PyTorch and Convolutional Neural Networks: Tensors to Transfer Learning

Ever wondered how a computer actually sees an image, and how a few lines of PyTorch can teach it to recognize one? In this episode we go from tensors and autograd to convolution computed by hand, then build, train and improve a real CNN. We cover the PyTorch training loop, datasets and data loaders, GPUs, saving models, debugging and reproducibility. Then we work through convolution with real numbers, stride, padding, the output size formula, channels, pooling and receptive fields, followed by transfer learning, augmentation and visualizing filters. This is episode 27 of the Machine Learning from Zero to Advanced series. Subscribe so you do not miss the next one on sequence models, RNNs, LSTMs and attention. 0:00 Intro 0:55 Tensors and Autograd 3:05 Data and Training Loop 5:16 GPU, Saving and Debugging 7:16 Images as Tensors 8:24 Convolution by Hand 10:50 Stride, Padding and Output Size 12:25 Channels and Pooling 14:49 Building and Training a CNN 16:51 Transfer Learning and Augmentation 18:33 Recap and Next Steps #PyTorch #CNN #DeepLearning #MachineLearning #ComputerVision