Claim actually close to 1% improvement on image net data set¹.

Classification accuracy from the paper¹

Architecture wise, its a very simple network resnet 50 having a 128-dimensional head. If you want you can add a few more layers as well.

Architecture and training process from the paper¹

def forward(self, x): 
feat = self.encoder(x) #normalizing the 128 vector is required Code self.encoder = resnet50() self.head = nn.Linear(2048, 128)  
feat = F.normalize(self.head(feat), dim=1) 
return feat

As shown in the figure training is done in two-stage.

• Train using contrastive loss (two variations)
• freeze the learned representations and then learn a classifier on a linear layer using a softmax loss. (From the paper)

The above is pretty self explanatory.

Loss, the main flavor of this paper is understanding the self supervised contrastive loss and supervised contrastive loss.

#programming #data-science