4.4 Defining Efficient Data Loaders (Part 1-4)

References

• Dataset API documentation
• DataLoader API documentation

Code

• Part 3: Data download & preparation, 4.4-dataloaders-part3-download-and-prep.ipynb
• Part 4: Run dataloaders, 4.4-dataloaders-part4-define-and-run.py

What we covered in this video lecture

Now that we are beginning to work with larger neural networks and datasets, it’s time to worry about optimizing our data-loading pipeline. In this lecture, we discussed PyTorch’s Dataset and DataLoader classes, which help us fetch the training batches in the background using multiple processes. To avoid computational bottlenecks, using these background processes ensures that the next batch of data is ready when the model finishes its backward pass.

Additional resources if you want to learn more

Using the Dataset and DataLoader classes is the most convenient way to load data in PyTorch. However, there is also a new DataPipe class that the PyTorch team is developing. You can find out more about these optional DataPipes in the official TorchData repository here. I also wrote a tutorial about DataPipes which you might find helpful here.

Watch Video 1 Mark complete and go to Unit 4.5 →