ResNet

About

Re-implementation of ResNet. Original paper.

Method

Paper Details (For ImageNet)

• Image scaled to 224x224 crop (same as AlexNet). Per-pixel mean subtracted.
• Batch norm after conv and before activation.
• Initiaise weights based on He initialisation.
• SGD Initialisation with Batch Size := 256.
• Start LR := 0.1, /= 10 on plateau. 6,000,000 iters.
• Weight decay := 0.0001. Momentum of 0.9.
• No dropout.
• 10 crop testing?

TODO

• ResNet
  • Model
  • Train Code
• MNIST (~98%)
• CIFAR-10 (83.94%)
• Tiny ImageNet
  • Initial Results (ResNet18(51% acc)) This seems to match other results.