def test_index_add():
A = np.zeros((2, INT_OVERFLOW))
ind = np.array([[0, 0], [0, 1]], dtype='int32')
val = np.array([100, 200])
A.attach_grad()
with mx.autograd.record():
B = npx.index_add(A, ind, val)
assert B.shape == (2, INT_OVERFLOW)
assert B[0][0] == 100 and B[0][1] == 200
B.backward()
assert A.grad.shape == (2, INT_OVERFLOW)
assert A.grad[0][0] == 1
def add_vectors_by_position(data, increment, positions):
"""Scatter each batch with the given positions.
data[i, positions[i, j], ...] += increment[i, j, ...]
Parameters
----------
F
data
Input tensor of the array to be updated.
Shape (batch_size, seq_length, ...)
increment
Input tensor of token ids
Shape (batch_size, num_disp_position, ...)
positions
Input tensor of the positions.
Shape (batch_size, num_disp_position).
For each sample in the batch, the values in this tensor must not exceed
the length of the sequence.
Returns
-------
out
The updated result.
Shape (batch_size, seq_length, ...)
"""
# Here, we use index_add to disperse the output from data:
# Need to compute
# out[i, masked_position[i, j], :] = in[i, j, :]
# Thus, construct an indices with shape [2, batch_size * num_masked_position], where
# indices[0, i * num_masked_position + j] = i
# indices[1, i * num_masked_position + j] = masked_position[i, j]
# And convert data to the shape of the (batch_size * num_masked_position, )
# Then, out = npx.index_add(data, indices, increment)
positions = positions.astype(np.int32)
# batch_idx.shape = (batch_size, 1) as [[0], [1], [2], ...]
batch_idx = np.expand_dims(npx.arange_like(positions, axis=0),
axis=1).astype(np.int32)
batch_idx = batch_idx + np.zeros_like(positions)
indices = np.stack([batch_idx.reshape((-1, )), positions.reshape((-1, ))])
out = npx.index_add(data, indices, npx.reshape(increment, (-5, -4)))
return out
