Skip to main content

Backbones

Backbones are also referred to as Feature Extractors in a 2D Convolutional Neural Network. At GQC, we have explored different backbones ranging from ResNet18, 34, 50 and 101.

You can read about the backbones over here

  1. ResNet18
  2. ResNet34
  3. ResNet50
  4. ResNet101

The number after the word ResNet refers to number of layers present in the network. Greater the number of layers, more the number of parameters in a network.

info

You can use any pretrained model from the torchvision models.

There is a lot of scope for exploring bigger and better backbones, but they would require a lot of GPU memory. Over 800 pretrained models are available on timm which is a library for training, fine-tuning, and evaluating image models.

You need to add some wrappers around the model so that you can use them with the fastai library.