add update_freq option for gradient accumulation

ppcls/configs/ImageNet/ConvNeXt/convnext_tiny.yaml

@@ -15,7 +15,7 @@ Global:
   save_inference_dir: ./inference
   # training model under @to_static
   to_static: False
-
+  update_freq: 8
 
 # model ema
 EMA:
@@ -51,7 +51,7 @@ Optimizer:
   one_dim_param_no_weight_decay: True
   lr:
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 4e-3  # lr 4e-3 for total_batch_size 4096
     eta_min: 1e-6
     warmup_epoch: 20
     warmup_start_lr: 0

ppcls/engine/engine.py

@@ -119,6 +119,9 @@ class Engine(object):
         # EMA model
         self.ema = "EMA" in self.config and self.mode == "train"
 
+        # gradient accumulation
+        self.update_freq = self.config["Global"].get("update_freq", 1)
+
         if "class_num" in config["Global"]:
             global_class_num = config["Global"]["class_num"]
             if "class_num" not in config["Arch"]:
@@ -229,7 +232,7 @@ class Engine(object):
         if self.mode == 'train':
             self.optimizer, self.lr_sch = build_optimizer(
                 self.config["Optimizer"], self.config["Global"]["epochs"],
-                len(self.train_dataloader),
+                len(self.train_dataloader) // self.update_freq,
                 [self.model, self.train_loss_func])
 
         # for amp training
@@ -312,6 +315,7 @@ class Engine(object):
 
         self.max_iter = len(self.train_dataloader) - 1 if platform.system(
         ) == "Windows" else len(self.train_dataloader)
+        self.max_iter = self.max_iter // self.update_freq * self.update_freq
         for epoch_id in range(best_metric["epoch"] + 1,
                               self.config["Global"]["epochs"] + 1):
             acc = 0.0

ppcls/engine/train/train.py

@@ -53,25 +53,32 @@ def train_epoch(engine, epoch_id, print_batch_step):
             out = forward(engine, batch)
             loss_dict = engine.train_loss_func(out, batch[1])
 
+        # loss
+        loss = loss_dict["loss"] / engine.update_freq
+
         # step opt
         if engine.amp:
-            scaled = engine.scaler.scale(loss_dict["loss"])
+            scaled = engine.scaler.scale(loss)
             scaled.backward()
-            for i in range(len(engine.optimizer)):
-                engine.scaler.minimize(engine.optimizer[i], scaled)
+            if (iter_id + 1) % engine.update_freq == 0:
+                for i in range(len(engine.optimizer)):
+                    engine.scaler.minimize(engine.optimizer[i], scaled)
         else:
-            loss_dict["loss"].backward()
+            loss.backward()
+            if (iter_id + 1) % engine.update_freq == 0:
+                for i in range(len(engine.optimizer)):
+                    engine.optimizer[i].step()
+
+        if (iter_id + 1) % engine.update_freq == 0:
+            # clear grad
             for i in range(len(engine.optimizer)):
-                engine.optimizer[i].step()
-        # clear grad
-        for i in range(len(engine.optimizer)):
-            engine.optimizer[i].clear_grad()
-        # step lr
-        for i in range(len(engine.lr_sch)):
-            engine.lr_sch[i].step()
-        # update ema
-        if engine.ema:
-            engine.model_ema.update(engine.model)
+                engine.optimizer[i].clear_grad()
+            # step lr
+            for i in range(len(engine.lr_sch)):
+                engine.lr_sch[i].step()
+            # update ema
+            if engine.ema:
+                engine.model_ema.update(engine.model)
 
         # below code just for logging
         # update metric_for_logger