Auto-encoder example out of date

[xidulu]

CNN Autoencoder under the example folder is till using the "old" api:

1 net = gluon.nn.HybridSequential(prefix='autoencoder_')
     2 with net.name_scope():
      3     # Encoder 1x28x28 -> 32x1x1
      4     encoder = gluon.nn.HybridSequential(prefix='encoder_')
      5     with encoder.name_scope():

which would cause:

TypeError: __init__() got an unexpected keyword argument 'prefix'

when using the nightly version of MXNet.

Reference

https://github.com/apache/incubator-mxnet/blob/master/example/autoencoder/convolutional_autoencoder.ipynb

[szha]

I think in addition to fixing the example, we need a mechanism to detect such problems. I started a discussion for managing examples in https://lists.apache.org/thread.html/rf0c95283d86cce60f98943f7d6e3387d6bad180a621bd939a1b2960f%40%3Cdev.mxnet.apache.org%3E

[iadi7ya]

Hi @szha Can I take this up?

[xidulu]

Hi @szha Can I take this up?

You are more than welcome to fix this issue.

[iadi7ya]

Hi @xidulu
There is no argument named as prefix in the HybridSequential class written in file https://github.com/apache/incubator-mxnet/blob/master/python/mxnet/gluon/nn/basic_layers.py

The code in the Auto-encoder example should be changed to:

with net.name_scope():
    # Encoder 1x28x28 -> 32x1x1
    encoder = gluon.nn.HybridSequential()
    with encoder.name_scope():
        encoder.add(
            gluon.nn.Conv2D(channels=4, kernel_size=3, padding=1, strides=(2,2), activation='relu'),
            gluon.nn.BatchNorm(),
            gluon.nn.Conv2D(channels=8, kernel_size=3, padding=1, strides=(2,2), activation='relu'),
            gluon.nn.BatchNorm(),
            gluon.nn.Conv2D(channels=16, kernel_size=3, padding=1, strides=(2,2), activation='relu'),
            gluon.nn.BatchNorm(),
            gluon.nn.Conv2D(channels=32, kernel_size=3, padding=0, strides=(2,2),activation='relu'),
            gluon.nn.BatchNorm()
        )
    decoder = gluon.nn.HybridSequential()

Am I doing it right?

[leezu]

You don't need the name_scope anymore. You can just use

encoder = gluon.nn.HybridSequential()
encoder.add(
            gluon.nn.Conv2D(channels=4, kernel_size=3, padding=1, strides=(2,2), activation='relu'),
            gluon.nn.BatchNorm(),
            gluon.nn.Conv2D(channels=8, kernel_size=3, padding=1, strides=(2,2), activation='relu'),
            gluon.nn.BatchNorm(),
            gluon.nn.Conv2D(channels=16, kernel_size=3, padding=1, strides=(2,2), activation='relu'),
            gluon.nn.BatchNorm(),
            gluon.nn.Conv2D(channels=32, kernel_size=3, padding=0, strides=(2,2),activation='relu'),
            gluon.nn.BatchNorm()
        )
decoder = gluon.nn.HybridSequential()