Python API not passing custom ops and rewriters properly

[vloncar]

Describe the bug
from_keras(), from_function() and from_graph_def() functions incorrectly process custom_op_handlers and custom rewriter parameters (maybe others as well, but these two I need). The custom_op_handlers is not used, and instead custom_ops is passed to process_tf_graph in its place. The custom_rewriter is ignored altogether, removing the possibility to use this feature.

Urgency
None

System information
Doesn't matter

To Reproduce
Just try to pass something custom_op_handlers or custom_rewriter to tf2onnx.convert.from_keras()

Screenshots
Not needed

Additional context
Not needed

[TomWildenhain-Microsoft]

Hi @vloncar, you are correct. Thanks for the investigation and report. #1708 should fix the issue.

By the way, custom_op_handlers is deprecated and you may find it easier to use @tf_op to register an op handler.

from tf2onnx.handler import tf_op

Example:

tensorflow-onnx/tf2onnx/custom_opsets/string_ops.py

Line 36 in 47acb8a

@tf_op("StringToHashBucketFast", domain=constants.CONTRIB_OPS_DOMAIN)

Just curious, what op handlers are you registering?

[vloncar]

Hi, thanks for the quick fix. I'm using the feature to parse quantized layers to a custom onnx op.

One extra question: is there maybe a way to use the graph manipulation tools on an existing ONNX model? For example I would like to insert nodes after the model was fully optimized (i.e. it has gemm nodes and the bias has been fused into the conv layer)

[TomWildenhain-Microsoft]

@vloncar Maybe the easiest thing would be cloning the tf2onnx repo. Then you can edit/add any editing steps you want.