📄️ PyTorch
Deep learning with PyTorch, from raw tensors to a served model. Duration: 9h, with an applied Fashion-MNIST project, a 40-question exam and a verifiable certificate.
📄️ 1. PyTorch tensors
Module 1 of the PyTorch premium course: create tensors, master shape, dtype and device, share memory with NumPy without silent copies, and dodge the in-place trap that breaks gradients.
📄️ 2. Autograd
Module 2 of the PyTorch premium course: how requires_grad, backward, zero_grad, no_grad and detach fit together, and why forgetting to zero the gradients silently ruins training.
📄️ 3. nn.Module
Module 3 of the PyTorch premium course: how nn.Module organises parameters, sub-modules and forward code, when to prefer nn.Sequential and how state_dict serialises everything.
📄️ 4. Dataset and DataLoader
Module 4 of the PyTorch premium course: implement a custom Dataset, configure DataLoader for batching, shuffling and workers, and avoid the normalisation leak between train and validation.
📄️ 5. Training loop
Module 5 of the PyTorch premium course: the five canonical steps of a PyTorch loop, why model.eval() and torch.no_grad() are both needed and how to detect overfitting from the metric curves.
📄️ 6. Optimizers and schedulers
Module 6 of the PyTorch premium course: choose between SGD, Adam and AdamW, tune weight decay, apply StepLR, CosineAnnealingLR, OneCycleLR and ReduceLROnPlateau without stepping on your own gradients.
📄️ 7. GPU and mixed precision
Module 7 of the PyTorch premium course: move a model and its data to a GPU, avoid transfer bottlenecks and enable mixed precision with autocast and GradScaler.
📄️ 8. Checkpoints and resume
Module 8 of the PyTorch premium course: save model, optimiser, scheduler and RNG state to resume exactly where a run left off, and choose between best epoch and last epoch.
📄️ 9. Transfer learning
Module 9 of the PyTorch premium course: load a pretrained ResNet18, freeze the backbone, replace the head, then fine-tune with differentiated learning rates without breaking normalisation.
📄️ 10. TorchScript, ONNX and serving
Module 10 of the PyTorch premium course: export a trained model with torch.jit.trace or script, generate an ONNX file, verify numerical parity and stand up a minimal serving endpoint.
📄️ 11. Recap and exam
Complete recap of the PyTorch premium course: tensors, autograd, nn.Module, DataLoader, training loop, optimisers, GPU, checkpoints, transfer learning, TorchScript and ONNX, then the 40-question exam.