Page History

...

sym, params = mx.serde.import(input_file, ininput_format=‘onnx’, outoutput_format=’gluon’)

• input_file : input model file (e.g., protobuf model file for onnx)
• ininput_format : (optional) onnx, coreml
• outoutput_format : (optional) “gluon/symbolic”. by default, gluon will be used.

...

1) mx.serde.export(sym, params, input_format='symbolic', output_format=‘onnx’, filename_prefix=”model_name”)

2) mx.serde.export(module, input_format='symbolic', output_format=‘onnx’, filename_prefix=”model_name”)

3) mx.serde.export(gluon_model, input_format='gluon', output_format=‘onnx’, filename_prefix=”model_name”)

• sym : model definition
• module : mxnet module object
• gluon_model : model definition (HybridBlock)
• params : weights
• input_format : symbolic/gluon 
• output_format : onnx, coreml
• filename_prefix: a filename prefix to be used to model save files. E.g., for onnx, a binary protobuf will be written to output file with “.onnx” extension.

...

mx.gluon.import(input_file, ininput_format=‘onnx’)

mx.gluon.export(output_format=’onnx’, filename_prefix=”model_name”)

...

e.g. `mx.gluon.import(input_file, ininput_format=‘onnx’)` will internally call `mx.serde.import(input_file, ininput_format=‘onnx’, outoutput_format='gluon')`

•  For Symbolic interface

sym, params = mx.mod.Module.import(input_file, ininput_format=‘onnx’)

mx.mod.Module.export(output_format=‘onnx’, filename_prefix=”model_name”)

...

The whole implementation will go under MXNet repo.

def import(..., input_format='onnx'):
    # returns mxnet graph with parameters
    ...
    return sym, params

def export(..., output_format='onnx'):
  # returns ONNX protobuf object
  ...
  return onnx_proto

...

import nnvm.frontend

def import(..., input_format='onnx'):
  # convert from onnx to nnvm graph
  nnvm_graph, params = nnvm.frontend.from_onnx(...) # Exists
  # convert fron nnvm graph to mxnet graph
  mxnet_graph, params = nnvm.frontend.to_mxnet(...) # Need to implement
  return mxnet_graph, params

def export(..., output_format='onnx'):
  # convert from mxnet to nnvm graph
  nnvm_graph, params = nnvm.frontend.from_mxnet(...) # Exists
  # convert fron nnvm graph to onnx proto format
  onnx_proto = nnvm.frontend.to_onnx(...) # Need to implement
  return onnx_proto

...

# This will internally call serde package API `serde.export`
mx.mod.Module.export(sym, params, output_format=‘onnx’, filename=”model.onnx”)

# OR save into different format using module directly

mx.mod.Module.export(output_format=‘onnx’, filename=”model.onnx”)

...

sym, params = mxnet.mod.Module.import(input_file, ininput_format=‘onnx’)

# create module for inference
mod = mx.mod.Module(symbol=sym, data_names=..., context=mx.cpu(), label_names=None)
mod.bind(for_training=False, data_shapes=..., label_shapes=None)

# set parameters
mod.set_params(arg_params=params, aux_params=None, allow_missing=True)

# forward on the provided data batch
mod.forward(Batch([data_batch]))

...