Deep Learning with PyTorch

Simple holdout validation

Set apart a fraction of the data as your test dataset. What fraction to keep may be very problem-specific and could largely depend on the amount of data available. For problems particularly in the fields of computer vision and NLP, collecting labeled data could be very expensive, so to hold out a large fraction of 30% may make it difficult for the algorithm to learn, as it will have less data to train on. So, depending on the data availability, choose the fraction of it wisely. Once the test data is split, keep it apart until you freeze the algorithm and its hyper parameters. For choosing the best hyper parameters for the problem, choose a separate validation dataset. To avoid overfitting, we generally divide available ...