Zha0q1 opened a new issue #20453: URL: https://github.com/apache/incubator-mxnet/issues/20453
Hi MXNet community, I am opening this issue to track the development progress of the onnx export module, mx2onnx, on MXNet 2.0. In MXNet 1.x, mx2onnx covers popular operators that's exposed under the `Symbol` namespace, and the operator unit tests utilize Gluon blocks to construct test models that consist of only one operator. However, in MXNet 2.0, the development has steered toward NumPy compatible operators and NumPy Extension operators and Gluon also moves on from using the `Symbol` namespace. With that said, mx2onnx will be consistent with this design choice and move toward support operators exposed under the `mx.np` and `mx.npx` namespaces. Here's a [PR](https://github.com/apache/incubator-mxnet/pull/20355) to forward port mx2onnx from v1.x to master. This PR also adapts existing operator conversion functions to their equivalent `np` or `npx` operators whenever possible. With this PR we now support the following operators ``` ## np absolute zeros ones zeors_like ones_like full_like arange reshape(naive support) concatenate transpose expand_dims add equal (missing scalar case) not_equal (missing scalar case) greater (missing scalar case) less (missing scalar case) greater_equal (missing scalar case) less_equal (missing scalar case) minimum maximum where add subtract multiply divide power argmin argmax minimum maximum min max mean sum prod mod swapaxes log2 reciprocal power sqrt square logical_and logical_xor logical_or broadcast_to clip logical_not tile sin cos tan tanh arcsin arccos arctan exp log ceil floor squeeze negative ## npx slice arange_like layer_norm sequence_mask reshape(naive support) fully_connected leaky_relu activation dropout cast softmax box_nms broadcast_greater topk box_decode reshape_like gather_nd slice_channel broadcast_like pooling roi_pooling convolution decovolution pad instance_norm shape_array log_softmax norm sequence_reverse one_hot batch_norm slice_channel ``` Further works on mx2onnx should include: 1. Add more conversion functions for the rest of `np` and `npx` operators so we have a better coverage 2. Add model level tests. Currently we only have operator unit tests as there are no model developed with only the new `np` and `npx` operators. 3. Code clean up: we can remove the deprecated operator conversion functions -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
