在 Google Colab 上与在本地机器上训练 DeepLab ResNet V3 之间的巨大差异

Question

我正在尝试训练 Deeplab Resnet V3 对自定义数据集执行语义分割。 我一直在我的本地机器上工作，但是我的 GPU 只是一个小型 Quadro T1000，所以我决定将我的 model 移到 Google Colab 上，以利用他们的 Z52F9EC21735243AD9917CDA3CA077D32 实例并获得更好的结果。

虽然我得到了我希望的速度提升，但与我的本地机器相比，我在 colab 上的训练损失大不相同。 我复制并粘贴了完全相同的代码，所以我能找到的唯一区别在于数据集中。 我使用的是完全相同的数据集，除了 colab 上的数据集是 Google Drive 上本地数据集的副本。 我注意到 Windows 上的 Drive orders 文件不同，但我看不出这是一个问题，因为我随机打乱了数据集。 我知道这些随机分裂可能会导致输出的微小差异，但是训练损失的大约 10 倍的差异是没有意义的。

我还尝试使用不同的随机种子、不同的批量大小、不同的 train_test_split 参数以及将优化器从 SGD 更改为 Adam 在 colab 上运行该版本，但是，这仍然会导致 model 很早就收敛，损失约为 0.5。

这是我的代码：

import torch
from torch.utils import data
from torchvision import transforms
from customdatasets import SegmentationDataSet
import pathlib
from sklearn.model_selection import train_test_split
from customtransforms import Compose, AlbuSeg2d, DenseTarget
from customtransforms import MoveAxis, Normalize01, Resize
import albumentations
import matplotlib.pyplot as plt
import time
import GPUtil



def get_filenames_of_path(path: pathlib.Path, ext: str = '*'):
    """Returns a list of files in a directory/path. Uses pathlib."""
    filenames = [file for file in path.glob(ext) if file.is_file()]
    return filenames


if __name__ == '__main__':

    root = pathlib.Path.cwd() / 'train'
    inputs = get_filenames_of_path(root / 'input')
    targets = get_filenames_of_path(root / 'target')

    

# training transformations and augmentations
    transforms_training = Compose([
        Resize(input_size=(128, 128, 3), target_size=(128, 128)),
        AlbuSeg2d(albu=albumentations.HorizontalFlip(p=0.5)),
        MoveAxis(),
        Normalize01()
    ])
# validation transformations
    transforms_validation = Compose([
        Resize(input_size=(128, 128, 3), target_size=(128, 128)),
        MoveAxis(),
        Normalize01()
    ])
    if torch.cuda.is_available():
        device = torch.device('cuda')
    else:
        device = torch.device('cpu')

    random_seed = 142
    train_size = 0.8

    inputs_train, inputs_valid = train_test_split(
        inputs,
        random_state=random_seed,
        train_size=train_size,
        shuffle=True)
    targets_train, targets_valid = train_test_split(
        targets,
        random_state=random_seed,
        train_size=train_size,
        shuffle=True)

    dataset_train = SegmentationDataSet(inputs=inputs_train,
                                    targets=targets_train,
                                    transform=transforms_training,
                                    device=device)

    dataset_valid = SegmentationDataSet(inputs=inputs_valid,
                                    targets=targets_valid,
                                    transform=transforms_validation,
                                    device=device)


    dataloader_training = data.DataLoader(dataset=dataset_train,
                                      batch_size=15,
                                      shuffle=True,
                                      num_workers=4,
                                      pin_memory=True)

    dataloader_validation = data.DataLoader(dataset=dataset_valid,
                                        batch_size=15,
                                        shuffle=True,
                                        num_workers=4,
                                        pin_memory=True)


    model = torch.hub.load('pytorch/vision:v0.6.0', 'deeplabv3_resnet101', pretrained=False)


    criterion = torch.nn.CrossEntropyLoss()

    model = model.to(device)

    optimizer = torch.optim.SGD(model.parameters(), lr=0.001, momentum=0.99)


    epochs = 10
    steps = 0
    running_loss = 0
    print_every = 10
    train_losses, valid_losses = [], []

    start_time = time.time()
    prev_time = time.time()


    for epoch in range(epochs):
        #Training
        for inputs, labels in dataloader_training:
            steps += 1
            inputs, labels = inputs.to(device, non_blocking=True), labels.to(device,non_blocking=True)
            optimizer.zero_grad()
            logps = model(inputs)
            loss = criterion(logps['out'], labels)
            loss.backward()
            optimizer.step()
            running_loss += loss.item()

            if steps % print_every == 0:
                train_losses.append(running_loss / len(dataloader_training))
                epoch_time = time.time()
                elasped_time = epoch_time - prev_time
                prev_time = epoch_time
                print(f"Epoch {epoch + 1}/{epochs}.. "
                    f"Train loss: {running_loss / print_every:.3f}.. "
                    f"Elapsed time: {elasped_time}")

                running_loss = 0
                model.train()
        # Evaluation
        valid_loss = 0
        accuracy = 0
        model.eval()
        with torch.no_grad():
            for inputs, labels in dataloader_validation:
                inputs, labels = inputs.to(device, non_blocking=True), labels.to(device, non_blocking=True)
                logps = model.forward(inputs)
                batch_loss = criterion(logps['out'], labels)
                valid_loss += batch_loss.item()

                ps = torch.exp(logps['out'])
                top_p, top_class = ps.topk(1, dim=1)
                equals = top_class == labels.view(*top_class.shape)
                accuracy += torch.mean(equals.type(torch.FloatTensor)).item()
        valid_losses.append(valid_loss / len(dataloader_validation))
        print(f"Epoch {epoch + 1}/{epochs}.. "
            f"Validation loss: {valid_loss / len(dataloader_training):.3f}.. "
            f"Validation accuracy: {accuracy / len(dataloader_training):.3f} ")
        model.train()
    torch.save(model, 'model.pth')

    end_time = time.time()
    total_time = end_time - start_time
    print("Total Time: ", total_time)
    plt.plot(train_losses, label='Training loss')
    plt.plot(valid_losses, label='Validation loss')
    plt.legend(frameon=False)
    plt.show()

这是 Colab 上一个 epoch 的 output：

Epoch 1/10.. Train loss: 2.080.. Elapsed time: 12.156640768051147
Epoch 1/10.. Train loss: 1.231.. Elapsed time: 8.76858925819397
Epoch 1/10.. Train loss: 1.051.. Elapsed time: 8.315532445907593
Epoch 1/10.. Train loss: 0.890.. Elapsed time: 8.249168634414673
Epoch 1/10.. Train loss: 0.839.. Elapsed time: 8.248667478561401
Epoch 1/10.. Train loss: 0.807.. Elapsed time: 8.120820999145508
Epoch 1/10.. Train loss: 0.742.. Elapsed time: 8.298616886138916
Epoch 1/10.. Train loss: 0.726.. Elapsed time: 8.170734167098999
Epoch 1/10.. Train loss: 0.677.. Elapsed time: 8.221246004104614
Epoch 1/10.. Train loss: 0.698.. Elapsed time: 8.124614000320435
Epoch 1/10.. Train loss: 0.675.. Elapsed time: 8.197462558746338
Epoch 1/10.. Train loss: 0.682.. Elapsed time: 8.263437509536743
Epoch 1/10.. Train loss: 0.626.. Elapsed time: 8.156179189682007
Epoch 1/10.. Train loss: 0.632.. Elapsed time: 8.268096446990967
Epoch 1/10.. Train loss: 0.616.. Elapsed time: 8.214547872543335
Epoch 1/10.. Train loss: 0.585.. Elapsed time: 8.31475019454956
Epoch 1/10.. Train loss: 0.598.. Elapsed time: 8.388074398040771
Epoch 1/10.. Train loss: 0.626.. Elapsed time: 8.179292440414429
Epoch 1/10.. Train loss: 0.612.. Elapsed time: 8.252359390258789
Epoch 1/10.. Train loss: 0.592.. Elapsed time: 8.284745693206787
Epoch 1/10.. Train loss: 0.597.. Elapsed time: 8.31213927268982
Epoch 1/10.. Train loss: 0.566.. Elapsed time: 8.164374113082886
Epoch 1/10.. Train loss: 0.556.. Elapsed time: 8.300082206726074
Epoch 1/10.. Train loss: 0.568.. Elapsed time: 8.26304841041565
Epoch 1/10.. Train loss: 0.572.. Elapsed time: 8.309881448745728
Epoch 1/10.. Train loss: 0.586.. Elapsed time: 8.211671352386475
Epoch 1/10.. Train loss: 0.586.. Elapsed time: 8.321797609329224
Epoch 1/10.. Train loss: 0.535.. Elapsed time: 8.318871021270752
Epoch 1/10.. Train loss: 0.543.. Elapsed time: 8.152915239334106
Epoch 1/10.. Train loss: 0.569.. Elapsed time: 8.251380205154419
Epoch 1/10.. Train loss: 0.526.. Elapsed time: 8.29153847694397
Epoch 1/10.. Train loss: 0.565.. Elapsed time: 8.15071702003479
Epoch 1/10.. Train loss: 0.542.. Elapsed time: 8.253364562988281
Epoch 1/10.. Validation loss: 0.182.. Validation accuracy: 0.271 

这是我本地机器上的 output：

Epoch 1/10.. Train loss: 2.932.. Elapsed time: 32.148621797561646
Epoch 1/10.. Train loss: 1.852.. Elapsed time: 14.120505809783936
Epoch 1/10.. Train loss: 0.887.. Elapsed time: 14.210048198699951
Epoch 1/10.. Train loss: 0.618.. Elapsed time: 14.23294186592102
Epoch 1/10.. Train loss: 0.549.. Elapsed time: 14.212541103363037
Epoch 1/10.. Train loss: 0.519.. Elapsed time: 14.047481775283813
Epoch 1/10.. Train loss: 0.506.. Elapsed time: 14.060708045959473
Epoch 1/10.. Train loss: 0.347.. Elapsed time: 14.301624059677124
Epoch 1/10.. Train loss: 0.399.. Elapsed time: 13.9844491481781
Epoch 1/10.. Train loss: 0.361.. Elapsed time: 13.957871913909912
Epoch 1/10.. Train loss: 0.305.. Elapsed time: 14.164010763168335
Epoch 1/10.. Train loss: 0.296.. Elapsed time: 14.001536846160889
Epoch 1/10.. Train loss: 0.298.. Elapsed time: 14.019971132278442
Epoch 1/10.. Train loss: 0.271.. Elapsed time: 13.951345443725586
Epoch 1/10.. Train loss: 0.252.. Elapsed time: 14.037938594818115
Epoch 1/10.. Train loss: 0.283.. Elapsed time: 13.944657564163208
Epoch 1/10.. Train loss: 0.299.. Elapsed time: 13.977224826812744
Epoch 1/10.. Train loss: 0.219.. Elapsed time: 13.941975355148315
Epoch 1/10.. Train loss: 0.242.. Elapsed time: 13.936140060424805
Epoch 1/10.. Train loss: 0.244.. Elapsed time: 13.942122459411621
Epoch 1/10.. Train loss: 0.216.. Elapsed time: 13.960899114608765
Epoch 1/10.. Train loss: 0.186.. Elapsed time: 13.956881523132324
Epoch 1/10.. Train loss: 0.241.. Elapsed time: 13.944581985473633
Epoch 1/10.. Train loss: 0.203.. Elapsed time: 13.934357404708862
Epoch 1/10.. Train loss: 0.189.. Elapsed time: 13.938358306884766
Epoch 1/10.. Train loss: 0.181.. Elapsed time: 13.944468021392822
Epoch 1/10.. Train loss: 0.186.. Elapsed time: 13.946297407150269
Epoch 1/10.. Train loss: 0.164.. Elapsed time: 13.940366744995117
Epoch 1/10.. Train loss: 0.165.. Elapsed time: 13.938241720199585
Epoch 1/10.. Train loss: 0.176.. Elapsed time: 14.015569925308228
Epoch 1/10.. Train loss: 0.165.. Elapsed time: 14.019208669662476
Epoch 1/10.. Train loss: 0.175.. Elapsed time: 14.149503469467163
Epoch 1/10.. Train loss: 0.159.. Elapsed time: 14.128302097320557
Epoch 1/10.. Train loss: 0.155.. Elapsed time: 13.935027837753296
Epoch 1/10.. Train loss: 0.137.. Elapsed time: 13.937382221221924
Epoch 1/10.. Train loss: 0.127.. Elapsed time: 13.929635524749756
Epoch 1/10.. Train loss: 0.133.. Elapsed time: 13.935472011566162
Epoch 1/10.. Train loss: 0.152.. Elapsed time: 13.922808647155762
Epoch 1/10.. Validation loss: 0.032.. Validation accuracy: 0.239

我不会粘贴更多，因为它很长并且需要一段时间才能运行，但到第三个时期结束时，Colab model 的损失仍然在 0.5 左右反弹，而在本地它达到 0.02。

如果有人可以帮助我解决这个问题，将不胜感激。

Answer 1

我通过将训练数据解压缩到 Google Drive 并从那里读取文件而不是使用 Colab 命令将文件夹直接解压缩到我的工作区来解决了这个问题。 我完全不知道为什么会导致问题； 对图像及其相应的张量进行快速目视检查看起来不错，但我无法通过 6,000 张左右的图像中的每一张来检查每一张。 如果有人知道为什么这会导致问题，请告诉我！