MFM

Unofficial code for paper "Masked Feature Prediction for Self-Supervised Visual Pre-Training" (https://arxiv.org/pdf/2206.07706.pdf)

Below are experiments with resnet50. Though better result is achieved, it seems that the baseline is also much higher than in paper.

	top-1 acc	pretrain	finetune
paper scratch	78.1	-	-
paper mfm pretrain	78.5	-	-
scratch	78.542	-	link
supervised pretrain	78.942	-	link
mfm pretrain	78.826	link	link

Note: Supervised pretrain means finetune from torchvision resnet weights (by setting pretrained=True). It seems that supervised pretrain is better than the proposed mfm pretrain.

Platform

• pytorch 1.13.1
• torchvision 0.14.1
• dali 1.21.0
• cuda 11.6
• V100 GPU(32G) x 8
• driver: 470.82.01

Dataset

Prepare imagenet val set in same method as pytorch official classification example, and then link them to the folder of this repo:

    $ mkdir -p imagenet
    $ ln -s /path/to/imagenet/train ./imagenet/train
    $ ln -s /path/to/imagenet/val ./imagenet/val

Train

Pretraining and finetuning Command is here.

More ablations

Here are some points that affects the results:

1. finetune --val-resize-size
   When we eval the model after finetuning, we always resize the short side of the image to a fixed value before a center crop operation. Here I find sometimes the value of fixed short side size affects the acc by a noticeable margin. Take the "supervised pretrain" as example:

   val-resize-size	234	235	236
   top-1 acc	78.856	78.942	78.794

2. finetune with bce loss is important
   We can see this by finetuning from scratch with CE(cross entropy) loss and BCE(binary cross entropy) loss, the result is:

   loss	CE	BCE
   top-1 acc	78.542	78.952

3. pretrain random crop area
   We usually crop a part of the image with certain area ratio from the original image, and the default value of this ratio is 0.08-1.0 with torchvision RandomResizedCrop. Different self-supervised learning methods tend to prefer different random area ratios. For example, MAE uses 0.2-1.0, MAE3d uses 0.5-1.0, and SimMIM uses 0.67-1.0. Here I find a smaller lower bound of 0.2-1.0 is better:

   random area ratio	0.67-1.0	0.2-1.0	0.1-1.0
   top-1 acc	78.770	78.826	78.842

   Though here 0.1-1.0 is better than 0.2-1.0, I still use the latter, since, with 0.1-1.0, the finetuning eval result is more affacted by val-resize-size:

   val-resize-size	234	235	236
   0.2-1.0	78.816	78.826	78.796
   0.1-1.0	78.730	78.842	78.738

4. model variance
   Here I pretrain the model for 4 times(2 on 8 v100 gpu, and 2 on 8 p40 gpu) with identical configuration. Then I finetune 3 times for each of the pretrained model(with 8 p40). Results are listed below. We can see that the results varies between a big margin. Maybe the above good results are brought by a good luck. Hence, I cannot say that I have certainly reproduced the results in the paper now.

   pretrain	finetune	acc1(235)	mean/std
   round 1	round 1	78.654	78.644/0.024	78.621/0.08
   	round 2	78.61
   	round 3	78.668
   round 2	round 1	78.646	78.642/0.122
   	round 2	78.79
   	round 3	78.49
   round 3	round 1	78.516	78.612/0.073
   	round 2	78.626
   	round 3	78.694
   round 4	round 1	78.608	78.584/0.080
   	round 2	78.668
   	round 3	78.476