Torchlira: attempt microbatch

lira-pytorch/train.py

@@ -21,6 +21,8 @@ from tqdm import tqdm
 from opacus.validators import ModuleValidator
 from opacus import PrivacyEngine
 from opacus.utils.batch_memory_manager import BatchMemoryManager
+import pyvacy
+#from pyvacy import optim#, analysis, sampling
 
 from wide_resnet import WideResNet
 
@@ -113,6 +115,12 @@ def run():
     ModuleValidator.validate(m, strict=True)
 
     optim = torch.optim.SGD(m.parameters(), lr=args.lr, momentum=0.9, weight_decay=5e-4)
+    #optim = pyvacy.DPSGD(
+    #    params=m.parameters(),
+    #    lr=args.lr,
+    #    momentum=0.9,
+    #    weight_decay=5e-4,
+    #)
     sched = torch.optim.lr_scheduler.CosineAnnealingLR(optim, T_max=args.epochs)
 
     privacy_engine = PrivacyEngine()
@@ -124,12 +132,13 @@ def run():
         target_epsilon=1,
         target_delta=1e-4,
         max_grad_norm=1.0,
+        batch_first=True,
     )
 
     print(f"Device: {DEVICE}")
+    accumulation_steps = 10
 
     # Train
-    # max_physical_batch_size=MAX_PHYSICAL_BATCH_SIZE,
     with BatchMemoryManager(
         data_loader=train_dl,
         max_physical_batch_size=1000,
@@ -142,14 +151,24 @@ def run():
             #pbar = tqdm(train_dl, leave=False)
             for itr, (x, y) in enumerate(pbar):
                 x, y = x.to(DEVICE), y.to(DEVICE)
+                if False:
+                    loss = F.cross_entropy(m(x), y) / accumulation_steps
+                    loss_norm = loss / accumulation_steps
+                    loss_total += loss_norm
+                    loss_norm.backward()
+                    pbar.set_postfix_str(f"loss: {loss:.2f}")
 
-                loss = F.cross_entropy(m(x), y)
-                loss_total += loss
+                    if ((itr + 1) % accumulation_steps == 0) or (itr + 1 == len(memory_safe_data_loader)):
+                        optim.step()
+                        optim.zero_grad()
+                else:
+                    loss = F.cross_entropy(m(x), y)
+                    loss_total += loss
 
-                pbar.set_postfix_str(f"loss: {loss:.2f}")
-                optim.zero_grad()
-                loss.backward()
-                optim.step()
+                    pbar.set_postfix_str(f"loss: {loss:.2f}")
+                    optim.zero_grad()
+                    loss.backward()
+                    optim.step()
             sched.step()
 
             wandb.log({"loss": loss_total / len(train_dl)})