The whole of deep learning on one page, every line explained: tensors and autograd, models as classes, activations and losses, the training loop, initialisation and normalisation, regularisation, CNNs, RNNs, attention, metrics, tuning and deployment.
Deep learning is a small number of ideas surrounded by a large number of details that fail silently. The maths of a layer is a matrix product, a bias and a nonlinearity, and backpropagation is the chain rule organised well. What actually costs people days is everything around that: a tensor on the wrong device, a missing zero_grad() so every step carries the last one’s gradients, model.eval() forgotten before validation so dropout is still firing, or a softmax added by hand on top of a loss function that already applies one. None of those raise an error. They just make the numbers worse.
This sheet puts twenty-three topics side by side, each with the handful of lines that matter and a plain explanation of what every one of them does, including a card for reading the training curves and one for the debugging moves that actually work: overfit a single batch, print shapes between layers, check the loss at initialisation against the value chance would give. It is wide rather than long by design, so the whole subject stays in one view. Keep it open while you build, or save the image and pin it nearby.
Hope this helps.
