Tensors and shapes
For images, read [batch, channels, height, width]. Shape, dtype, and device are part of the program.
PyTorch · start here
Build your first image model.
Start with the three things that make PyTorch click: tensor shapes, one complete training loop, and how a CNN reads an image. This page is the short version; the detailed reference shows the why and the runnable code.
tensor.inspect
shape [32, 3, 224, 224]
dtype torch.float32
device cuda:0
requires_grad false
dtype torch.float32
device cuda:0
requires_grad false
01 / START HERE
The five steps every model repeats.
A model learns by seeing a batch, making logits, measuring error, calculating gradients, and updating its weights. Select a stage to see its job in plain language.
02 / CORE CONCEPTS
What to understand before changing code.
You do not need to memorize every PyTorch function. You do need to recognise what enters a model, what leaves it, where learning happens, and how to check whether it worked.
Samples become batches
Dataset returns one prepared example. DataLoader groups examples into batches and shuffles training order.
Layers create logits
nn.Module turns a batch into raw class scores. Activations make a network capable of learning non-linear patterns.
Loss changes weights
Clear gradients, predict, measure the loss, calculate gradients, then update the parameters.
Read the training loop ↗Evaluation is different
Use model.eval() and torch.no_grad() before measuring accuracy or inspecting errors.
CNNs learn local patterns
Convolutions create feature maps. Pooling reduces spatial size. Validation curves reveal whether those features generalize.
See CNNs visually ↗
Image data path. The model receives a batch tensor, not an image file.
Convolution. A small filter creates a feature map; a CNN learns useful filters from data.
CNN shapes. Channels can grow while pooling reduces height and width.
Why activations matter. This retained result shows a line cannot fit every pattern.
training_loop.py
for images, labels in train_loader:
images, labels = images.to(device), labels.to(device)
optimizer.zero_grad() # clear old gradients
logits = model(images) # forward pass
loss = loss_fn(logits, labels) # measure error
loss.backward() # calculate gradients
optimizer.step() # update parameters03 / VISUAL TOOLS
Change an input and see what changes.
These small tools make shapes, batches, broadcasting, and generalization visible without running a model or storing anything in your browser.
Batch calculator
CNN shape tracer
Broadcasting checker
Generalization curves
04 / RUN EXAMPLES
See each idea in a complete script.
Each example downloads public data when needed, trains a real model, and writes its own predictions, curves, metrics, and checkpoint. Start small on CPU; use the longer GPU mode when you want it.
Linear vs nonlinear features
A deterministic curved function exposes what a linear model cannot represent and why activation functions change capacity.
TensorMSEAutogradAdam
Open project ↗EMNIST letter classifier
Downloads real letters, trains a CNN, and saves predictions, a confusion matrix, curves, metrics, and a checkpoint—small on CPU, longer on GPU.
EMNISTDownloadTrainArtifacts
Open project ↗Robust image pipeline
Downloads public images, builds a checked folder dataset, records a corrupt file, then trains only on valid samples and saves diagnostics.
FoldersValidationTrainReport
Open project ↗Nature CNN and overfitting
Downloads selected CIFAR-100 classes, trains a regularized CNN, and saves predictions, curves, a confusion matrix, metrics, and a checkpoint.
CIFAR-100DownloadCNNArtifacts
Open project ↗Detailed reference
Need the next explanation?
Continue through two connected guides, jump to a specific section, or copy a complete maintained project script.
Open the detailed Fundamentals guide ↗