Roadmap For ML Models
Data Curation,Preprocessing and Distribution
A very important step to get the best accuracy from the model. We must have clean data or we can make efforts to create clean data.
Preprocessing is a required step for a model to learn features that suits the algorithms in terms of sizes and dimensions.
Distribution is again plays important role in performance for classification problems. If the distribution is not equal then the role of class weights, sampling, or more data collection comes.
Transfer Learning
Transfer learning has always proved a boost in accuracy for my model. Playing with freeze or unfreeze is another part but yes using a pre-trained model and some more layers is always a good practice and we must take advantage of weights too. So using a pre-trained model is highly recommended.
Avoid Overfitting
Have many things to try with your model, Play with regularization, data augmentation as it upsamples your data. Hyperparameters tuning can save you from overfitting. Overfitting is always the main problem that many of us have been stuck but we can find a way out just to have chains of experiments.
Different Performance Metrics
Never stop with accuracy as the only important performance metrics, precision-recall, and f1 score are holding much importance for you to tell which thing is getting wrong. If you analyze this thing you will have better inference.
Mistakes Analysis
‣ Look at those mistakes made by models with their confidence.
‣ Manually review those mistakes and correct their label or curate those if required
‣ Retrain the model and analyze it again
