HybridBlock.export parameter naming scheme is incompatible with Block.load_parameters

[haojin2]

Desciption

Gluon will throw exception when trying to load back model weights that were just saved seconds ago

Version

cu102mkl 1.6.0.post0

Minimal Reproduction

from mxnet.gluon import HybridBlock
from mxnet.gluon.nn import Dense
from mxnet import nd

class MLP(HybridBlock):
    def __init__(self, units, prefix=None):
        super(MLP, self).__init__(prefix=prefix)

        with self.name_scope():
            self.layer = Dense(units, activation='sigmoid')

    def hybrid_forward(self, F, x):
        return self.layer(x)

model = MLP(100, prefix=None)

model.initialize()

test = nd.random.uniform(shape=(1, 100))

model.hybridize()
model.forward(test)
nd.waitall()

print(model.collect_params().keys())

model.export("mlp")

model.load_parameters('mlp-0000.params')

Exception & Stack Trace

odict_keys(['mlp0_dense0_weight', 'mlp0_dense0_bias'])
Traceback (most recent call last):
  File "test.py", line 29, in <module>
    model.load_parameters('mlp-0000.params')
  File "/home/ubuntu/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/mxnet/gluon/block.py", line 530, in load_parameters
    cast_dtype=cast_dtype, dtype_source=dtype_source)
  File "/home/ubuntu/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/mxnet/gluon/parameter.py", line 1012, in load
    ignore_extra, restore_prefix, filename, cast_dtype, dtype_source)
  File "/home/ubuntu/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/mxnet/gluon/parameter.py", line 1047, in load_dict
    name[lprefix:], error_str, _brief_print_list(arg_dict.keys()))
AssertionError: Parameter 'dense0_weight' is missing in file: mlp-0000.params, which contains parameters: 'mlp0_mlp0_dense0_weight', 'mlp0_mlp0_dense0_bias'. Please make sure source and target networks have the same prefix.

[leezu]

If you use export, you want to use SymbolBlock.imports to load the model back. If you're just interested in the parameters, you can use save_parameters / load_parameters. In summary, it's expected that your example doesn't work as it mixes the two APIs. It's unfortunate that the two APIs use different naming schemes, but may not change for backwards compat reason.

https://github.com/apache/incubator-mxnet/blob/c76593162c2a5e421243dd09a8c3cde044efaa3f/python/mxnet/gluon/block.py#L1077-L1079