100+ Free PyTorch Certified Associate Practice Questions

PyTorch fundamentals

Create and reshape tensors, run tensor operations with broadcasting, track gradients with autograd and requires_grad, use torch.no_grad and detach, and write device-agnostic code that moves tensors between CPU and CUDA.

Building neural networks

Subclass nn.Module and define forward(), compose layers like nn.Linear and nn.Conv2d, apply activations such as ReLU and Softmax, choose losses like CrossEntropyLoss and MSELoss, and configure optimizers such as SGD and Adam with learning rate and weight decay.

Data handling

Implement custom Dataset classes with __len__ and __getitem__, batch and shuffle with DataLoader, apply torchvision transforms including ToTensor and Normalize, and customize collate_fn, num_workers, and pin_memory for efficient loading.

Training and evaluation loops

Run the zero_grad, backward, step cycle in the correct order, switch between model.train() and model.eval(), evaluate under torch.no_grad(), and compute metrics such as running loss and accuracy with argmax.

Model saving, loading, deployment, and inference

Save and load models with state_dict, build resumable checkpoints, map storages across devices with map_location, compile to TorchScript using trace versus script, and export models to ONNX for cross-runtime inference.

Debugging and best practices

Diagnose vanishing, exploding, and NaN gradients with anomaly detection and gradient clipping, fix device-mismatch and CUDA out-of-memory errors, seed all RNGs for reproducibility, and regularize against overfitting.