This module introduces custom architectures that go beyond Sequential models, showing how PyTorch’s dynamic graphs support multi-input/multi-output design, parameter sharing, conditional execution, and dynamic creation. You’ll build Siamese Networks, ResNet, and DenseNet to see how architectural choices solve real challenges like similarity comparison, vanishing gradients, and information reuse.

This module explores specialized vision approaches in PyTorch, starting with how receptive fields grow in CNNs and moving into interpretability tools like saliency maps and Grad-CAM to reveal what drives model predictions. You’ll then dive into generative models, using diffusion techniques with Hugging Face’s diffusers library and Stable Diffusion to create images while experimenting with parameters that shape the output.

This module demystifies transformer architectures by showing how modern NLP models are built from familiar PyTorch components like linear layers, embeddings, and attention. You’ll explore encoder-only, decoder-only, and encoder-decoder designs step by step, learning how attention, positional encoding, and cross-attention make these models so powerful for tasks from classification to translation.

This module bridges the gap between training models and deploying them in the real world, covering how to save, track, and manage experiments with PyTorch serialization and MLflow. You’ll then make models portable with ONNX and optimize them for production using pruning and quantization techniques that shrink size and boost speed without losing accuracy.