I haved downloaded a Graph convolution neural network (GCNN) code from the github, on which the authors extend the traditional CNN to graph data and implement a user-defined convolution/pooling layer, see github ).
I would like to use this model on my problem, but this code is implemented as a classifier. How should I change it for a regression problem?
Best Answer
It depends on what possible values your regression can take, but likely you want to change the activation of the final layer from what it is now (likely softmax or sigmoid) to linear (no activation function).
(I say it depends because if your values are non-negative, you could use ReLU, if they are bounded on both sides you may consider using a scaled sigmoid or tang, etc.)
You will also need to update the loss function. It is likely cross entropy, usually RMSE is what you want for regressions.
Similar Posts:
- Solved – Number of trainable parameters in Convolution models (Keras)
- Solved – Softmax in multi-class in deep NNs
- Solved – Why zero-centered output affected the backpropagation
- Solved – what makes neural networks a nonlinear classification model
- Solved – Sigmoid activation hurts training a NN on pyTorch