[RELAY][OP] expand_dims #1819
[RELAY][OP] expand_dims #1819
Conversation
|
|
Please also request reviews from reviewers |
|
@tqchen @srkreddy1238 Hey could you help review this PR? |
|
Rebased to current master. |
|
sorry, still need another rebase after #1821 |
|
|
||
| num_newaxis : int | ||
| Number of axises to be inserted. Should be >= 0. | ||
| """ |
There was a problem hiding this comment.
The result. The number of dimensions is num_newaxis greater than that of the input.
There was a problem hiding this comment.
@tqchen Sorry for my poor English...What does this sentence mean: "The number of dimensions is num_newaxis greater than that of the input."?
There was a problem hiding this comment.
maybe result_dimension = source dimension + num_newaxis
There was a problem hiding this comment.
@tqchen Good point! Like this?
result : relay.Expr
The reshaped result. `result.ndim = data.ndim + num_newaxis`.
|
Please also add the operator to https://github.com/dmlc/tvm/blob/master/docs/langref/relay_op.rst |
|
|
||
| num_newaxis : int | ||
| Number of axises to be inserted. Should be >= 0. | ||
| """ |
There was a problem hiding this comment.
The result. The number of dimensions is num_newaxis greater than that of the input.
|
@tqchen Should I add examples somewhere? |
This PR contains a relay operator,
expand_dims.TODO:
axisin topi. Did I miss anything?expand_dims(data, axis, num_newaxis=1), but numpy API isexpand_dims(a, axis). I think it is okay because ours seem to be a superset.axisandnum_newaxisasint. Wondering if it is possible to make themIndexExpr, so thataxiscould depend on some intermediate result of computation.I am working on the docstring, and @tqchen could you help comment on the TODO 2, 3, 4?
Related issue: #1799