add instance loss

Author: zhedong

.instance_loss.py.swp

@@ -0,0 +1,36 @@
+from typing import Tuple
+
+import torch
+from torch import nn, Tensor
+import torch.nn.functional as F
+
+def l2_norm(v):
+    fnorm = torch.norm(v, p=2, dim=1, keepdim=True) + 1e-6
+    v = v.div(fnorm.expand_as(v))
+    return v
+
+class InstanceLoss(nn.Module):
+    def __init__(self) -> None:
+        super(InstanceLoss, self).__init__()
+
+    def forward(self, feature, label) -> Tensor:
+        # Dual-Path Convolutional Image-Text Embeddings with Instance Loss, ACM TOMM 2020 
+        # https://arxiv.org/abs/1711.05535 
+        # using cross-entropy loss for every class 
+        normed_feature = l2_norm(feature)
+        sim1 = torch.mm(normed_feature, torch.t(normed_feature)) 
+        sim2 = sim1.t()
+        #sim_label = torch.arange(sim1.size(0)).cuda().detach()
+        _, sim_label = torch.unique(label, return_inverse=True)
+        loss = F.cross_entropy(sim1, sim_label) + F.cross_entropy(sim2, sim_label)
+        return loss
+
+
+if __name__ == "__main__":
+    feat = nn.functional.normalize(torch.rand(256, 64, requires_grad=True))
+    lbl = torch.randint(high=10, size=(256,))
+
+    criterion = InstanceLoss()
+    instance_loss = criterion(feat, lbl)
+
+    print(instance_loss)

instance_loss.py

@@ -22,6 +22,7 @@
 import yaml
 from shutil import copyfile
 from circle_loss import CircleLoss, convert_label_to_similarity
+from instance_loss import InstanceLoss
 
 version =  torch.__version__
 #fp16
@@ -61,6 +62,7 @@
 parser.add_argument('--circle', action='store_true', help='use Circle loss' )
 parser.add_argument('--cosface', action='store_true', help='use CosFace loss' )
 parser.add_argument('--contrast', action='store_true', help='use contrast loss' )
+parser.add_argument('--instance', action='store_true', help='use instance loss' )
 parser.add_argument('--triplet', action='store_true', help='use triplet loss' )
 parser.add_argument('--lifted', action='store_true', help='use lifted loss' )
 parser.add_argument('--sphere', action='store_true', help='use sphere loss' )
@@ -213,6 +215,8 @@ def train_model(model, criterion, optimizer, scheduler, num_epochs=25):
         criterion_lifted = losses.GeneralizedLiftedStructureLoss(neg_margin=1, pos_margin=0)
     if opt.contrast: 
         criterion_contrast = losses.ContrastiveLoss(pos_margin=0, neg_margin=1)
+    if opt.instance:
+        criterion_instance = InstanceLoss()
     if opt.sphere:
         criterion_sphere = losses.SphereFaceLoss(num_classes=opt.nclasses, embedding_size=512, margin=4)
     for epoch in range(num_epochs):
@@ -258,7 +262,7 @@ def train_model(model, criterion, optimizer, scheduler, num_epochs=25):
 
                 sm = nn.Softmax(dim=1)
                 log_sm = nn.LogSoftmax(dim=1)
-                return_feature = opt.arcface or opt.cosface or opt.circle or opt.triplet or opt.contrast or opt.lifted or opt.sphere
+                return_feature = opt.arcface or opt.cosface or opt.circle or opt.triplet or opt.contrast or opt.instance or opt.lifted or opt.sphere
                 if return_feature: 
                     logits, ff = outputs
                     fnorm = torch.norm(ff, p=2, dim=1, keepdim=True)
@@ -278,6 +282,8 @@ def train_model(model, criterion, optimizer, scheduler, num_epochs=25):
                         loss +=  criterion_lifted(ff, labels) #/now_batch_size
                     if opt.contrast:
                         loss +=  criterion_contrast(ff, labels) #/now_batch_size
+                    if opt.instance:
+                        loss += criterion_instance(ff, labels)
                     if opt.sphere:
                         loss +=  criterion_sphere(ff, labels)/now_batch_size
                 elif opt.PCB:  #  PCB
@@ -421,7 +427,7 @@ def save_network(network, epoch_label):
 # Load a pretrainied model and reset final fully connected layer.
 #
 
-return_feature = opt.arcface or opt.cosface or opt.circle or opt.triplet or opt.contrast or opt.lifted or opt.sphere
+return_feature = opt.arcface or opt.cosface or opt.circle or opt.triplet or opt.contrast or opt.instance or opt.lifted or opt.sphere
 
 if opt.use_dense:
     model = ft_net_dense(len(class_names), opt.droprate, circle = return_feature)