onnxmltools.convert_keras leads to parameters missing from target model

[luludak]

I converted InceptionV3 from Keras to ONNX, using onnxmltools.convert_keras(....). The original model is one obtained from the pretrained models included in Keras API.

Upon deeper inspection I noticed that some pad hyperparameters are missing to specific Conv nodes. For example, Pads is missing in target.Conv_34 vs model/conv2d_5/BiasAdd, as well as from target.Conv_36 node compared to model/conv2d_6/BiasAdd. In some cases, there are Pad nodes added preceding the related Conv nodes, but that is not always the case and even if they are present, the padding value is different across the original and the converter model. Any ideas why this happens?