Skip to content

Conversation

PawelGlomski-Intel
Copy link
Contributor

@PawelGlomski-Intel PawelGlomski-Intel commented May 27, 2021

Summary

Improving MXNet integration. This change will focus on improving the current features.

Motivation

Changes

In mxnet.yaml:

  • using only "int8" for operators that are meant to be quantized (and discard uint8) until it is possible in MXNet to select dtype of quantization per node/op (now "auto" dtype is hardcoded in MXNet adaptor)
  • removed granularity until it is possible in MXNet to select the granularity of quantization per node/op (now "tensor-wise" is hardcoded in MXNet adaptor)
  • removed bf16 until it is supported in LPOT for MXNet
  • fixed names of operators

In mxnet.py:

  • quantizable_ops selection
  • calib_<kl/minmax>_layers selection

Please review the structure of my 1.9.0 configuration and decide whether this can be applied to the rest of the yaml file.

guomingz and others added 11 commits May 10, 2021 10:39
intel#12)

* Fix the graph optimization bug that inputs/outputs field is not optional.

Signed-off-by: Zhang, Guoming <guoming.zhang@intel.com>

* Add postfix.
intel#16)

* Fix the graph optimization issue that setting precision to bf16 led crash on non-avx512-bf16 enabled platform.

* Enhance the graph optimization by sorting the specified precisions.

* Update the doc.

Co-authored-by: Tian, Feng <feng.tian@intel.com>
Signed-off-by: Deb Taylor <deb.taylor@intel.com>
Reviewed-by: Feng Tian <feng.tian@intel.com>

Co-authored-by: Deb Taylor <deb.taylor@intel.com>
Signed-off-by: Deb Taylor <deb.taylor@intel.com>
Reviewed-by: Feng Tian <feng.tian@intel.com>

Co-authored-by: Deb Taylor <deb.taylor@intel.com>
Proposition of yaml structure for MXNet
'int8': {
'_sg_mkldnn_conv': {
'activation': {
'dtype': ['int8', 'fp32'],
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Would it be possible to add "autoint8" to the possible values of "dtype" attribute, which would correspond to the automatic selection of dtype according to the calibration result of each tensor?

Suggested change
'dtype': ['int8', 'fp32'],
'dtype': ['autoint8', 'fp32'],

This way we could describe more clearly what is actually happening in MXNet.

@ftian1 ftian1 force-pushed the master branch 2 times, most recently from 42ec628 to 4939eb6 Compare May 31, 2021 02:33
chensuyue pushed a commit that referenced this pull request Oct 17, 2022
VincyZhang added a commit that referenced this pull request Feb 12, 2023
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants