add loss_cfc

Author: SiLangWHL

.DS_Store

@@ -231,7 +231,7 @@ def batch_forward(self, inputs, labels=None, feature_ext=False, phase='train'):
                     var_dic['feat'], var_dic['gt'] = self.features, labels
                     var_dic['op_opt'] = self.op_opt
                     kappa = var_dic['kappa'].detach()
-                    self.loss_icd, bias, op, self.sim_c = multi_get_loss(**var_dic)
+                    self.loss_icd, self.loss_cfc, bias, op, self.sim_c = multi_get_loss(**var_dic)
                     self.networks['classifier'].module.bias = bias
                     self.networks['classifier'].module.op = op
                     self.logits = self.networks['classifier'](self.features, bias.repeat(self.num_gpus)) 
@@ -263,9 +263,11 @@ def batch_loss(self, labels):
             self.loss_perf = self.criterions['PerformanceLoss'](self.logits, labels)
             self.loss_perf *=  self.criterion_weights['PerformanceLoss']
             self.loss += self.loss_perf
+        weight = self.op_opt['auxlossweight'] * (1 - self.current_epoch / self.total_epoch) ** 0.9
         if 'icd' in self.op_opt['auxloss']:
-            weight = self.op_opt['auxlossweight'] * (1 - self.current_epoch / self.total_epoch) ** 0.9
             self.loss += weight * self.loss_icd 
+        if 'cfc' in self.op_opt['auxloss']:
+            self.loss += weight * self.loss_cfc 
 
 
     def shuffle_batch(self, x, y):
@@ -361,6 +363,9 @@ def train(self):
 
                         if 'icd' in self.op_opt['auxloss']:
                             minibatch_loss_icd = self.loss_icd.item()
+                        if 'cfc' in self.op_opt['auxloss']:
+                            minibatch_loss_cfc = self.loss_cfc.item()
+
 
                         minibatch_acc = mic_acc_cal(preds, labels)
 
@@ -377,6 +382,8 @@ def train(self):
                         if 'vmf' in self.config['networks']['classifier']['def_file']:
                             if 'icd' in self.op_opt['auxloss']:
                                 print_str += ['Licd: %.3f'% (minibatch_loss_icd)]
+                            if 'cfc' in self.op_opt['auxloss']:
+                                print_str += ['Lcfc: %.3f'% (minibatch_loss_cfc)]
                         print_write(print_str, self.log_file)
 
                         loss_info = {

classification/.DS_Store

@@ -96,19 +96,26 @@ def get_all_overlap(op_opt, apk, bias, kappa, feat, gt, num_classes, mu, ni):
         loss_icd = ( ni * valid_op ).mean().mean() #.sum().sum() / ni.sum().sum()
     else:
         loss_icd = 0
+    if 'cfc' in op_opt['auxloss']:
+        sim_feat_class = ( mu[mask] * mu_feat ).sum(1)
+        kl = apk[mask] * kappa[mask] * ( 1 - sim_feat_class )
+        op = 1 / ( 1 + kl ) 
+        loss_cfc = ( 1 - op ).mean() 
+    else:
+        loss_cfc = 0
     with torch.no_grad():
         sim_c = (ni * sim_class).mean(1).mean(0)
         op = valid_op.mean(1)
-    return loss_icd, op, sim_c
+    return loss_icd, loss_cfc, op, sim_c
 
 ni = read_ni()
 def multi_get_loss(op_opt, p, kappa, weight, feat, gt, num_classes):
     mu = get_mu(weight)
     apk = get_apk(p, kappa)
     bias = get_bias(p, kappa, apk)
     if op_opt['auxloss'] != []:
-        loss_icd, op, sim_c = get_all_overlap(op_opt, apk, bias, kappa, feat, gt, num_classes, mu, ni)
+        loss_icd, loss_cfc, op, sim_c = get_all_overlap(op_opt, apk, bias, kappa, feat, gt, num_classes, mu, ni)
     else:
-        loss_icd, sim_c, op = 0.0, 0.0, bias.detach()
-    return loss_icd, bias, op, sim_c
+        loss_icd, loss_cfc, sim_c, op = 0.0, 0.0, 0.0, bias.detach()   #  hard code to avoid happening errors when running codes, I will rewrite it.
+    return loss_icd, loss_cfc, bias, op, sim_c