Are you ready to lift the veil from your broken PyTorch model and prevent it from ever breaking again? This presentation covers strategies to
- Create synthetic data for your custom ML model and
- Setup an adequate test suite to speed up your ML dev process
To run the streamlit app, follow the steps below:
Activate the poetry environment
poetry shell
and run the presentation
streamlit run Home.py
The presentation should now be available in a new browser tab
All PyTorch code is contained in pages/torch_examples
- Jeremy Jordan's blogpost about testing philosophy for ML
- Testing for PyTorch with torchcheck
- Deprecated, but interesting test suites for < TF2.0: mltest and for PyTorch: torchtest
- Differences between Software Engineering and Machine Learning Engineering workflows at TensorFlow explained by TF team lead
- Weight analysis without training or validation data: weightwatcher
- Post-training label quality analysis: cleanlab
- Behavioral testing for NLP models
- Andrej Kaparthy's recipe for training NNs
- and if you're done debugging check out Google's Deep Learning tuning playbook