Closed PythEsc closed 7 years ago
Would be the best if Bruno would also push his RNN code, that he AND me can compare the code
I aligned the two. I pushed the RNN and the changes I did for both.
@brunolubascher @Naxter Is it currently possible to train the network and use the training progress later for the prediction of a single post?
So basically the networks should support a way to simply call a function
def predict(post: str) -> dict
which gets a single post and returns the predicted reactions based on the former trained model
For the CNN this works with the saved checkpoints that are written while training. The function can be easily written. Maybe we provide a base class so it is easier for the API to call the function for both networks. Also, the modularity is better and easier extendable.
This is a really good idea. I've created a base class called "NeuralNetwork". Moreover, I've moved both directories ("convolutional_neural_network" and "recurrent_neural_network") to a base directory ("neural_network"). In this base directory we'll add all files that are used by both networks. Everything specific for one or the other network will be added to its sub-directory
Ok I have started to adjust the RNN to the base class. I think it would be nice to have the possibility to save and restore the models so that one can continue training at a later point. Moreover, this is necessary for the predict-method anyways. So basically this is still open:
CNN prediciton method is implemented
Is aligned.
We should make sure that the CNN and RNN actually are using the same structure of input and output data. Moreover, they should use the same internal methods (e.g. for calculating the accuracy)