moco-tf2.x

• Unofficial reimplementation of MoCo: Momentum Contrast for Unsupervised Visual Representation Learning
• Found many helpful implementations from moco.tensorflow
• Augmentation code are copied from SimCLR - A Simple Framework for Contrastive Learning of Visual Representations
• Trying to implement as much as possible in Tensorflow 2.x
  • Used MirroredStrategy with custom training loop (tensorflow-tutorial)

Note

• Difference between official implementation
  • 8 GPUs vs 4 GPUs
  • 53 Hours vs 147 hours (Unsupervised training time) - much slower than official one
• Batch normalization - tf
  • If one sets batch normalization layer as un-trainable, tf will normalize input with their moving mean & var, even though you use training=True
• Lack of information about how to properly apply weight regularization within distributed environment

Result (Result lower than official one)

• MoCo v1
  • Could not reproduce same accuracy (Linear classification protocol on Imagenet) result as official one.

	InfoNCE	(K+1) Accuracy
MoCo V1

	Train Accuracy	Validation Accuracy
lincls

• Comparison with official result

ResNet-50	pre-train epochs	pre-train time	MoCo v1 top-1 acc.
Official Result	200	53 hours	60.6
This repo Result	200	147 hours	50.8