Enum ResNetVariant
Defines the available ResNet (Residual Network) architecture variants.
public enum ResNetVariantFields
ResNet101 = 3ResNet-101: 101 weight layers using BottleneckBlock. For complex tasks requiring more capacity.
Architecture: conv1 + [3, 4, 23, 3] BottleneckBlocks + FC = 101 layers
For Beginners: ResNet101 has approximately 44.5 million parameters and provides more capacity than ResNet50. The additional depth (especially in the third stage with 23 blocks) allows it to learn more complex patterns. Use this when ResNet50 underfits your data or when you're working with very complex visual recognition tasks. Requires more GPU memory and longer training times than ResNet50.
ResNet152 = 4ResNet-152: 152 weight layers using BottleneckBlock. Maximum capacity ResNet variant.
Architecture: conv1 + [3, 8, 36, 3] BottleneckBlocks + FC = 152 layers
For Beginners: ResNet152 is the deepest standard ResNet variant with approximately 60.2 million parameters. It significantly increases depth in the second and third stages. This model is typically used for competition-level performance on very large and complex datasets like ImageNet. It requires substantial computational resources and is often slower to train and run inference than smaller variants. For most practical applications, ResNet50 or ResNet101 provide better speed/accuracy tradeoffs.
ResNet18 = 0ResNet-18: 18 weight layers using BasicBlock. Suitable for smaller datasets.
Architecture: conv1 + [2, 2, 2, 2] BasicBlocks + FC = 18 layers
For Beginners: ResNet18 is the smallest ResNet variant with approximately 11.7 million parameters. It's a good choice for smaller datasets, limited computational resources, or when you need faster inference times. Despite being relatively shallow, it still benefits from residual connections and typically outperforms VGG networks with similar parameter counts.
ResNet34 = 1ResNet-34: 34 weight layers using BasicBlock. Good balance of depth and efficiency.
Architecture: conv1 + [3, 4, 6, 3] BasicBlocks + FC = 34 layers
For Beginners: ResNet34 has approximately 21.8 million parameters and provides a good balance between model capacity and training efficiency. It uses the same BasicBlock structure as ResNet18 but with more blocks, allowing it to learn more complex features. Often used when ResNet18 underfits but computational resources are limited.
ResNet50 = 2ResNet-50: 50 weight layers using BottleneckBlock. The most commonly used ResNet variant.
Architecture: conv1 + [3, 4, 6, 3] BottleneckBlocks + FC = 50 layers
For Beginners: ResNet50 is the most popular ResNet variant with approximately 25.6 million parameters. It switches to BottleneckBlocks (1x1-3x3-1x1 convolutions) which are more parameter efficient than BasicBlocks. Despite having more layers than ResNet34, the bottleneck design keeps the parameter count manageable while significantly increasing depth. This is the go-to architecture for most image classification tasks and transfer learning.
Remarks
ResNet architectures are a family of deep convolutional neural networks that introduced skip connections (residual connections) to enable training of very deep networks. The key innovation is learning residual functions with reference to the layer inputs, rather than learning unreferenced functions.
For Beginners: ResNet networks are named after their total number of weight layers. For example, ResNet50 has 50 convolutional and fully-connected layers. These networks can be much deeper than earlier architectures (like VGG) because the skip connections allow gradients to flow more easily during training, solving the "vanishing gradient" problem.
Architecture Types:
- ResNet18/34 use "BasicBlock" with two 3x3 convolutions
- ResNet50/101/152 use "BottleneckBlock" with 1x1, 3x3, 1x1 convolutions for efficiency