Q1: In the Week3 assignment, it mentioned the data we practised is well pre-processed applied on training. What kind of pre-processing does it require?
(Since I am working on a related MRI segmentation project and have raw MRI data to use.)
Q2: In Assignment Section 4.2, we loaded a pre-trained model for later works.
I would like to know more about the pre-training details and methods for my own models, hope you can provide some knowledge of it.
It may be a little bit difficult to answer clearly to your question, but I mention here my opinion.
I’m glad if you can use it as a reference.
- We can suppose various cases. For example, image resolution or machine resolution is assumed. In addition, the condition of the light source also has an effect.
When the deviation between the training and test data, the model may NOT work well.
- Assuming the same model structure and training data, similar results may be reproduced by training again with the model parameters initialized.
As for the parameter optimization method, the famous method is to try Adam or SGD.