This checkpoint contains weights of a fully trained ImageNet encoder, optimized with respect to 3 * L_align(alpha=2) + L_uniform(t=3) with the MoCo v2 improvements (strong augmentation, MLP head, cosine learning rate). It achieves 67.694% validation linear classification top1 accuracy.

imagenet_align_uniform.pth.tar contains weights for both encoder_q and encoder_k, their MLP heads, the MoCo queue, and the optimizer state. This is the same format as is saved at https://github.com/SsnL/moco_align_uniform/blob/02b6c3fa24d47da06ea3a53713500f809f496238/main_moco.py#L315-L319.

imagenet_align_uniform_with_linear_clf.pth.tar contains weights for encoder_q (the one used in downstream task), with the MLP head replaced with a linear classifier already optimized for ImageNet classification.