updat

odlab/utils/lr_scheduler.py

@@ -11,9 +11,8 @@ class LinearWarmUpScheduler(object):
 
     def set_lr(self, optimizer, lr):
         for param_group in optimizer.param_groups:
-            init_lr = param_group['initial_lr']
-            ratio = init_lr / self.base_lr
-            param_group['lr'] = lr * ratio
+            param_group['lr'] = lr
+
 
     def __call__(self, iter, optimizer):
         # warmup

yolo/engine.py

@@ -62,7 +62,8 @@ class YoloTrainer(object):
         self.scaler = torch.cuda.amp.GradScaler(enabled=args.fp16)
 
         # ---------------------------- Build Optimizer ----------------------------
-        cfg.base_lr = cfg.per_image_lr * args.batch_size
+        self.grad_accumulate = max(256 // args.batch_size, 1)
+        cfg.base_lr = cfg.per_image_lr * args.batch_size * self.grad_accumulate
         cfg.min_lr  = cfg.base_lr * cfg.min_lr_ratio
         self.optimizer, self.start_epoch = build_yolo_optimizer(cfg, model, args.resume)
 
@@ -208,21 +209,23 @@ class YoloTrainer(object):
                 loss_dict = self.criterion(outputs=outputs, targets=targets)
                 losses = loss_dict['losses']
                 loss_dict_reduced = distributed_utils.reduce_dict(loss_dict)
+                losses /= self.grad_accumulate
 
             # Backward
             self.scaler.scale(losses).backward()
 
             # Optimize
-            if self.cfg.clip_max_norm > 0:
-                self.scaler.unscale_(self.optimizer)
-                torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=self.cfg.clip_max_norm)
-            self.scaler.step(self.optimizer)
-            self.scaler.update()
-            self.optimizer.zero_grad()
-
-            # ModelEMA
-            if self.model_ema is not None:
-                self.model_ema.update(model)
+            if (iter_i + 1) % self.grad_accumulate == 0:
+                if self.cfg.clip_max_norm > 0:
+                    self.scaler.unscale_(self.optimizer)
+                    torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=self.cfg.clip_max_norm)
+                self.scaler.step(self.optimizer)
+                self.scaler.update()
+                self.optimizer.zero_grad()
+
+                # ModelEMA
+                if self.model_ema is not None:
+                    self.model_ema.update(model)
 
             # Update log
             metric_logger.update(**loss_dict_reduced)
@@ -343,7 +346,8 @@ class RTDetrTrainer(object):
         self.scaler = torch.cuda.amp.GradScaler(enabled=args.fp16)
 
         # ---------------------------- Build Optimizer ----------------------------
-        cfg.base_lr = cfg.per_image_lr * args.batch_size
+        self.grad_accumulate = max(16 // args.batch_size, 1)
+        cfg.base_lr = cfg.per_image_lr * args.batch_size * self.grad_accumulate
         cfg.min_lr  = cfg.base_lr * cfg.min_lr_ratio
         self.optimizer, self.start_epoch = build_rtdetr_optimizer(cfg, model, args.resume)
 
@@ -477,22 +481,24 @@ class RTDetrTrainer(object):
                 outputs = model(images, targets)    
                 loss_dict = self.criterion(outputs, targets)
                 losses = sum(loss_dict.values())
+                losses /= self.grad_accumulate
                 loss_dict_reduced = distributed_utils.reduce_dict(loss_dict)
 
             # Backward
             self.scaler.scale(losses).backward()
 
             # Optimize
-            if self.cfg.clip_max_norm > 0:
-                self.scaler.unscale_(self.optimizer)
-                torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=self.cfg.clip_max_norm)
-            self.scaler.step(self.optimizer)
-            self.scaler.update()
-            self.optimizer.zero_grad()
-
-            # ModelEMA
-            if self.model_ema is not None:
-                self.model_ema.update(model)
+            if (iter_i + 1) % self.grad_accumulate == 0:
+                if self.cfg.clip_max_norm > 0:
+                    self.scaler.unscale_(self.optimizer)
+                    torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=self.cfg.clip_max_norm)
+                self.scaler.step(self.optimizer)
+                self.scaler.update()
+                self.optimizer.zero_grad()
+
+                # ModelEMA
+                if self.model_ema is not None:
+                    self.model_ema.update(model)
 
             # Update log
             metric_logger.update(**loss_dict_reduced)