Design a Network

In the Github repo, you will find an example of a possible network structure at RVSS2019-WS/on_laptop/steer_net/steerNet.py

This file contains the SteerNet class, which details an example network architecture. Read through the network structure: it uses convolutional layers, rectified linear units, max pooling layers and fully connected layers.

Regression or Classification? You pick!

The example network is a regression network, where a single output is produced - see line 16: self.fc3 = nn.Linear(10,1) , there is only 1 output from the final layer. You could follow this approach and train the network to produce the steering direction, or you could adapt the network to be a classification task, where it gives more than one output (as an example, your network could classify between three classes: 'left', 'right' or 'straight' - this would involve changing line 16: self.fc3 = nn.Linear(10, 3) to allow for 3 outputs from the final layer).

Once you have designed a network, you are ready to train it!