- Create the Dataset object and test if it can load the data correctly
- Create the transformation (in DataLoader) for the Dataset
- Create a model and test if the DataLoader can feed the data to the model
- Create the training procedure
- Check the input and output distribution. Ideally they are as close as possible to normal distribution. If not, then apply transformation to make it close to normal distribution.
- If you can load the dataset in the memory, please do so. It could speed up the training quite significantly.
- I usually check if untrained model gives reasonably random output (e.g. equal chance of being positive or negative).
- Find the model that could overfit the training data, then apply regularization.