bug fix

Learner.py

@@ -25,7 +25,7 @@ def __init__(self, conf, inference=False):
             print('{}_{} model generated'.format(conf.net_mode, conf.net_depth))
 
         if not inference:
-#             self.milestones = conf.milestones
+            self.milestones = conf.milestones
             self.loader, self.class_num = get_train_loader(conf)        
 
             self.writer = SummaryWriter(conf.log_path)
@@ -48,7 +48,7 @@ def __init__(self, conf, inference=False):
                                     {'params': paras_only_bn}
                                 ], lr = conf.lr, momentum = conf.momentum)
             print(self.optimizer)
-            self.scheduler = optim.lr_scheduler.ReduceLROnPlateau(self.optimizer, patience=40, verbose=True)
+#             self.scheduler = optim.lr_scheduler.ReduceLROnPlateau(self.optimizer, patience=40, verbose=True)
 
             print('optimizers generated')    
             self.board_loss_every = len(self.loader)//100
@@ -183,9 +183,15 @@ def find_lr(self,
 
     def train(self, conf, epochs):
         self.model.train()
-        running_loss = 0.
+        running_loss = 0.            
         for e in range(epochs):
             print('epoch {} started'.format(e))
+            if e == self.milestones[0]:
+                self.schedule_lr()
+            if e == self.milestones[1]:
+                self.schedule_lr()      
+            if e == self.milestones[2]:
+                self.schedule_lr()                                 
             for imgs, labels in tqdm(iter(self.loader)):
                 imgs = imgs.to(conf.device)
                 labels = labels.to(conf.device)
@@ -200,7 +206,6 @@ def train(self, conf, epochs):
                 if self.step % self.board_loss_every == 0 and self.step != 0:
                     loss_board = running_loss / self.board_loss_every
                     self.writer.add_scalar('train_loss', loss_board, self.step)
-                    self.scheduler.step(loss_board)
                     running_loss = 0.
 
                 if self.step % self.evaluate_every == 0 and self.step != 0:
@@ -216,18 +221,6 @@ def train(self, conf, epochs):
 
                 self.step += 1
 
-#             if e == self.milestones[0]:
-#                 self.schedule_lr()               
-#                 self.save_state(conf, accuracy, to_save_folder=True, extra='{} epochs'.format(e))
-
-#             if e == self.milestones[1]:
-#                 self.schedule_lr()               
-#                 self.save_state(conf, accuracy, to_save_folder=True, extra='{} epochs'.format(e))
-
-#             if e == self.milestones[2]:
-#                 self.schedule_lr()                                 
-#                 self.save_state(conf, accuracy, to_save_folder=True, extra='{} epochs'.format(e))
-
         self.save_state(conf, accuracy, to_save_folder=True, extra='final')
 
     def schedule_lr(self):

config.py

@@ -34,8 +34,8 @@ def get_config(training = True):
         conf.save_path = conf.work_path/'save'
     #     conf.weight_decay = 5e-4
         conf.lr = 1e-3
-#         conf.milestones = [3,4,5] # mobildefacenet
-        conf.milestones = [4,6,7] # arcface
+        conf.milestones = [10,13,15] # mobildefacenet
+#         conf.milestones = [12,15,18] # arcface
         conf.momentum = 0.9
         conf.pin_memory = True
 #         conf.num_workers = 4 # when batchsize is 200

model.py

@@ -269,12 +269,13 @@ def forward(self, embbedings, label):
         #      0<=theta+m<=pi
         #     -m<=theta<=pi-m
         cond_v = cos_theta - self.threshold
-        cond_mask = cond_v > 0
+        cond_mask = cond_v <= 0
         keep_val = (cos_theta - self.mm) # when theta not in [0,pi], use cosface instead
         cos_theta_m[cond_mask] = keep_val[cond_mask]        
         label = label.view(-1,1) #size=(B,1) 
         output = cos_theta * 1.0 # a little bit hacky way to prevent in_place operation on cos_theta
-        output[torch.range(0,nB-1).to(torch.long),label] = cos_theta_m[torch.range(0,nB-1).to(torch.long),label]
+        idx_ = torch.arange(0, nB, dtype=torch.long)
+        output[idx_, label] = cos_theta_m[idx_, label]
         output *= self.s # scale up in order to make softmax work, first introduced in normface
         return output
 

train.py

@@ -2,9 +2,11 @@
 from Learner import face_learner
 import argparse
 
+# python train.py -net mobilefacenet -b 200 -w 4 
+
 if __name__ == '__main__':
     parser = argparse.ArgumentParser(description='for face verification')
-    parser.add_argument("-e", "--epochs", help="training epochs", default=8, type=int)
+    parser.add_argument("-e", "--epochs", help="training epochs", default=18, type=int)
     parser.add_argument("-net", "--net_mode", help="which network, [ir, ir_se, mobilefacenet]",default='ir_se', type=str)
     parser.add_argument("-depth", "--net_depth", help="how many layers [50,100,152]", default=50, type=int)
     parser.add_argument('-lr','--lr',help='learning rate',default=1e-3, type=float)

utils.py

@@ -13,7 +13,7 @@
 
 def separate_bn_paras(modules):
     if not isinstance(modules, list):
-        modules = [*module.modules()]
+        modules = [*modules.modules()]
     paras_only_bn = []
     paras_wo_bn = []
     for layer in modules: