PyTorch

Deep learning from tensors to fine-tuning. You practise the maths in NumPy and the API in quizzes. 16 modules of 10 minutes each, with questions that check your answers in the browser.

Start the track

Foundations

• 1. Tensors: creation and basicsFree
  Meet the tensor, PyTorch's array, create tensors in the common ways, and read their shape and dtype, with NumPy practice that runs right here.
• 2. Tensor operations and broadcastingFree
  Do element-wise maths, matrix multiplication and reductions along dimensions on tensors, with the same broadcasting rules as NumPy and a few PyTorch-specific habits.
• 3. NumPy ↔ tensors; devices (CPU/GPU)Free
  Move data between NumPy, pandas and PyTorch without surprises, understand shared memory, and put tensors and models on the right device.
• 4. Autograd: how gradients workFree
  See how PyTorch records operations and computes gradients with backward(), why gradients accumulate, when to switch tracking off, and check gradients numerically in NumPy.

Intermediate

• 5. Building a model with nn.ModuleFree
  Define networks as nn.Module classes with layers and a forward method, understand what a Linear layer computes, add activations, and count parameters.
• 6. Loss functions and optimizersFree
  Pick the right loss for regression, binary and multi-class problems, understand logits and BCEWithLogitsLoss, and let an optimizer like SGD or Adam update the weights.
• 7. The training loop, step by stepFree
  Write a complete training loop with epochs, mini-batches, a validation pass in eval mode and loss logging, and know what each line is for.
• 8. Datasets and DataLoadersFree
  Wrap data in a Dataset, let a DataLoader handle batching and shuffling, and write the cleaner training loop that every PyTorch project uses.

Advanced

• 9. A full example: regression from scratchPro
  Build a complete PyTorch regression project for coffee demand, from data preparation through training, evaluation against baselines, and prediction for new days.
• 10. A full example: classificationPro
  Train a PyTorch churn classifier end to end with logits and BCEWithLogitsLoss, turn outputs into probabilities, choose a threshold, and evaluate with AUC and a confusion matrix.
• 11. Saving and loading modelsPro
  Save a model's state_dict, write full training checkpoints, load them back correctly, and store the preprocessing and config that predictions depend on.
• 12. Reducing overfitting (dropout, early stopping)Pro
  Keep networks from memorising the training data with weight decay, dropout and early stopping, and see each technique's effect on validation loss.

Expert

• 13. Custom datasets and transformsPro
  Write Datasets that build examples on demand (sliding windows, files read lazily) and transforms that preprocess or augment each example, keeping training and evaluation behaviour separate.
• 14. Structuring a clean training projectPro
  Split a training script into config, data, model, engine and entry-point modules, with seeding, command-line options, checkpoints and tests, so experiments are reproducible and easy to change.
• 15. Fine-tuning a pretrained modelPro
  Reuse a network trained on a large task for a small one by freezing its body, replacing its head, and optionally fine-tuning all layers with a small learning rate.
• 16. Debugging training (common pitfalls)Pro
  Recognise and fix the bugs that make training silently fail, including shape broadcasting in the loss, forgotten zero_grad or eval, exploding losses and unscaled inputs, and use the overfit-one-batch check.

Track project

• PyTorch Project: Toko Kita's demand network, by handPro
  Three graded tasks that rebuild what PyTorch does for you, in NumPy. Run a small network forward and measure its loss, train with early stopping that keeps the best weights, and debug a training script that has four classic bugs.

Pro modules unlock with one payment of Rp 149.000, yours to keep.