Exploring simple siamese representation learning

This is a PyTorch re-implementation of the SimSiam paper on ImageNet dataset. The results match that reported in the paper. The implementation is based on the codes of MOCO.

Unsupervised pre-training

To run unsupervised pre-training on ImageNet,

sh train_simsiam.sh

This is to do the unsupervised pre-training for 100 epochs. Please modify the path to your ImageNet data folder.

Note 1: I try to follow the setting in the paper, which is bs=512 and lr=0.1 on 8-GPU, but somehow I can not fit it. So I used the max batch_size that I can fit (432) while kept the lr unchaged (0.1).

Note 2: In pre-training, I didn’t fix the lr of prediction MLP. According to the paper (Table. 1), fixing the lr of prediction MLP can give slightly improvements (67.7% -> 68.1%). You can try it if interested.

Linear evaluation

To run linear evaluation,

sh train_lincls.sh

The linear evaluation is done using NVIDIA LARC optimizer by setting trus_coefficient=0.001 and clip=False. The batch size is 4096.

Note: I first followed the setting in the paper, which is Lr=0.32 (0.02*4096/256). But I can only got a result of 66.0%. Then I increased the learning rate to Lr=1.6 (0.1*4096.256) and achieved the result of 67.8%. The results and models are given below.

SimSiam pretrained batchsize lincls Lr Top-1 Acc
Reported 512 0.32 67.7%
Reproduced 432 (Model) 1.6 67.8% (Model)
Reproduced 432 0.32 66.0%


Thank Xinlei for his help on some implementation details.