Exercises

Load a pre-trained ResNet-18 from torchvision and print the architecture.

Replace the classifier head of ResNet-18 for a 5-class task. Count trainable parameters.

Freeze all layers except the last two and verify with requires_grad checks.

Export a simple MLP to ONNX format and verify the exported file loads.

Apply post-training dynamic quantisation (INT8) to a linear model and compare inference time.

Fine-tune a pre-trained ResNet-18 on CIFAR-10 using discriminative learning rates.

Implement LoRA for nn.Linear and apply it to a pre-trained model. Compare parameter counts.

Implement PnP deblurring using a pre-trained DnCNN denoiser.

Compare feature extraction vs full fine-tuning on a small dataset (100 images per class).

Export a CNN to TorchScript via tracing and scripting. Compare outputs.

Implement gradual unfreezing: unfreeze one layer group per epoch during fine-tuning.

Implement PnP channel estimation using DRUNet on the OFDM resource grid.

Implement domain-adversarial training for sim-to-real channel model transfer.

Fine-tune a pre-trained transformer for channel prediction and compare to from-scratch training.

Build a complete deployment pipeline: train, quantise, export to ONNX, benchmark inference.