Training
... TODO ...
See README for now.
The idea here was to train the same architecture on classification as we did on the RL task. This was used to compare RFs with previous literature (Lindsey et al. 2019) and also to test attribution algorithms (see Analysis page). So basically here the network is initialized and trained from scratch. The architectures are exactly the same if the purely feedforward network is used in the RL task, the only difference being the output layer (6 + 1 for RL/actor+critic and 10 for classification - 1 for each digit). To maintain the same number of parameters in the FC layer, the digits are also padded with a 'doom-like' background, but the actual background shouldn't really matter at least in the MNIST case.
This is all currently implemented in the class_encoder_train.ipynb notebook. To get optimal performance and nicer-looking receptive fields it is also better to increase the number of epochs (currently 30) to somewhere around 100-500. At the end the notebook saves the trained model in pytorch/models/name_of_the_model.npy, which can then be accessed by the analysis scripts.
Accessing nodes
Building a container
Running TensorBoard
WandB
MAPINFO
DECORATE
TEXTURE1 and PNAMES
MAP01 and TEXTMAP
SCRIPTS and BEHAVIOUR
Preparing dataset
Setting offsets
Renaming and DECORATE
Modifying SCRIPTS
Importing to SLADE
Testing
Conv layers
FC layer
RNN layer
Attribution analysis (simulation)
Visualisation of encoder output
Classification (decoding) analysis
Attribution analysis (dataset)
Action analysis
