| 
 | 1 | +# Copyright (c) 2017 Facebook, Inc. All rights reserved.  | 
 | 2 | +# BSD 3-Clause License  | 
 | 3 | +#  | 
 | 4 | +# Script adapted from:  | 
 | 5 | +# https://pytorch.org/tutorials/beginner/blitz/cifar10_tutorial.html  | 
 | 6 | +# ==============================================================================  | 
 | 7 | + | 
 | 8 | +# imports  | 
 | 9 | +import torch  | 
 | 10 | +import torchvision  | 
 | 11 | +import torchvision.transforms as transforms  | 
 | 12 | +import torch.nn as nn  | 
 | 13 | +import torch.nn.functional as F  | 
 | 14 | +import torch.optim as optim  | 
 | 15 | +import os  | 
 | 16 | +import argparse  | 
 | 17 | + | 
 | 18 | + | 
 | 19 | +# define network architecture  | 
 | 20 | +class Net(nn.Module):  | 
 | 21 | +    def __init__(self):  | 
 | 22 | +        super(Net, self).__init__()  | 
 | 23 | +        self.conv1 = nn.Conv2d(3, 32, 3)  | 
 | 24 | +        self.pool = nn.MaxPool2d(2, 2)  | 
 | 25 | +        self.conv2 = nn.Conv2d(32, 64, 3)  | 
 | 26 | +        self.conv3 = nn.Conv2d(64, 128, 3)  | 
 | 27 | +        self.fc1 = nn.Linear(128 * 6 * 6, 120)  | 
 | 28 | +        self.dropout = nn.Dropout(p=0.2)  | 
 | 29 | +        self.fc2 = nn.Linear(120, 84)  | 
 | 30 | +        self.fc3 = nn.Linear(84, 10)  | 
 | 31 | + | 
 | 32 | +    def forward(self, x):  | 
 | 33 | +        x = F.relu(self.conv1(x))  | 
 | 34 | +        x = self.pool(F.relu(self.conv2(x)))  | 
 | 35 | +        x = self.pool(F.relu(self.conv3(x)))  | 
 | 36 | +        x = x.view(-1, 128 * 6 * 6)  | 
 | 37 | +        x = self.dropout(F.relu(self.fc1(x)))  | 
 | 38 | +        x = F.relu(self.fc2(x))  | 
 | 39 | +        x = self.fc3(x)  | 
 | 40 | +        return x  | 
 | 41 | + | 
 | 42 | + | 
 | 43 | +def train(train_loader, model, criterion, optimizer, epoch, device, print_freq, rank):  | 
 | 44 | +    running_loss = 0.0  | 
 | 45 | +    for i, data in enumerate(train_loader, 0):  | 
 | 46 | +        # get the inputs; data is a list of [inputs, labels]  | 
 | 47 | +        inputs, labels = data[0].to(device), data[1].to(device)  | 
 | 48 | + | 
 | 49 | +        # zero the parameter gradients  | 
 | 50 | +        optimizer.zero_grad()  | 
 | 51 | + | 
 | 52 | +        # forward + backward + optimize  | 
 | 53 | +        outputs = model(inputs)  | 
 | 54 | +        loss = criterion(outputs, labels)  | 
 | 55 | +        loss.backward()  | 
 | 56 | +        optimizer.step()  | 
 | 57 | + | 
 | 58 | +        # print statistics  | 
 | 59 | +        running_loss += loss.item()  | 
 | 60 | +        if i % print_freq == 0:  # print every print_freq mini-batches  | 
 | 61 | +            print(  | 
 | 62 | +                "Rank %d: [%d, %5d] loss: %.3f"  | 
 | 63 | +                % (rank, epoch + 1, i + 1, running_loss / print_freq)  | 
 | 64 | +            )  | 
 | 65 | +            running_loss = 0.0  | 
 | 66 | + | 
 | 67 | + | 
 | 68 | +def evaluate(test_loader, model, device):  | 
 | 69 | +    classes = (  | 
 | 70 | +        "plane",  | 
 | 71 | +        "car",  | 
 | 72 | +        "bird",  | 
 | 73 | +        "cat",  | 
 | 74 | +        "deer",  | 
 | 75 | +        "dog",  | 
 | 76 | +        "frog",  | 
 | 77 | +        "horse",  | 
 | 78 | +        "ship",  | 
 | 79 | +        "truck",  | 
 | 80 | +    )  | 
 | 81 | + | 
 | 82 | +    model.eval()  | 
 | 83 | + | 
 | 84 | +    correct = 0  | 
 | 85 | +    total = 0  | 
 | 86 | +    class_correct = list(0.0 for i in range(10))  | 
 | 87 | +    class_total = list(0.0 for i in range(10))  | 
 | 88 | +    with torch.no_grad():  | 
 | 89 | +        for data in test_loader:  | 
 | 90 | +            images, labels = data[0].to(device), data[1].to(device)  | 
 | 91 | +            outputs = model(images)  | 
 | 92 | +            _, predicted = torch.max(outputs.data, 1)  | 
 | 93 | +            total += labels.size(0)  | 
 | 94 | +            correct += (predicted == labels).sum().item()  | 
 | 95 | +            c = (predicted == labels).squeeze()  | 
 | 96 | +            for i in range(10):  | 
 | 97 | +                label = labels[i]  | 
 | 98 | +                class_correct[label] += c[i].item()  | 
 | 99 | +                class_total[label] += 1  | 
 | 100 | + | 
 | 101 | +    # print total test set accuracy  | 
 | 102 | +    print(  | 
 | 103 | +        "Accuracy of the network on the 10000 test images: %d %%"  | 
 | 104 | +        % (100 * correct / total)  | 
 | 105 | +    )  | 
 | 106 | + | 
 | 107 | +    # print test accuracy for each of the classes  | 
 | 108 | +    for i in range(10):  | 
 | 109 | +        print(  | 
 | 110 | +            "Accuracy of %5s : %2d %%"  | 
 | 111 | +            % (classes[i], 100 * class_correct[i] / class_total[i])  | 
 | 112 | +        )  | 
 | 113 | + | 
 | 114 | + | 
 | 115 | +def main(args):  | 
 | 116 | +    # get PyTorch environment variables  | 
 | 117 | +    world_size = int(os.environ["WORLD_SIZE"])  | 
 | 118 | +    rank = int(os.environ["RANK"])  | 
 | 119 | +    local_rank = int(os.environ["LOCAL_RANK"])  | 
 | 120 | + | 
 | 121 | +    distributed = world_size > 1  | 
 | 122 | + | 
 | 123 | +    # set device  | 
 | 124 | +    if distributed:  | 
 | 125 | +        device = torch.device("cuda", local_rank)  | 
 | 126 | +    else:  | 
 | 127 | +        device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")  | 
 | 128 | + | 
 | 129 | +    # initialize distributed process group using default env:// method  | 
 | 130 | +    if distributed:  | 
 | 131 | +        torch.distributed.init_process_group(backend="nccl")  | 
 | 132 | + | 
 | 133 | +    # define train and test dataset DataLoaders  | 
 | 134 | +    transform = transforms.Compose(  | 
 | 135 | +        [transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]  | 
 | 136 | +    )  | 
 | 137 | + | 
 | 138 | +    train_set = torchvision.datasets.CIFAR10(  | 
 | 139 | +        root=args.data_dir, train=True, download=False, transform=transform  | 
 | 140 | +    )  | 
 | 141 | + | 
 | 142 | +    if distributed:  | 
 | 143 | +        train_sampler = torch.utils.data.distributed.DistributedSampler(train_set)  | 
 | 144 | +    else:  | 
 | 145 | +        train_sampler = None  | 
 | 146 | + | 
 | 147 | +    train_loader = torch.utils.data.DataLoader(  | 
 | 148 | +        train_set,  | 
 | 149 | +        batch_size=args.batch_size,  | 
 | 150 | +        shuffle=(train_sampler is None),  | 
 | 151 | +        num_workers=args.workers,  | 
 | 152 | +        sampler=train_sampler,  | 
 | 153 | +    )  | 
 | 154 | + | 
 | 155 | +    test_set = torchvision.datasets.CIFAR10(  | 
 | 156 | +        root=args.data_dir, train=False, download=False, transform=transform  | 
 | 157 | +    )  | 
 | 158 | +    test_loader = torch.utils.data.DataLoader(  | 
 | 159 | +        test_set, batch_size=args.batch_size, shuffle=False, num_workers=args.workers  | 
 | 160 | +    )  | 
 | 161 | + | 
 | 162 | +    model = Net().to(device)  | 
 | 163 | + | 
 | 164 | +    # wrap model with DDP  | 
 | 165 | +    if distributed:  | 
 | 166 | +        model = nn.parallel.DistributedDataParallel(  | 
 | 167 | +            model, device_ids=[local_rank], output_device=local_rank  | 
 | 168 | +        )  | 
 | 169 | + | 
 | 170 | +    # define loss function and optimizer  | 
 | 171 | +    criterion = nn.CrossEntropyLoss()  | 
 | 172 | +    optimizer = optim.SGD(  | 
 | 173 | +        model.parameters(), lr=args.learning_rate, momentum=args.momentum  | 
 | 174 | +    )  | 
 | 175 | + | 
 | 176 | +    # train the model  | 
 | 177 | +    for epoch in range(args.epochs):  | 
 | 178 | +        print("Rank %d: Starting epoch %d" % (rank, epoch))  | 
 | 179 | +        if distributed:  | 
 | 180 | +            train_sampler.set_epoch(epoch)  | 
 | 181 | +        model.train()  | 
 | 182 | +        train(  | 
 | 183 | +            train_loader,  | 
 | 184 | +            model,  | 
 | 185 | +            criterion,  | 
 | 186 | +            optimizer,  | 
 | 187 | +            epoch,  | 
 | 188 | +            device,  | 
 | 189 | +            args.print_freq,  | 
 | 190 | +            rank,  | 
 | 191 | +        )  | 
 | 192 | + | 
 | 193 | +    print("Rank %d: Finished Training" % (rank))  | 
 | 194 | + | 
 | 195 | +    if not distributed or rank == 0:  | 
 | 196 | +        os.makedirs(args.output_dir, exist_ok=True)  | 
 | 197 | +        model_path = os.path.join(args.output_dir, "cifar_net.pt")  | 
 | 198 | +        torch.save(model.state_dict(), model_path)  | 
 | 199 | + | 
 | 200 | +        # evaluate on full test dataset  | 
 | 201 | +        evaluate(test_loader, model, device)  | 
 | 202 | + | 
 | 203 | + | 
 | 204 | +if __name__ == "__main__":  | 
 | 205 | +    # setup argparse  | 
 | 206 | +    parser = argparse.ArgumentParser()  | 
 | 207 | +    parser.add_argument(  | 
 | 208 | +        "--data-dir", type=str, help="directory containing CIFAR-10 dataset"  | 
 | 209 | +    )  | 
 | 210 | +    parser.add_argument("--epochs", default=10, type=int, help="number of epochs")  | 
 | 211 | +    parser.add_argument(  | 
 | 212 | +        "--batch-size",  | 
 | 213 | +        default=16,  | 
 | 214 | +        type=int,  | 
 | 215 | +        help="mini batch size for each gpu/process",  | 
 | 216 | +    )  | 
 | 217 | +    parser.add_argument(  | 
 | 218 | +        "--workers",  | 
 | 219 | +        default=2,  | 
 | 220 | +        type=int,  | 
 | 221 | +        help="number of data loading workers for each gpu/process",  | 
 | 222 | +    )  | 
 | 223 | +    parser.add_argument(  | 
 | 224 | +        "--learning-rate", default=0.001, type=float, help="learning rate"  | 
 | 225 | +    )  | 
 | 226 | +    parser.add_argument("--momentum", default=0.9, type=float, help="momentum")  | 
 | 227 | +    parser.add_argument(  | 
 | 228 | +        "--output-dir", default="outputs", type=str, help="directory to save model to"  | 
 | 229 | +    )  | 
 | 230 | +    parser.add_argument(  | 
 | 231 | +        "--print-freq",  | 
 | 232 | +        default=200,  | 
 | 233 | +        type=int,  | 
 | 234 | +        help="frequency of printing training statistics",  | 
 | 235 | +    )  | 
 | 236 | +    args = parser.parse_args()  | 
 | 237 | + | 
 | 238 | +    main(args)  | 
0 commit comments