Debugging Machine Learning Models
Related:
The main issues with debugging Machine Learning Models are:
- Even if the code is totally correct, things can go wrong (incorrect data, poorly chosen hyper parameters, etc...)
- Training takes a long time, so catching and reproducing a bug is slow.
Debugging tipsβ
- Start with a simpler model (less layers, less data, etc...)
- Add features one step at a time and check each time that performance stays consistent.
- Use data visualisation to inspect behaviour
Common problems, possible causes and solutionsβ
Be sure to know what a Loss function is.
The loss never goes down.β
- Make sure your data can predict the output
- Make sure that the layers of your network are well-suited to your task (using a softmax layer for classification)
- Tweak your hyper-parameters, your learning rate might be too small
The loss quickly oscillatesβ
- Your learning rate might be too high
The loss suddently goes upβ
- Your data might not be homogeneous. Clean your dataset and train with random sampling.
The memory usage is too highβ
- Reduce the batch size.
Source:
Tabnine Link: https://www.tabnine.com/ Whole-line and full-function code completions right in your editor Tabnine uses generative AI technology to predict and suggests your next lines of code based on context & syntax.
- Whole line code completions
- Full-function code completions
- Natural language to code Similar to Captain Stack but instead of using AI, it sends your search query to Google, then retrieves StackOverflow and Github Gist answers and autocompletes them for you