support cumprod op #2133
support cumprod op #2133
Conversation
Signed-off-by: Francesco Salvetti <[email protected]>
Signed-off-by: Francesco Salvetti <[email protected]>
This reverts commit 6ad3231. Signed-off-by: Francesco Salvetti <[email protected]>
Signed-off-by: Francesco Salvetti <[email protected]>
f-salvetti
force-pushed
the
support-cumprod-op
branch
from
565b28c to
d791944
Compare
Signed-off-by: Francesco Salvetti <[email protected]>
Signed-off-by: Francesco Salvetti <[email protected]>
| pad_const_node = one_node | ||
| pad_node = graph.make_node("Pad", inputs=[slice_node.output[0], pads_node.output[0], pad_const_node.output[0]], | ||
| op_name_scope=loop_name, outputs=[utils.make_name("pad")]) | ||
| mul_node = graph.make_node("Mul", inputs=[graph.input_names[4], pad_node.output[0]], |
There was a problem hiding this comment.
Is it possible to leverage reduce_prod to avoid multiplying those values one by one?
There was a problem hiding this comment.
sure, didn't think about that. I will refactor the code in order to use that operator
There was a problem hiding this comment.
@fatcat-z Actually, using ReduceProd, would require to loop over slices of the input tensor to be reduced and then concatenate results to create the output tensor. However, in order to support a non-constant axis for the CumProd op as in TF, I don't think it is possible to use the Concat op, since it requires the axis as an attribute. Due to this limitation, I don't see how I can use ReduceProd instead of Mul.
|
The 2 CI failures are not supposed to be relative to your changes. I will figure them out. |
- update current opset to 18 - added Cumprod (onnx#2133) - added ReduceJoin (onnx#2091) Signed-off-by: f-salvetti <[email protected]>
Supports cumprod TF op. Both reverse and exclusive flags are supported.