle.scatter_nd(paddle.reshape(update_inds, (-1, 1)), updates,
(sparseX.shape[0], denseY.shape[1]))
return paddorch.convertTensor(ret_Mat2)
def segment_padding(data, segment_ids):
"""
Segment padding operator.
This operator padding the input elements which with the same index in 'segment_ids' to a common length ,
and reshape its into [uniq_segment_id, max_padding, dim].
Args:
data (tensor): a tensor, available data type float32, float64.
segment_ids (tensor): a 1-d tensor, which have the same size
with the first dimension of input data.
available data type is int32, int64.
Returns:
output (Tensor): the padding result with shape [uniq_segment_id, max_padding, dim].
seq_len (Tensor): the numbers of elements grouped same segment_ids
max_padding: the max number of elements grouped by same segment_ids
Examples:
.. code-block:: python
import paddle
import pgl
data = paddle.to_tensor([[1, 2, 3], [3, 2, 1], [4, 5, 6]], dtype='float32')
segment_ids = paddle.to_tensor([0, 0, 1], dtype='int64')
out = pgl.math.segment_softmax(data, segment_ids)
#Outputs: [[[1., 2., 3.], [3., 2., 1.]], [[4., 5., 6.], [0., 0., 0.]]], [2,1], 2
"""
idx_a = segment_ids
idx_b = paddle.arange(segment_ids.shape[0])
temp_idx = paddle.ones([segment_ids.shape[0]], dtype='float32')
temp_idx = segment_sum(temp_idx, segment_ids).astype('int32')
seq_len = temp_idx
max_padding = temp_idx.max().numpy()[0]
temp_idx = paddle.cumsum(temp_idx)
temp_idx_i = paddle.zeros([temp_idx.shape[0] + 1], dtype='int32')
temp_idx_i[1: ] = temp_idx
temp_idx = temp_idx_i[: -1]
temp_idx = paddle.gather(temp_idx, segment_ids)
idx_b = idx_b - temp_idx
index = paddle.stack([idx_a, idx_b], axis=1)
bz = segment_ids.max().numpy()[0] + 1
shape = [bz, max_padding, data.shape[-1]]
output = paddle.scatter_nd(index, data, shape)
return output, seq_len, index
def segment_padding(data, segment_ids):
"""
Segment padding operator.
This operator padding the input elements which with the same index in 'segment_ids' to a common length ,
and reshape its into [uniq_segment_id, max_padding, dim].
Args:
data (tensor): a tensor, available data type float32, float64.
segment_ids (tensor): a 1-d tensor, which have the same size
with the first dimension of input data.
available data type is int32, int64.
Returns:
output (Tensor): the padding result with shape [uniq_segment_id, max_padding, dim].
seq_len (Tensor): the numbers of elements grouped same segment_ids
index: The index of elements for gather_nd or scatter_nd operation
Examples:
.. code-block:: python
import paddle
import pgl
data = paddle.to_tensor([[1, 2, 3], [3, 2, 1], [4, 5, 6]], dtype='float32')
segment_ids = paddle.to_tensor([0, 0, 1], dtype='int64')
output, seq_len, index = pgl.math.segment_padding(data, segment_ids)
"""
idx_a = segment_ids
idx_b = paddle.arange(paddle.shape(segment_ids)[0])
temp_idx = paddle.ones_like(segment_ids, dtype='float32')
segment_len = segment_sum(temp_idx, segment_ids).astype('int32')
max_padding = paddle.max(segment_len)
segment_shift = get_index_from_counts(segment_len)[:-1]
segment_shift = paddle.gather(segment_shift, segment_ids)
idx_b = idx_b - segment_shift
index = paddle.stack([idx_a, idx_b], axis=1)
shape = [paddle.shape(segment_len)[0], max_padding, data.shape[-1]]
output = paddle.scatter_nd(index, data, shape)
return output, segment_len, index
def to_dense(self):
ret_mat = paddle.scatter_nd(paddle.transpose(self.indices, (1, 0)),
self.values, self.shape)
return ret_mat
def scatter_nd(index, updates):
shape = [3, 5, 9, 10]
return paddle.scatter_nd(index, updates, shape)