Typorithm learning path
PyTorch: tensors to a trained model
Tensors, broadcasting, storage, autograd, registered modules, logits and losses, loaders, SGD, evaluation, train-only preprocessing, metrics, checkpoints, convolution, embeddings, torch.export and a held-out classifier capstone. Six sections per lesson, retained exercise baselines, two hints and small gradual tasks. Verified with Python 3.13 and PyTorch 2.15.0. Reading and reasoning checks work in the app; real framework execution uses the downloadable computer lab. CPU only, owned fixtures, no external datasets or production feeds.
Study with a practical destination
Use the linked lessons in order, complete the lab and explain the decisions without the reference. Select the path and depth that match your role.
PyTorch: a complete tiny CPU training experiment
Before starting: Python functions and classes plus basic arithmetic and a willingness to name tensor axes. Reading and reasoning checks work in the app; framework execution uses the isolated pinned computer lab.
Build and explain: Explain and verify a small training pipeline with held-out evaluation, safe owned checkpoints and explicit export guards.
- 1. PyTorch: tensors, shapes, dtypes and explicit CPU execution
- 2. PyTorch: broadcasting and reduction axes
- 3. PyTorch: views, clones and shared storage
- 4. PyTorch: autograd and deliberate gradient accumulation
- 5. PyTorch: nn.Module and registered parameters
- 6. PyTorch: logits, class labels and cross-entropy
- 7. PyTorch: datasets, loaders and complete batches
- 8. PyTorch: the zero-grad, forward, backward and step sequence
- 9. PyTorch: eval mode and gradient-disabled inference
- 10. PyTorch: train-only preprocessing and leakage prevention
- 11. PyTorch: confusion counts and imbalanced evaluation
- 12. PyTorch: state_dict checkpoints and explicit safe loading
- 13. PyTorch: convolution shapes and preserving the batch axis
- 14. PyTorch: embeddings, padding and masked sequence pooling
- 15. PyTorch: torch.export and explicit dynamic shape contracts
- 16. PyTorch capstone: a bounded classifier with held-out evaluation
Review evidence: Attempt the retained small exercise before revealing its solution. Use the two hints, compare observed framework state and run the positive and negative cases. Optional stretches do not block the core. Local fixtures do not establish production readiness or professional certification.
Your learning path
0 of 16 lessons completed
Saved on this device
Module 1 · 1 lessons
PyTorch: tensors, shapes, dtypes and explicit CPU execution
Read a tensor’s structural contract before writing a model.
Module 2 · 1 lessons
PyTorch: broadcasting and reduction axes
Center feature columns without accidentally mixing examples and features.
Module 3 · 1 lessons
PyTorch: views, clones and shared storage
Choose intentionally whether an operation shares or owns tensor storage.
Module 4 · 1 lessons
PyTorch: autograd and deliberate gradient accumulation
Build a fresh graph and reset gradients when an update should be independent.
Module 5 · 1 lessons
PyTorch: nn.Module and registered parameters
Make layers visible to optimizers, checkpoints and device moves.
Module 6 · 1 lessons
PyTorch: logits, class labels and cross-entropy
Match output and target contracts before interpreting a loss value.
Module 7 · 1 lessons
PyTorch: datasets, loaders and complete batches
Preserve all examples and distinguish reproducible order from shuffled training.
Module 8 · 1 lessons
PyTorch: the zero-grad, forward, backward and step sequence
Observe a real SGD parameter update with an analytic one-example target.
Module 9 · 1 lessons
PyTorch: eval mode and gradient-disabled inference
Use both model behavior mode and an appropriate inference context.
Module 10 · 1 lessons
PyTorch: train-only preprocessing and leakage prevention
Fit normalization on training data and apply it unchanged to validation.
Module 11 · 1 lessons
PyTorch: confusion counts and imbalanced evaluation
Inspect class-specific failures instead of trusting accuracy alone.
Module 12 · 1 lessons
PyTorch: state_dict checkpoints and explicit safe loading
Restore owned weights into a matching architecture without loading arbitrary objects.
Module 13 · 1 lessons
PyTorch: convolution shapes and preserving the batch axis
Trace a small NCHW image through a real convolution and flatten operation.
Module 14 · 1 lessons
PyTorch: embeddings, padding and masked sequence pooling
Average real tokens without letting padding change the representation.
Module 15 · 1 lessons
PyTorch: torch.export and explicit dynamic shape contracts
Distinguish an exported program’s input guards from a normal eager call.
Module 16 · 1 lessons
PyTorch capstone: a bounded classifier with held-out evaluation
Connect a complete tiny training loop to a separate observed prediction set.