Player not loading? Watch on YouTube
This live PyTorch session introduces TensorBoard through image logging. The first part explores pooling, batch normalization and dropout with small Python experiments and frequent audience questions. TensorBoard work begins around 1:30:09, after these exercises; training scalar plots are left for the next session.
The speaker imports SummaryWriter from torch.utils.tensorboard, creates an experiment under runs/exp1 and writes a grid of Fashion-MNIST images. He launches TensorBoard with the runs directory as its log source and opens the browser interface. Adding another image grid demonstrates that the dashboard can update while its server remains running.
The exercise then switches to the Oxford pet dataset. Images have different dimensions, causing a batching error. Reducing the batch size to one lets the example continue. Attempts to show multiple cats and dogs lead to more debugging: the speaker removes old experiment logs, checks the iteration and gives each image a unique label. The final dashboard displays multiple images.
This is a first-use walkthrough with mistakes and corrections, rather than a finished training dashboard. The speaker reports DataLoader worker trouble on his Mac; that is his setup experience. Casual references to Gemini, OpenRouter and Claude Code do not form tutorials for those products.
