shows a problem with exporting then importing a model in gluon

gluon-export-model

import mxnet as mx
from mxnet.test_utils import download
from mxnet.gluon.model_zoo import vision as models
gpus = 0
contexts = [mx.gpu(i) for i in range(gpus)] if gpus > 0 else [mx.cpu()]

deep_dog_net = models.squeezenet1_1(prefix='deep_dog_', classes=2)
deep_dog_net.collect_params().initialize(ctx=contexts)

download('https://apache-mxnet.s3-accelerate.amazonaws.com/gluon/models/deep-dog-5a342a6f.params',
         overwrite=True)

F = mx.nd.random.uniform(shape=(1,3,224,224), ctx=contexts[0])
deep_dog_net.hybridize()
deep_dog_net(F)
deep_dog_net.load_params('deep-dog-5a342a6f.params', contexts)

# this generates nothotdog-0000.params and nothotdog-symbol.json
deep_dog_net.export('nothotdog')

# then loading the exported params file will fail with
# Parameter conv0_weight is missing in file nothotdog-0000.params
deep_dog_net.load_params('nothotdog-0000.params', contexts)

# try to allow missing, but fails with
# Parameter squeezenet1_conv0_weight loaded from file nothotdog-0000.params is not present in ParameterDict
deep_dog_net.load_params('nothotdog-0000.params', ctx=contexts, allow_missing=True)

# add ignore_extra=True and it "works" but I image the network is not really loaded properly
deep_dog_net.load_params('nothotdog-0000.params', ctx=contexts, allow_missing=True, ignore_extra=True)

When originally fetching the model from the zoo, if you don't set a prefix, you'll get the default "squeezenet1_", and this will fail too when you try to load the exported params file.

It seems that the load_params isn't using the prefix and isn't expecting the network definition to have the prefix.

If you set prefix to be blank, things seem to start working.

deep_dog_net = models.squeezenet1_1(prefix='', classes=2)

The exported params and symbols file now have the prefix now, and now the load_params call will work.