Complete PyTorch projects¶
Choose the workflow you want to run. Each page explains the important steps before showing the complete maintained script.
Selected Course 2 workflows¶
Three examples combine useful ideas across the course rather than reproduce each lab.
Controlled training comparison
Metrics, learning-rate choice, plateau scheduling, gradient accumulation
Run / output: Small synthetic classification problem on CPU; saves a comparison report.
Image augmentation and head training
Training transforms, impulse noise, frozen ResNet backbone, replacement head
Run / output: Synthetic images with random weights; no downloads. Optional own images + pretrained weights.
Variable-length text classifier
Training-only vocabulary, offsets, mean pooling, class weights
Run / output: Tiny original phrases on CPU; saves predictions and the model contract.
Foundation projects¶
Linear versus nonlinear representation, MSE, autograd
Run / output: Retained prediction chart and validation MSE.
CNN shapes, classification, evaluation
Run / output: Predictions, learning curves and a confusion matrix.
Stable class IDs, readable files, rejected samples
Run / output: Validated manifest and diagnostics.
Reusable blocks, regularization, overfitting
Run / output: Training versus validation curves.
Running and interpreting the examples¶
Commands run from the repository root after preparing the documented project dependencies. On Windows, use py -3.14 in place of python if that is your installed launcher.
The three new workflows use CPU and small offline defaults. The original image projects use --device auto, selecting CUDA when available and otherwise CPU; they may download public TorchVision data. Their optional --full --device cuda mode increases the run budget.
What a small run proves
Synthetic images, toy phrases and smoke tests verify the code path and make the mechanics visible. They do not establish real dataset quality, pretrained transfer performance or an efficiency benchmark. A measured claim needs retained evidence from the corresponding run.