As far as I'm aware it's not possible to create a `Sequential` Block going from
`1DConv` to `RNN` without implementing a custom transpose block. Although
simple for the user to implement, I think this network configuration is common
enough for us to think about finding an easier way.
## Issue
One potential solution would be to select compatible layouts, but `1DConv` only
supports 'NCW' (a.k.a. 'NCT'), and `RNN` only supports 'TNC' and 'NTC'. So they
are incompatible (without a transpose).
Although the following code runs without error, the final shape is incorrect
since incompatible layouts have been used: i.e. 'NCT' with 'TNC'. Would
intuitively expect the output shape to be (8, 1, 3) where T=8, N=1, C=3. But we
get (1, 2, 3).
```
import mxnet as mx
net = mx.gluon.nn.Sequential()
net.add(mx.gluon.nn.Conv1D(channels=2, kernel_size=3))
net.add(mx.gluon.rnn.RNN(hidden_size=3))
net.initialize()
data = mx.ndarray.random.uniform(shape=(1,1,10))
output = net(data)
print(output.shape)
```
## Options/Ideas
1) Support more layout types in `1DConv` or `RNN`. Which could perform the
transposes.
2) Create a 'Transpose' Block, so can add between the `Conv1D` and `RNN`. e.g.
```
net = mx.gluon.nn.Sequential()
net.add(mx.gluon.nn.Conv1D(channels=2, kernel_size=3))
net.add(mx.gluon.nn.Transpose(axes=(2,0,1))
net.add(mx.gluon.rnn.RNN(hidden_size=3))
```
3) Create a generic 'Operator' Block, so can add arbitrary operator in
`Sequential` Blocks. e.g.
```
net = mx.gluon.nn.Sequential()
net.add(mx.gluon.nn.Conv1D(channels=2, kernel_size=3))
net.add(mx.gluon.nn.Operator('transpose', axes=(2,0,1))
net.add(mx.gluon.rnn.RNN(hidden_size=3))
```
[ Full content available at:
https://github.com/apache/incubator-mxnet/issues/12506 ]
This message was relayed via gitbox.apache.org for [email protected]
