feat: single machine DDP (test)

train/train.py

@@ -6,34 +6,54 @@ from model import Network
 from torch.utils.data import DataLoader
 import matplotlib.pyplot as plt
 
-device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+import torch.multiprocessing as mp
+from torch.utils.data.distributed import DistributedSampler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.distributed import init_process_group, destroy_process_group
+import os
 
-model = Network().to(device)
-dataset = Cifar10Dataset('./dataset_dir/cifar-10-batches-py')
-loader = DataLoader(dataset, batch_size=32, shuffle=True)
+def ddp_init(rank, world_size):
+    os.environ['MASTER_ADDR'] = 'localhost'
+    os.environ['MASTER_PORT'] = '21046'
+    init_process_group('nccl', rank=rank, world_size=world_size)
 
-optimizer = optim.Adam(model.parameters(), lr=0.001)
-criterion = nn.CrossEntropyLoss()
 
-for epoch in range(50):
-    model.train()
-    train_loss_sum = 0
-    train_correct_sum = 0
-    for x, y in loader:
-        x = x.float()
-        x, y = x.to(device), y.to(device)
+def main(rank, world_size):
+    ddp_init(rank, world_size)
 
-        predict = model(x)
-        loss = criterion(predict, y)
-        loss.backward()
+    model = Network()
+    model = DDP(model, device_ids=rank)
 
-        # evaluate
-        train_loss_sum += loss.item()
-        predicted_classes = torch.argmax(predict, dim=1)
-        train_correct_sum += (predicted_classes == y).sum()
+    dataset = Cifar10Dataset('./dataset_dir/cifar-10-batches-py')
+    loader = DataLoader(dataset, batch_size=32, shuffle=False, sampler=DistributedSampler(dataset, ))
+
+    optimizer = optim.Adam(model.parameters(), lr=0.001)
+    criterion = nn.CrossEntropyLoss()
+
+    for epoch in range(50):
+        model.train()
+        train_loss_sum = 0
+        train_correct_sum = 0
+        for x, y in loader:
+            x = x.float()
+            x, y = x.to(rank), y.to(rank)
+
+            predict = model(x)
+            loss = criterion(predict, y)
+            loss.backward()
+
+            # evaluate
+            train_loss_sum += loss.item()
+            predicted_classes = torch.argmax(predict, dim=1)
+            train_correct_sum += (predicted_classes == y).sum()
+
+            optimizer.step()
+            optimizer.zero_grad()
+        print(train_loss_sum / len(loader))
+        print((train_correct_sum / len(dataset)).item(),'%')
+        print()
+
+if __name__ == '__main__':
+    world_size = torch.cuda.device_count()
+    mp.spawn(main, args=(world_size, ))
     
-        optimizer.step()
-        optimizer.zero_grad()
-    print(train_loss_sum / len(loader))
-    print((train_correct_sum / len(dataset)).item(),'%')
-    print()