Save / Load model (Python)

[abieler]

After training my feed forward model I want to save it to disk so I can load it
back up at a later point. The way I do it seems rather cumbersome, but I could
not figure out a better way. Any Advice would be appreciated.

# train_iter and test_iter are NDArrayIter and taken care
# of before this code snippet

# model definition
data = mx.sym.var('data')

fc1 = mx.sym.FullyConnected(data=data, num_hidden=10)
relu1 = mx.sym.Activation(data=fc1, act_type="relu")

fc2 = mx.sym.FullyConnected(data=relu1, num_hidden=12)
relu2 = mx.sym.Activation(data=fc2, act_type="relu")

yhat = mx.sym.FullyConnected(data=relu1, num_hidden=n_classes)
prediction = mx.sym.SoftmaxOutput(data=yhat, name="softmax")

model = mx.mod.Module(symbol=prediction, context=mx.cpu())
model.fit(train_iter,
          eval_data=test_iter,
          optimizer="sgd",
          optimizer_params={'learning_rate': 0.1},
          eval_metric="acc",
          batch_end_callback = mx.callback.Speedometer(BATCH_SIZE,300),
          num_epoch=50)

model.save_checkpoint("myfile", 50)

#now assume this is a new session where I want to load the model from before
sym, arg_params, aux_params = mx.model.load_checkpoint("myfile", 50)
new_model = mx.mod.Module(symbol=sym)
new_model.bind(train_iter.provide_data, train_iter.provide_data)
new_model.set_params(arg_params, aux_params)

This all works, but there is the parameter epoch which has to be specified for loading the
files. Is there a way independent on epoch to save/load models?
I can think of two obvious but "hacky" solutions:
a) Save with a fixed number for epoch that actually does not reflect the actual epoch of training.
b) Add an additional model.save_params() after the save_checkpoint()

Environment info

Package used: Python

MXNet version: 0.10.0

[CNevd]

try write your own save and load function:

def save(mod, prefix):
    mod._symbol.save('%s-symbol.json' % prefix)
    mod.save_params('%s.params' % prefix)

def load(prefix):
    symbol = mx.sym.load('%s-symbol.json' % prefix)
    save_dict = mx.nd.load('%s.params' % prefix)
    arg_params = {}
    aux_params = {}
    for k, v in save_dict.items():
        tp, name = k.split(':', 1)
        print name
        if tp == 'arg':
            arg_params[name] = v
        if tp == 'aux':
            aux_params[name] = v
    return (symbol, arg_params, aux_params)

[szha]

This issue is closed due to lack of activity in the last 90 days. Feel free to ping me to reopen if this is still an active issue. Thanks!