The goal of this project is to feel more comfortable with PyTorch by practicing image classification on a simple resnet.
My main reference is the original paper. In the paper, the authors develop residual networks with different amounts of residual blocks, of which I chose ResNet18 because it would be good enough for this project.
For the dataset, originally I planned on using ImageNet since that is used in many computer vision papers, but due to hardware constraints I went with CIFAR-10 instead. The dataset contains 50,000 32x32 training images belonging to 10 classes.
By default the model is trained for 90 epochs using the SGD optimizer and a cosine annealing learning rate scheduler. The performance is as follows
| Model | Epochs | Dataset | Accuracy |
|---|---|---|---|
| resnet18 | 90 | CIFAR10 |
- Remove the hard-coded block sizes to allow for larger models (e.g. R50).