Python 使用来自谷歌云存储的数据在谷歌计算引擎上培训Resnet时的可变性能

我正在尝试从标记的Google Streeview数据中对大量作物训练resnet18图像分类器。我跟在后面。我有两个数据集，一个约20k图像，一个约100k图像。这两个数据集以相同的格式存储，并且都已上载到各自的Google云存储桶中。然后，我使用带有

--implicit dirs

标志的

gcsfuse

将这两个bucket装载到VM的主目录中

然后，我在我的Google Compute Engine VM上运行我的

train.py

文件，它是从Google的云市场上创建的。虚拟机有一个vCPU、一个Nvidia Tesla K80 GPU、3.75gb内存和一个100gb永久磁盘

当我运行培训脚本时，除了将dataset_dir变量指向VM上正确的

gcsfuse

-挂载目录外，我不做任何更改

当我在100k crops目录上运行

train.py

时，它运行得相对较快，单个历元大约需要30分钟。当它运行时，我跳入

top

，CPU利用率相当高，保持在90%左右

但是，使用相同的VM，当我在20K crops目录上运行

train.py

时，运行速度要慢得多，一个历元需要6-7个小时，尽管数据集的大小较小。在这种情况下，CPU利用率从未超过5%

我不知道是什么原因导致了减速，因为两次运行之间没有什么不同（据我所知），除了数据集，它们的格式都相同。我使用相同的

pytorch

dataloader和相同数量的线程。两个GCS存储桶位于同一区域，

us-west1

，这与我的VM实例位于同一区域

似乎有可能一个桶相对于另一个桶受到IO限制，但我不知道为什么

任何想法都很感激

下面是我的

train.py

文件

import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
import numpy as np
import torchvision
from torchvision import datasets, models, transforms
import matplotlib.pyplot as plt
import time
import os
import copy
from collections import defaultdict



data_transforms = {
    'Test': transforms.Compose([
        transforms.RandomResizedCrop(224),
        transforms.RandomHorizontalFlip(),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ]),
    'Val': transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ]),
}


data_dir = 'home/gweld/sliding_window_dataset/'
image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x),
                                          data_transforms[x])
                  for x in ['Test', 'Val']}
dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=4,
                                             shuffle=True, num_workers=4)
              for x in ['Test', 'Val']}
dataset_sizes = {x: len(image_datasets[x]) for x in ['Test', 'Val']}
class_names = image_datasets['Test'].classes

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")




def train_model(model, criterion, optimizer, scheduler, num_epochs=25):
    since = time.time()

    best_model_wts = copy.deepcopy(model.state_dict())
    best_acc = 0.0

    for epoch in range(num_epochs):
        print('Epoch {}/{}'.format(epoch, num_epochs - 1))
        print('-' * 10)

        # Each epoch has a training and validation phase
        for phase in ['Test', 'Val']:
            if phase == 'Test':
                scheduler.step()
                model.train()  # Set model to training mode
            else:
                model.eval()   # Set model to evaluate mode

            running_loss = 0.0
            running_corrects = 0

            class_corrects = defaultdict(int)
            class_totals   = defaultdict(int)

            # Iterate over data.
            for inputs, labels in dataloaders[phase]:
                inputs = inputs.to(device)
                labels = labels.to(device)

                # zero the parameter gradients
                optimizer.zero_grad()

                # forward
                # track history if only in train
                with torch.set_grad_enabled(phase == 'Test'):
                    outputs = model(inputs)
                    _, preds = torch.max(outputs, 1)
                    loss = criterion(outputs, labels)

                    # backward + optimize only if in training phase
                    if phase == 'Test':
                        loss.backward()
                        optimizer.step()

                # statistics
                running_loss += loss.item() * inputs.size(0)
                running_corrects += torch.sum(preds == labels.data)

                for index, pred in enumerate(preds):
                    actual = labels.data[index]
                    class_name = class_names[actual]

                    if actual == pred: class_corrects[class_name] += 1
                    class_totals[class_name] += 1

            epoch_loss = running_loss / dataset_sizes[phase]
            epoch_acc = running_corrects.double() / dataset_sizes[phase]

            print('{} Loss: {:.4f} Acc: {:.4f}'.format(
                phase, epoch_loss, epoch_acc))

            if phase == 'Val':
                print("Validation Class Accuracies")

                for class_name in class_totals:
                    class_acc = float(class_corrects[class_name])
                    class_acc = class_acc/class_totals[class_name]

                    print("{:20}{}%".format(class_name, 100*class_acc))
                print("\n")

            # deep copy the model
            if phase == 'Val' and epoch_acc > best_acc:
                best_acc = epoch_acc
                best_model_wts = copy.deepcopy(model.state_dict())

        print()

    time_elapsed = time.time() - since
    print('Training complete in {:.0f}m {:.0f}s'.format(
        time_elapsed // 60, time_elapsed % 60))
    print('Best val Acc: {:4f}'.format(best_acc))

    # load best model weights
    model.load_state_dict(best_model_wts)
    return model




model_ft = models.resnet18(pretrained=True)
num_ftrs = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_ftrs, 5) # last arg here, # classes? -gw

model_ft = model_ft.to(device)

criterion = nn.CrossEntropyLoss()

# Observe that all parameters are being optimized
optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9)

# Decay LR by a factor of 0.1 every 7 epochs
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)


# Train and evaluate
# ^^^^^^^^^^^^^^^^^^

print('Beginning Training on {} train and {} val images.'.format(dataset_sizes['Test'], dataset_sizes['Val']))


model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler,
                       num_epochs=25)




torch.save(model_ft.state_dict(), 'models/test_run_resnet18.pt')

---

两个数据集在同一个存储桶中？如果不选中gsutil ls-L-b gs://| grep Storage。如果您在两个存储桶上运行相同的CPU/GPU ML进程，并且其中一个进程的速度要慢得多，那么很有可能是某种原因导致了两个存储桶之间的IO性能不同。慢的是使用近线/冷线的吗？是否启用了fast？FUSE文件系统的安装/配置方式有什么不同吗？