fix bugs and norm only for cl feature

PreResNet_rours.py

@@ -179,9 +179,10 @@ def forward(self, x, filter=None, lin=0, lout=5):
             elif filter == 'dct':
                 out += self.filter_DCT(out)
 
+            logits = self.linear(out)
+            # Feature for CL
             if self.norm:
                 out = F.normalize(out, dim=1)
-            logits = self.linear(out)
 
         return logits, out
 

loss.py

@@ -132,7 +132,7 @@ def __init__(self, args, contrast_mode='all', base_temperature=0.07):
         # Linear
         # gamma_schedule = torch.linspace(0, 1, args.warm_up)
         # Exponent
-        gamma_schedule[:args.warm_up] = torch.logspace(-args.warm_up, 0, args.warmup, np.e)
+        gamma_schedule[:args.warm_up] = torch.logspace(-args.warm_up, 0, args.warm_up, np.e)
         self.gamma = gamma_schedule * args.gamma
         self.eta = args.eta
         self.warmup = args.warm_up

main_ce1.py

@@ -252,6 +252,8 @@ def main(args, params={}):
         params = params['best']
         args.nrun = True
     args = update_args(params)
+    # TODO
+    args.record = True
     res = main(args, params=params)
     # TODO
     if args.out_tmp:

utils.py

@@ -165,6 +165,8 @@ def train_ours(args, model, loader, optimizer, epoch, scheduler, criterion, net_
     log_value('train_detail/delta_smooth/std', delta_smooth.std(), step=epoch)
     log_value('train_detail/net_record/avg', net_record.mean(), step=epoch)
     log_value('train_detail/net_record/std', net_record.std(), step=epoch)
+    if args.model_type == 'ours_cl':
+        log_value('train_detail/gamma', criterion.gamma[epoch], step=epoch)
     # Print and log stats for the epoch
     log_value('train/loss', train_loss.avg, step=epoch)
     log(args.logpath, 'Time for Train-Epoch-{}/{}:{:.1f}s Acc:{}, Loss:{}\n'.format(epoch, args.n_epoch, time.time() - t0, correct.avg, train_loss.avg))