Resume training hybridized model

I would like to know if there is a way to resume training serialized Hybrid networks (saved as .JSON and .params file). Thanks a lot!

Hi MrDough,

yes, that should be possible.

If you constructed a net using Gluon and then saved it using the export function, you can:

  1. construct the same net using Gluon and import its saved parameters using net.load_parameters().
  2. not construct the net using Gluon, but let MXNet construct it automatically from the JSON and .params file using mx.mod.Module. This is what I do when deploying a model in production.

You can find an example in tutorial [1]. If you have files mymodel-symbol.json and mymodel-0012.params, you can construct the network as follows:

sym, arg_params, aux_params = mx.model.load_checkpoint('mymodel', 12)

# Load the network into an MXNet module and bind the corresponding parameters
mod = mx.mod.Module(symbol=sym, context=mx.cpu())
mod.bind(for_training=False, data_shapes=[('data', (1,3,224,224))])
mod.set_params(arg_params, aux_params)


# Use the forward function for inference with a batch of data.




1 Like