MP6: RNN

This MP will use pytorch, which is an extremely powerful and flexible neural net training package.

NOTE EXTENDED DEADLINE: Due date is Wednesday, 12/8, 11:59pm.

• Pytorch is intentionally easy to learn, but even so, it’s worth spending some time with the pytorch tutorial.

In this MP, you will train an LSTM voice conversion system.

• mp6.zip contains the code. You should download the code, unzip it into some directory, and then type jupyter-lab to get started.

• mp6_notebook.html is an example of what the Jupyter notebook will look like, once you’ve finished everything.

• The MP uses a lot of torch-specific functions. You can find documentation for all of those functions here.

  • Your LineEar.__init__ method will use torch.nn.parameter.Parameter.

  • Your LineEar.forward method will use torch.matmul.

  • Your module EllEssTeeEmm will put torch.nn.LSTMCells. into two torch.nn.ModuleLists – one called self.forward_layers, one called self.reverse_layers.

  • Your module GeeArrYou will put torch.nn.GRUCells. into two torch.nn.ModuleLists – one called self.forward_layers, one called self.reverse_layers.

  • Your Encoder.__init__, Decoder.__init__, and Postnet.__init__ will each put Conv1d, BatchNorm1d, and ReLU layers into a torch.nn.ModuleLists – called self.convolutions. They will then create an EllEssTeeEmm module called self.recurrent, and a LineEar called self.fc_projection.

• python grade.py is how you will grade your code on your local machine.

• When it works on your machine, then you can try uploading only the file submitted.py to the MP4 assignment on Gradescope.