gradient accumulation

configs/wilds_vith16-448_ep300.yaml

@@ -1,5 +1,5 @@
 data:
-  batch_size: 512
+  batch_size: 16
   color_jitter_strength: 0.0
   crop_scale:
     - 1.0
@@ -40,6 +40,7 @@ meta:
   use_bfloat16: true
   use_gradient_checkpointing: true
 optimization:
+  gradient_accumulation_steps: 32
   ema:
     - 0.996
     - 1.0

src/helper.py

@@ -167,5 +167,5 @@ def init_opt(
             ref_wd=wd,
             final_wd=final_wd,
             T_max=int(ipe_scale*num_epochs*iterations_per_epoch))
-    scaler = torch.cuda.amp.GradScaler() if use_bfloat16 else None
+    scaler = None
     return optimizer, scaler, scheduler, wd_scheduler

src/train.py

@@ -114,6 +114,7 @@ def main(args, resume_preempt=False):
     start_lr = args["optimization"]["start_lr"]
     lr = args["optimization"]["lr"]
     final_lr = args["optimization"]["final_lr"]
+    accum_steps = args["optimization"].get("gradient_accumulation_steps", 1)
 
     # -- LOGGING
     folder = args["logging"]["folder"]
@@ -275,6 +276,7 @@ def main(args, resume_preempt=False):
                 torch.save(save_dict, save_path.format(epoch=f"{epoch + 1}"))
 
     # -- TRAINING LOOP
+    optimizer.zero_grad()
     for epoch in range(start_epoch, num_epochs):
         logger.info("Epoch %d" % (epoch + 1))
 
@@ -300,10 +302,6 @@ def main(args, resume_preempt=False):
             maskB_meter.update(len(masks_pred[0][0]))
 
             def train_step():
-                _new_lr = scheduler.step()
-                _new_wd = wd_scheduler.step()
-                # --
-
                 def forward_target():
                     with torch.no_grad():
                         h = target_encoder(imgs)
@@ -332,24 +330,28 @@ def main(args, resume_preempt=False):
                     z = forward_context()
                     loss = loss_fn(z, h)
 
-                #  Step 2. Backward & step
+                # Step 2. Backward (accumulate gradients)
-                if use_bfloat16:
+                (loss / accum_steps).backward()
-                    scaler.scale(loss).backward()
+
-                    scaler.step(optimizer)
+                # Step 3. Optimizer / scheduler / momentum (every accum_steps)
-                    scaler.update()
+                is_accumulated = ((itr + 1) % accum_steps == 0)
-                else:
+                if is_accumulated:
-                    loss.backward()
+                    _new_lr = scheduler.step()
-                    optimizer.step()
+                    _new_wd = wd_scheduler.step()
-                grad_stats = grad_logger(encoder.named_parameters())
+                    optimizer.step()
-                optimizer.zero_grad(set_to_none=True)
+                    optimizer.zero_grad(set_to_none=True)
+                    grad_stats = grad_logger(encoder.named_parameters())
 
-                # Step 3. momentum update of target encoder
                     with torch.no_grad():
                         m = next(momentum_scheduler)
                         for param_q, param_k in zip(
                             encoder.parameters(), target_encoder.parameters()
                         ):
                             param_k.data.mul_(m).add_((1.0 - m) * param_q.detach().data)
+                else:
+                    _new_lr = None
+                    _new_wd = None
+                    grad_stats = None
 
                 return (float(loss), _new_lr, _new_wd, grad_stats)
 
@@ -363,6 +365,8 @@ def main(args, resume_preempt=False):
                     epoch + 1, itr, loss, maskA_meter.val, maskB_meter.val, etime
                 )
                 if (itr % log_freq == 0) or np.isnan(loss) or np.isinf(loss):
+                    lr_to_log = _new_lr if _new_lr is not None else 0.
+                    wd_to_log = _new_wd if _new_wd is not None else 0.
                     logger.info(
                         "[%d, %5d] loss: %.3f "
                         "masks: %.1f %.1f "
@@ -375,8 +379,8 @@ def main(args, resume_preempt=False):
                             loss_meter.avg,
                             maskA_meter.avg,
                             maskB_meter.avg,
-                            _new_wd,
+                            wd_to_log,
-                            _new_lr,
+                            lr_to_log,
                             torch.cuda.max_memory_allocated() / 1024.0**2,
                             time_meter.avg,
                         )