How to measure convergence? about moco HOT 1 CLOSED

The code release in this repo is meant for reproducing the results of research papers.

We help users who have trouble using the code to reproduce results in the paper, but in general we do not provide suggestions on other issues such as how to apply the method to a new task or dataset.

Regarding ImageNet, as we showed in the MoCo v2 paper, the models are far from convergence at 200ep (67.5% accuracy) and can be improved by a lot when training 800ep (71.1%) or perhaps longer. I suggest to monitor the truth performance of the feature by training a linear classifier or transferring to a downstream task, because the loss/accuracy of the pretext task is not indicative (an easier pretext task does not mean better features).

from moco.