/tf-resnet

TensorFlow/Keras ResNet on CIFAR-10

Primary LanguagePython

TensorFlow CIFAR-10 ResNet

Basic residual network for CIFAR-10 classification from Deep Residual Learning for Image Recognition. Uses TensorFlow and Keras.

Results

Wow, it's true! Residual connections really do make a magical difference for deep nets.

  • plain20: 20-layer network without residual connections. Trains fine.
  • plain56: 56-layer network without residual connections. Does terribly.
  • resnet56: 56-layers network with residual connections. Performance matches and exceeds 20-layer network.

Usage

Setup

$ pipenv sync

Training

$ pipenv run python -m resnet.train with net=resnet20