Add RequantClipFn to enable proper ONNX export of clipping: this allows exporting proper Clip nodes in ONNX instead of Min,Max symbolics.
Add FlattenCanonicaliser to transform a few known "bad" flatten templates into a single "good" one based on nn.Flatten.
Add an editor that can remove the final EpsTunnel, e.g., to transfer to a backend deployment tool that is not capable of interpreting the EpsTunnels.
Add AddCalibrator editor to harmonise HarmonisedAdd layers post-calibration. Before it was performed only during training.
Add DropoutRemover transform to remove nn.Dropout layers.
Add a AddRequantisationMerger to merge the requant resulting from HarmonisedAdd blocks with the previous ones (e.g., from regular nn.Relu). This can be useful, e.g., when the backend deployment tool does not support back-to-back requantisation layers.
Fixes to DORY ONNX exporter.
Use _ char instead of [,] to index modules inserted by editors (makes it easier to address them in code).
Clipnodes in ONNX instead ofMin,Maxsymbolics.nn.Flatten.EpsTunnel, e.g., to transfer to a backend deployment tool that is not capable of interpreting theEpsTunnels.AddCalibratoreditor to harmoniseHarmonisedAddlayers post-calibration. Before it was performed only during training.DropoutRemovertransform to removenn.Dropoutlayers.AddRequantisationMergerto merge the requant resulting fromHarmonisedAddblocks with the previous ones (e.g., from regularnn.Relu). This can be useful, e.g., when the backend deployment tool does not support back-to-back requantisation layers._char instead of[,]to index modules inserted by editors (makes it easier to address them in code).