Allow passing optimizers to from_keras conversion function

[shelkesagar29]

_convert_common function in convert.py has optimizers parameter which allows user to select a subset of _optimizers from tf2onnx.optimizer while conversion is being done.

One way of using it to convert a TF saved model can be as follows.

from tf2onnx.optimizer import _optimizers
# Get all optimizers
updated_optimizers = copy.deepcopy(_optimizers)
# Remove two optimizers and then apply remaining optimizers during conversion
del updated_optimizers["q_dq_optimizer"]
del updated_optimizers["const_dequantize_optimizer"]

# 1. Load TensorFlow saved model
graph_def, inputs, outputs = tf_loader.from_saved_model(
    model_path=saved_model_path,
    input_names=None,
    output_names=None,
    tag="serve",
    signatures=["serving_default"]
)

# 2. Convert tf2onnx and save onnx file
model_proto, _ = convert._convert_common(
    graph_def,
    name=saved_model,
    opset=13,
    input_names=inputs,
    output_names=outputs,
    output_path=onnx_output_path,
    optimizers=updated_optimizers
)

However, currently there is no way to pass optimizers when using from_keras function to directly convert a keras model (without saving and loading). Converting from Keras directly is important since TF gives error many times for saving custom layers (this is especially true for QAT).

This PR adds optimizers parameter to from_keras function and passes it to _convert_common.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

Sagar Shelke seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[hwangdeyu]

Thanks for contribution!
Could you sign off every commit message following DCO required?

Signed-off-by: Author Name <[email protected]>

[hwangdeyu]

LGTM.