PyTorch一小时掌握之迁移学习篇-巨人网络通讯

主页 > 知识库 > PyTorch一小时掌握之迁移学习篇

PyTorch一小时掌握之迁移学习篇

概述

迁移学习 (Transfer Learning) 是把已学训练好的模型参数用作新训练模型的起始参数. 迁移学习是深度学习中非常重要和常用的一个策略.

为什么使用迁移学习

更好的结果

迁移学习 (Transfer Learning) 可以帮助我们得到更好的结果.

当我们手上的数据比较少的时候, 训练非常容易造成过拟合的现象. 使用迁移学习可以帮助我们通过更少的训练数据达到更好的效果. 使得模型的泛化能力更强, 训练过程更稳定.

节省时间

迁移学习 (Transfer Learning) 可以帮助我们节省时间.

通过迁徙学习, 我们站在了巨人的肩膀上. 利用前人花大量时间训练好的参数, 能帮助我们在模型的训练上节省大把的时间.

加载模型

首先我们需要加载模型, 并指定层数. 常用的模型有:

VGG
ResNet
SqueezeNet
DenseNet
Inception
GoogLeNet
ShuffleNet
MobileNet

官网 API

ResNet152

我们将使用 ResNet 152 和 CIFAR 100 来举例.

冻层实现

def set_parameter_requires_grad(model, feature_extracting):
    """
    是否保留梯度, 实现冻层
    :param model: 模型
    :param feature_extracting: 是否冻层
    :return: 无返回值
    """
    if feature_extracting:  # 如果冻层
        for param in model.parameters():  # 遍历每个权重参数
            param.requires_grad = False  # 保留梯度为False

模型初始化

def initialize_model(model_name, num_classes, feature_exact, use_pretrained=True):
    """
    初始化模型
    :param model_name: 模型名字
    :param num_classes: 类别数
    :param feature_exact: 是否冻层
    :param use_pretrained: 是否下载模型
    :return: 返回模型,
    """

    model_ft = None

    if model_name == "resnet":
        """Resnet152"""

        # 加载模型
        model_ft = models.resnet152(pretrained=use_pretrained)  # 下载参数
        set_parameter_requires_grad(model_ft, feature_exact)  # 冻层

        # 修改全连接层
        num_features = model_ft.fc.in_features
        model_ft.fc = torch.nn.Sequential(
            torch.nn.Linear(num_features, num_classes),
            torch.nn.LogSoftmax(dim=1)
        )

    # 返回初始化好的模型
    return model_ft

获取需更新参数

def parameter_to_update(model):
    """
    获取需要更新的参数
    :param model: 模型
    :return: 需要更新的参数列表
    """

    print("Params to learn")
    param_array = model.parameters()

    if feature_exact:
        param_array = []
        for name, param, in model.named_parameters():
            if param.requires_grad == True:
                param_array.append(param)
                print("\t", name)
    else:
        for name, param, in model.named_parameters():
            if param.requires_grad == True:
                print("\t", name)

    return param_array

训练模型

def train_model(model, dataloaders, citerion, optimizer, filename, num_epochs=25):
    # 获取起始时间
    since = time.time()

    # 初始化参数
    best_acc = 0
    val_acc_history = []
    train_acc_history = []
    train_losses = []
    valid_losses = []
    LRs = [optimizer.param_groups[0]["lr"]]
    best_model_weights = copy.deepcopy(model.state_dict())

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

        # 训练和验证
        for phase in ["train", "valid"]:
            if phase == "train":
                model.train()  # 训练
            else:
                model.eval()  # 验证

            running_loss = 0.0
            running_corrects = 0

            # 遍历数据
            for inputs, labels in dataloaders[phase]:
                inputs = inputs.to(device)
                labels = labels.to(device)

                # 梯度清零
                optimizer.zero_grad()

                # 只有训练的时候计算和更新梯度
                with torch.set_grad_enabled(phase == "train"):
                    outputs = model(inputs)
                    _, preds = torch.max(outputs, 1)

                    # 计算损失
                    loss = criterion(outputs, labels)

                    # 训练阶段更新权重
                    if phase == "train":
                        loss.backward()
                        optimizer.step()

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

            epoch_loss = running_loss / len(dataloaders[phase].dataset)
            epoch_acc = running_corrects.double() / len(dataloaders[phase].dataset)

            time_eplased = time.time() - since
            print("Time elapsed {:.0f}m {:.0f}s".format(time_eplased // 60, time_eplased % 60))
            print("{} Loss: {:.4f} Acc: {:.4f}".format(phase, epoch_loss, epoch_acc))

            # 得到最好的模型
            if phase == "valid" and epoch_acc > best_acc:
                best_acc = epoch_acc
                best_model_weights = copy.deepcopy(model.state_dict())
                state = {
                    "state_dict": model.state_dict(),
                    "best_acc": best_acc,
                    "optimizer": optimizer.state_dict(),
                }
                torch.save(state, filename)
            if phase == "valid":
                val_acc_history.append(epoch_acc)
                valid_losses.append(epoch_loss)
                scheduler.step(epoch_loss)
            if phase == "train":
                train_acc_history.append(epoch_acc)
                train_losses.append(epoch_loss)

        print("Optimizer learning rate: {:.7f}".format(optimizer.param_groups[0]["lr"]))
        LRs.append(optimizer.param_groups[0]["lr"])
        print()

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

    # 训练完后用最好的一次当做模型最终的结果
    model.load_state_dict(best_model_weights)

    # 返回
    return model, val_acc_history, train_acc_history, valid_losses, train_losses, LRs

获取数据

def get_data():
    """获取数据"""

    # 获取测试集
    train = torchvision.datasets.CIFAR100(root="./mnt", train=True, download=True,
                                          transform=torchvision.transforms.Compose([
                                              torchvision.transforms.ToTensor(),  # 转换成张量
                                              torchvision.transforms.Normalize((0.1307,), (0.3081,))  # 标准化
                                          ]))
    train_loader = DataLoader(train, batch_size=batch_size)  # 分割测试集

    # 获取测试集
    test = torchvision.datasets.CIFAR100(root="./mnt", train=False, download=True,
                                         transform=torchvision.transforms.Compose([
                                             torchvision.transforms.ToTensor(),  # 转换成张量
                                             torchvision.transforms.Normalize((0.1307,), (0.3081,))  # 标准化
                                         ]))
    test_loader = DataLoader(test, batch_size=batch_size)  # 分割训练

    data_loader = {"train": train_loader, "valid": test_loader}

    # 返回分割好的训练集和测试集
    return data_loader

完整代码

完整代码:

import copy
import torch
from torch.utils.data import DataLoader
import time
from torchsummary import summary
import torchvision
import torchvision.models as models


def set_parameter_requires_grad(model, feature_extracting):
    """
    是否保留梯度, 实现冻层
    :param model: 模型
    :param feature_extracting: 是否冻层
    :return: 无返回值
    """
    if feature_extracting:  # 如果冻层
        for param in model.parameters():  # 遍历每个权重参数
            param.requires_grad = False  # 保留梯度为False


def initialize_model(model_name, num_classes, feature_exact, use_pretrained=True):
    """
    初始化模型
    :param model_name: 模型名字
    :param num_classes: 类别数
    :param feature_exact: 是否冻层
    :param use_pretrained: 是否下载模型
    :return: 返回模型,
    """

    model_ft = None

    if model_name == "resnet":
        """Resnet152"""

        # 加载模型
        model_ft = models.resnet152(pretrained=use_pretrained)  # 下载参数
        set_parameter_requires_grad(model_ft, feature_exact)  # 冻层

        # 修改全连接层
        num_features = model_ft.fc.in_features
        model_ft.fc = torch.nn.Sequential(
            torch.nn.Linear(num_features, num_classes),
            torch.nn.LogSoftmax(dim=1)
        )

    # 返回初始化好的模型
    return model_ft


def parameter_to_update(model):
    """
    获取需要更新的参数
    :param model: 模型
    :return: 需要更新的参数列表
    """

    print("Params to learn")
    param_array = model.parameters()

    if feature_exact:
        param_array = []
        for name, param, in model.named_parameters():
            if param.requires_grad == True:
                param_array.append(param)
                print("\t", name)
    else:
        for name, param, in model.named_parameters():
            if param.requires_grad == True:
                print("\t", name)

    return param_array


def train_model(model, dataloaders, citerion, optimizer, filename, num_epochs=25):
    # 获取起始时间
    since = time.time()

    # 初始化参数
    best_acc = 0
    val_acc_history = []
    train_acc_history = []
    train_losses = []
    valid_losses = []
    LRs = [optimizer.param_groups[0]["lr"]]
    best_model_weights = copy.deepcopy(model.state_dict())

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

        # 训练和验证
        for phase in ["train", "valid"]:
            if phase == "train":
                model.train()  # 训练
            else:
                model.eval()  # 验证

            running_loss = 0.0
            running_corrects = 0

            # 遍历数据
            for inputs, labels in dataloaders[phase]:
                inputs = inputs.to(device)
                labels = labels.to(device)

                # 梯度清零
                optimizer.zero_grad()

                # 只有训练的时候计算和更新梯度
                with torch.set_grad_enabled(phase == "train"):
                    outputs = model(inputs)
                    _, preds = torch.max(outputs, 1)

                    # 计算损失
                    loss = criterion(outputs, labels)

                    # 训练阶段更新权重
                    if phase == "train":
                        loss.backward()
                        optimizer.step()

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

            epoch_loss = running_loss / len(dataloaders[phase].dataset)
            epoch_acc = running_corrects.double() / len(dataloaders[phase].dataset)

            time_eplased = time.time() - since
            print("Time elapsed {:.0f}m {:.0f}s".format(time_eplased // 60, time_eplased % 60))
            print("{} Loss: {:.4f} Acc: {:.4f}".format(phase, epoch_loss, epoch_acc))

            # 得到最好的模型
            if phase == "valid" and epoch_acc > best_acc:
                best_acc = epoch_acc
                best_model_weights = copy.deepcopy(model.state_dict())
                state = {
                    "state_dict": model.state_dict(),
                    "best_acc": best_acc,
                    "optimizer": optimizer.state_dict(),
                }
                torch.save(state, filename)
            if phase == "valid":
                val_acc_history.append(epoch_acc)
                valid_losses.append(epoch_loss)
                scheduler.step(epoch_loss)
            if phase == "train":
                train_acc_history.append(epoch_acc)
                train_losses.append(epoch_loss)

        print("Optimizer learning rate: {:.7f}".format(optimizer.param_groups[0]["lr"]))
        LRs.append(optimizer.param_groups[0]["lr"])
        print()

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

    # 训练完后用最好的一次当做模型最终的结果
    model.load_state_dict(best_model_weights)

    # 返回
    return model, val_acc_history, train_acc_history, valid_losses, train_losses, LRs


def get_data():
    """获取数据"""

    # 获取测试集
    train = torchvision.datasets.CIFAR100(root="./mnt", train=True, download=True,
                                          transform=torchvision.transforms.Compose([
                                              torchvision.transforms.ToTensor(),  # 转换成张量
                                              torchvision.transforms.Normalize((0.1307,), (0.3081,))  # 标准化
                                          ]))
    train_loader = DataLoader(train, batch_size=batch_size)  # 分割测试集

    # 获取测试集
    test = torchvision.datasets.CIFAR100(root="./mnt", train=False, download=True,
                                         transform=torchvision.transforms.Compose([
                                             torchvision.transforms.ToTensor(),  # 转换成张量
                                             torchvision.transforms.Normalize((0.1307,), (0.3081,))  # 标准化
                                         ]))
    test_loader = DataLoader(test, batch_size=batch_size)  # 分割训练

    data_loader = {"train": train_loader, "valid": test_loader}

    # 返回分割好的训练集和测试集
    return data_loader


# 超参数
filename = "checkpoint.pth"  # 模型保存
feature_exact = True  # 冻层
num_classes = 100  # 输出的类别数
batch_size = 1024  # 一次训练的样本数目
iteration_num = 10  # 迭代次数

# 获取模型
resnet152 = initialize_model(
    model_name="resnet",
    num_classes=num_classes,
    feature_exact=feature_exact,
    use_pretrained=True
)

# 是否使用GPU训练
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
if use_cuda: resnet152.cuda()  # GPU 计算
print("是否使用 GPU 加速:", use_cuda)

# 输出网络结构
print(summary(resnet152, (3, 32, 32)))

# 训练参数
params_to_update = parameter_to_update(resnet152)

# 优化器
optimizer = torch.optim.Adam(params_to_update, lr=0.01)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=10, gamma=0.1)  # 学习率每10个epoch衰减到原来的1/10
criterion = torch.nn.NLLLoss()

if __name__ == "__main__":
    data_loader = get_data()
    resnet152, val_acc_history, train_acc_history, valid_losses, train_losses, LRs = train_model(
        model=resnet152,
        dataloaders=data_loader,
        citerion=criterion,
        optimizer=optimizer,
        num_epochs=iteration_num,
        filename=filename
    )

输出结果:

是否使用 GPU 加速: True
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 16, 16] 9,408
BatchNorm2d-2 [-1, 64, 16, 16] 128
ReLU-3 [-1, 64, 16, 16] 0
MaxPool2d-4 [-1, 64, 8, 8] 0
Conv2d-5 [-1, 64, 8, 8] 4,096
BatchNorm2d-6 [-1, 64, 8, 8] 128
ReLU-7 [-1, 64, 8, 8] 0
Conv2d-8 [-1, 64, 8, 8] 36,864
BatchNorm2d-9 [-1, 64, 8, 8] 128
ReLU-10 [-1, 64, 8, 8] 0
Conv2d-11 [-1, 256, 8, 8] 16,384
BatchNorm2d-12 [-1, 256, 8, 8] 512
Conv2d-13 [-1, 256, 8, 8] 16,384
BatchNorm2d-14 [-1, 256, 8, 8] 512
ReLU-15 [-1, 256, 8, 8] 0
Bottleneck-16 [-1, 256, 8, 8] 0
Conv2d-17 [-1, 64, 8, 8] 16,384
BatchNorm2d-18 [-1, 64, 8, 8] 128
ReLU-19 [-1, 64, 8, 8] 0
Conv2d-20 [-1, 64, 8, 8] 36,864
BatchNorm2d-21 [-1, 64, 8, 8] 128
ReLU-22 [-1, 64, 8, 8] 0
Conv2d-23 [-1, 256, 8, 8] 16,384
BatchNorm2d-24 [-1, 256, 8, 8] 512
ReLU-25 [-1, 256, 8, 8] 0
Bottleneck-26 [-1, 256, 8, 8] 0
Conv2d-27 [-1, 64, 8, 8] 16,384
BatchNorm2d-28 [-1, 64, 8, 8] 128
ReLU-29 [-1, 64, 8, 8] 0
Conv2d-30 [-1, 64, 8, 8] 36,864
BatchNorm2d-31 [-1, 64, 8, 8] 128
ReLU-32 [-1, 64, 8, 8] 0
Conv2d-33 [-1, 256, 8, 8] 16,384
BatchNorm2d-34 [-1, 256, 8, 8] 512
ReLU-35 [-1, 256, 8, 8] 0
Bottleneck-36 [-1, 256, 8, 8] 0
Conv2d-37 [-1, 128, 8, 8] 32,768
BatchNorm2d-38 [-1, 128, 8, 8] 256
ReLU-39 [-1, 128, 8, 8] 0
Conv2d-40 [-1, 128, 4, 4] 147,456
BatchNorm2d-41 [-1, 128, 4, 4] 256
ReLU-42 [-1, 128, 4, 4] 0
Conv2d-43 [-1, 512, 4, 4] 65,536
BatchNorm2d-44 [-1, 512, 4, 4] 1,024
Conv2d-45 [-1, 512, 4, 4] 131,072
BatchNorm2d-46 [-1, 512, 4, 4] 1,024
ReLU-47 [-1, 512, 4, 4] 0
Bottleneck-48 [-1, 512, 4, 4] 0
Conv2d-49 [-1, 128, 4, 4] 65,536
BatchNorm2d-50 [-1, 128, 4, 4] 256
ReLU-51 [-1, 128, 4, 4] 0
Conv2d-52 [-1, 128, 4, 4] 147,456
BatchNorm2d-53 [-1, 128, 4, 4] 256
ReLU-54 [-1, 128, 4, 4] 0
Conv2d-55 [-1, 512, 4, 4] 65,536
BatchNorm2d-56 [-1, 512, 4, 4] 1,024
ReLU-57 [-1, 512, 4, 4] 0
Bottleneck-58 [-1, 512, 4, 4] 0
Conv2d-59 [-1, 128, 4, 4] 65,536
BatchNorm2d-60 [-1, 128, 4, 4] 256
ReLU-61 [-1, 128, 4, 4] 0
Conv2d-62 [-1, 128, 4, 4] 147,456
BatchNorm2d-63 [-1, 128, 4, 4] 256
ReLU-64 [-1, 128, 4, 4] 0
Conv2d-65 [-1, 512, 4, 4] 65,536
BatchNorm2d-66 [-1, 512, 4, 4] 1,024
ReLU-67 [-1, 512, 4, 4] 0
Bottleneck-68 [-1, 512, 4, 4] 0
Conv2d-69 [-1, 128, 4, 4] 65,536
BatchNorm2d-70 [-1, 128, 4, 4] 256
ReLU-71 [-1, 128, 4, 4] 0
Conv2d-72 [-1, 128, 4, 4] 147,456
BatchNorm2d-73 [-1, 128, 4, 4] 256
ReLU-74 [-1, 128, 4, 4] 0
Conv2d-75 [-1, 512, 4, 4] 65,536
BatchNorm2d-76 [-1, 512, 4, 4] 1,024
ReLU-77 [-1, 512, 4, 4] 0
Bottleneck-78 [-1, 512, 4, 4] 0
Conv2d-79 [-1, 128, 4, 4] 65,536
BatchNorm2d-80 [-1, 128, 4, 4] 256
ReLU-81 [-1, 128, 4, 4] 0
Conv2d-82 [-1, 128, 4, 4] 147,456
BatchNorm2d-83 [-1, 128, 4, 4] 256
ReLU-84 [-1, 128, 4, 4] 0
Conv2d-85 [-1, 512, 4, 4] 65,536
BatchNorm2d-86 [-1, 512, 4, 4] 1,024
ReLU-87 [-1, 512, 4, 4] 0
Bottleneck-88 [-1, 512, 4, 4] 0
Conv2d-89 [-1, 128, 4, 4] 65,536
BatchNorm2d-90 [-1, 128, 4, 4] 256
ReLU-91 [-1, 128, 4, 4] 0
Conv2d-92 [-1, 128, 4, 4] 147,456
BatchNorm2d-93 [-1, 128, 4, 4] 256
ReLU-94 [-1, 128, 4, 4] 0
Conv2d-95 [-1, 512, 4, 4] 65,536
BatchNorm2d-96 [-1, 512, 4, 4] 1,024
ReLU-97 [-1, 512, 4, 4] 0
Bottleneck-98 [-1, 512, 4, 4] 0
Conv2d-99 [-1, 128, 4, 4] 65,536
BatchNorm2d-100 [-1, 128, 4, 4] 256
ReLU-101 [-1, 128, 4, 4] 0
Conv2d-102 [-1, 128, 4, 4] 147,456
BatchNorm2d-103 [-1, 128, 4, 4] 256
ReLU-104 [-1, 128, 4, 4] 0
Conv2d-105 [-1, 512, 4, 4] 65,536
BatchNorm2d-106 [-1, 512, 4, 4] 1,024
ReLU-107 [-1, 512, 4, 4] 0
Bottleneck-108 [-1, 512, 4, 4] 0
Conv2d-109 [-1, 128, 4, 4] 65,536
BatchNorm2d-110 [-1, 128, 4, 4] 256
ReLU-111 [-1, 128, 4, 4] 0
Conv2d-112 [-1, 128, 4, 4] 147,456
BatchNorm2d-113 [-1, 128, 4, 4] 256
ReLU-114 [-1, 128, 4, 4] 0
Conv2d-115 [-1, 512, 4, 4] 65,536
BatchNorm2d-116 [-1, 512, 4, 4] 1,024
ReLU-117 [-1, 512, 4, 4] 0
Bottleneck-118 [-1, 512, 4, 4] 0
Conv2d-119 [-1, 256, 4, 4] 131,072
BatchNorm2d-120 [-1, 256, 4, 4] 512
ReLU-121 [-1, 256, 4, 4] 0
Conv2d-122 [-1, 256, 2, 2] 589,824
BatchNorm2d-123 [-1, 256, 2, 2] 512
ReLU-124 [-1, 256, 2, 2] 0
Conv2d-125 [-1, 1024, 2, 2] 262,144
BatchNorm2d-126 [-1, 1024, 2, 2] 2,048
Conv2d-127 [-1, 1024, 2, 2] 524,288
BatchNorm2d-128 [-1, 1024, 2, 2] 2,048
ReLU-129 [-1, 1024, 2, 2] 0
Bottleneck-130 [-1, 1024, 2, 2] 0
Conv2d-131 [-1, 256, 2, 2] 262,144
BatchNorm2d-132 [-1, 256, 2, 2] 512
ReLU-133 [-1, 256, 2, 2] 0
Conv2d-134 [-1, 256, 2, 2] 589,824
BatchNorm2d-135 [-1, 256, 2, 2] 512
ReLU-136 [-1, 256, 2, 2] 0
Conv2d-137 [-1, 1024, 2, 2] 262,144
BatchNorm2d-138 [-1, 1024, 2, 2] 2,048
ReLU-139 [-1, 1024, 2, 2] 0
Bottleneck-140 [-1, 1024, 2, 2] 0
Conv2d-141 [-1, 256, 2, 2] 262,144
BatchNorm2d-142 [-1, 256, 2, 2] 512
ReLU-143 [-1, 256, 2, 2] 0
Conv2d-144 [-1, 256, 2, 2] 589,824
BatchNorm2d-145 [-1, 256, 2, 2] 512
ReLU-146 [-1, 256, 2, 2] 0
Conv2d-147 [-1, 1024, 2, 2] 262,144
BatchNorm2d-148 [-1, 1024, 2, 2] 2,048
ReLU-149 [-1, 1024, 2, 2] 0
Bottleneck-150 [-1, 1024, 2, 2] 0
Conv2d-151 [-1, 256, 2, 2] 262,144
BatchNorm2d-152 [-1, 256, 2, 2] 512
ReLU-153 [-1, 256, 2, 2] 0
Conv2d-154 [-1, 256, 2, 2] 589,824
BatchNorm2d-155 [-1, 256, 2, 2] 512
ReLU-156 [-1, 256, 2, 2] 0
Conv2d-157 [-1, 1024, 2, 2] 262,144
BatchNorm2d-158 [-1, 1024, 2, 2] 2,048
ReLU-159 [-1, 1024, 2, 2] 0
Bottleneck-160 [-1, 1024, 2, 2] 0
Conv2d-161 [-1, 256, 2, 2] 262,144
BatchNorm2d-162 [-1, 256, 2, 2] 512
ReLU-163 [-1, 256, 2, 2] 0
Conv2d-164 [-1, 256, 2, 2] 589,824
BatchNorm2d-165 [-1, 256, 2, 2] 512
ReLU-166 [-1, 256, 2, 2] 0
Conv2d-167 [-1, 1024, 2, 2] 262,144
BatchNorm2d-168 [-1, 1024, 2, 2] 2,048
ReLU-169 [-1, 1024, 2, 2] 0
Bottleneck-170 [-1, 1024, 2, 2] 0
Conv2d-171 [-1, 256, 2, 2] 262,144
BatchNorm2d-172 [-1, 256, 2, 2] 512
ReLU-173 [-1, 256, 2, 2] 0
Conv2d-174 [-1, 256, 2, 2] 589,824
BatchNorm2d-175 [-1, 256, 2, 2] 512
ReLU-176 [-1, 256, 2, 2] 0
Conv2d-177 [-1, 1024, 2, 2] 262,144
BatchNorm2d-178 [-1, 1024, 2, 2] 2,048
ReLU-179 [-1, 1024, 2, 2] 0
Bottleneck-180 [-1, 1024, 2, 2] 0
Conv2d-181 [-1, 256, 2, 2] 262,144
BatchNorm2d-182 [-1, 256, 2, 2] 512
ReLU-183 [-1, 256, 2, 2] 0
Conv2d-184 [-1, 256, 2, 2] 589,824
BatchNorm2d-185 [-1, 256, 2, 2] 512
ReLU-186 [-1, 256, 2, 2] 0
Conv2d-187 [-1, 1024, 2, 2] 262,144
BatchNorm2d-188 [-1, 1024, 2, 2] 2,048
ReLU-189 [-1, 1024, 2, 2] 0
Bottleneck-190 [-1, 1024, 2, 2] 0
Conv2d-191 [-1, 256, 2, 2] 262,144
BatchNorm2d-192 [-1, 256, 2, 2] 512
ReLU-193 [-1, 256, 2, 2] 0
Conv2d-194 [-1, 256, 2, 2] 589,824
BatchNorm2d-195 [-1, 256, 2, 2] 512
ReLU-196 [-1, 256, 2, 2] 0
Conv2d-197 [-1, 1024, 2, 2] 262,144
BatchNorm2d-198 [-1, 1024, 2, 2] 2,048
ReLU-199 [-1, 1024, 2, 2] 0
Bottleneck-200 [-1, 1024, 2, 2] 0
Conv2d-201 [-1, 256, 2, 2] 262,144
BatchNorm2d-202 [-1, 256, 2, 2] 512
ReLU-203 [-1, 256, 2, 2] 0
Conv2d-204 [-1, 256, 2, 2] 589,824
BatchNorm2d-205 [-1, 256, 2, 2] 512
ReLU-206 [-1, 256, 2, 2] 0
Conv2d-207 [-1, 1024, 2, 2] 262,144
BatchNorm2d-208 [-1, 1024, 2, 2] 2,048
ReLU-209 [-1, 1024, 2, 2] 0
Bottleneck-210 [-1, 1024, 2, 2] 0
Conv2d-211 [-1, 256, 2, 2] 262,144
BatchNorm2d-212 [-1, 256, 2, 2] 512
ReLU-213 [-1, 256, 2, 2] 0
Conv2d-214 [-1, 256, 2, 2] 589,824
BatchNorm2d-215 [-1, 256, 2, 2] 512
ReLU-216 [-1, 256, 2, 2] 0
Conv2d-217 [-1, 1024, 2, 2] 262,144
BatchNorm2d-218 [-1, 1024, 2, 2] 2,048
ReLU-219 [-1, 1024, 2, 2] 0
Bottleneck-220 [-1, 1024, 2, 2] 0
Conv2d-221 [-1, 256, 2, 2] 262,144
BatchNorm2d-222 [-1, 256, 2, 2] 512
ReLU-223 [-1, 256, 2, 2] 0
Conv2d-224 [-1, 256, 2, 2] 589,824
BatchNorm2d-225 [-1, 256, 2, 2] 512
ReLU-226 [-1, 256, 2, 2] 0
Conv2d-227 [-1, 1024, 2, 2] 262,144
BatchNorm2d-228 [-1, 1024, 2, 2] 2,048
ReLU-229 [-1, 1024, 2, 2] 0
Bottleneck-230 [-1, 1024, 2, 2] 0
Conv2d-231 [-1, 256, 2, 2] 262,144
BatchNorm2d-232 [-1, 256, 2, 2] 512
ReLU-233 [-1, 256, 2, 2] 0
Conv2d-234 [-1, 256, 2, 2] 589,824
BatchNorm2d-235 [-1, 256, 2, 2] 512
ReLU-236 [-1, 256, 2, 2] 0
Conv2d-237 [-1, 1024, 2, 2] 262,144
BatchNorm2d-238 [-1, 1024, 2, 2] 2,048
ReLU-239 [-1, 1024, 2, 2] 0
Bottleneck-240 [-1, 1024, 2, 2] 0
Conv2d-241 [-1, 256, 2, 2] 262,144
BatchNorm2d-242 [-1, 256, 2, 2] 512
ReLU-243 [-1, 256, 2, 2] 0
Conv2d-244 [-1, 256, 2, 2] 589,824
BatchNorm2d-245 [-1, 256, 2, 2] 512
ReLU-246 [-1, 256, 2, 2] 0
Conv2d-247 [-1, 1024, 2, 2] 262,144
BatchNorm2d-248 [-1, 1024, 2, 2] 2,048
ReLU-249 [-1, 1024, 2, 2] 0
Bottleneck-250 [-1, 1024, 2, 2] 0
Conv2d-251 [-1, 256, 2, 2] 262,144
BatchNorm2d-252 [-1, 256, 2, 2] 512
ReLU-253 [-1, 256, 2, 2] 0
Conv2d-254 [-1, 256, 2, 2] 589,824
BatchNorm2d-255 [-1, 256, 2, 2] 512
ReLU-256 [-1, 256, 2, 2] 0
Conv2d-257 [-1, 1024, 2, 2] 262,144
BatchNorm2d-258 [-1, 1024, 2, 2] 2,048
ReLU-259 [-1, 1024, 2, 2] 0
Bottleneck-260 [-1, 1024, 2, 2] 0
Conv2d-261 [-1, 256, 2, 2] 262,144
BatchNorm2d-262 [-1, 256, 2, 2] 512
ReLU-263 [-1, 256, 2, 2] 0
Conv2d-264 [-1, 256, 2, 2] 589,824
BatchNorm2d-265 [-1, 256, 2, 2] 512
ReLU-266 [-1, 256, 2, 2] 0
Conv2d-267 [-1, 1024, 2, 2] 262,144
BatchNorm2d-268 [-1, 1024, 2, 2] 2,048
ReLU-269 [-1, 1024, 2, 2] 0
Bottleneck-270 [-1, 1024, 2, 2] 0
Conv2d-271 [-1, 256, 2, 2] 262,144
BatchNorm2d-272 [-1, 256, 2, 2] 512
ReLU-273 [-1, 256, 2, 2] 0
Conv2d-274 [-1, 256, 2, 2] 589,824
BatchNorm2d-275 [-1, 256, 2, 2] 512
ReLU-276 [-1, 256, 2, 2] 0
Conv2d-277 [-1, 1024, 2, 2] 262,144
BatchNorm2d-278 [-1, 1024, 2, 2] 2,048
ReLU-279 [-1, 1024, 2, 2] 0
Bottleneck-280 [-1, 1024, 2, 2] 0
Conv2d-281 [-1, 256, 2, 2] 262,144
BatchNorm2d-282 [-1, 256, 2, 2] 512
ReLU-283 [-1, 256, 2, 2] 0
Conv2d-284 [-1, 256, 2, 2] 589,824
BatchNorm2d-285 [-1, 256, 2, 2] 512
ReLU-286 [-1, 256, 2, 2] 0
Conv2d-287 [-1, 1024, 2, 2] 262,144
BatchNorm2d-288 [-1, 1024, 2, 2] 2,048
ReLU-289 [-1, 1024, 2, 2] 0
Bottleneck-290 [-1, 1024, 2, 2] 0
Conv2d-291 [-1, 256, 2, 2] 262,144
BatchNorm2d-292 [-1, 256, 2, 2] 512
ReLU-293 [-1, 256, 2, 2] 0
Conv2d-294 [-1, 256, 2, 2] 589,824
BatchNorm2d-295 [-1, 256, 2, 2] 512
ReLU-296 [-1, 256, 2, 2] 0
Conv2d-297 [-1, 1024, 2, 2] 262,144
BatchNorm2d-298 [-1, 1024, 2, 2] 2,048
ReLU-299 [-1, 1024, 2, 2] 0
Bottleneck-300 [-1, 1024, 2, 2] 0
Conv2d-301 [-1, 256, 2, 2] 262,144
BatchNorm2d-302 [-1, 256, 2, 2] 512
ReLU-303 [-1, 256, 2, 2] 0
Conv2d-304 [-1, 256, 2, 2] 589,824
BatchNorm2d-305 [-1, 256, 2, 2] 512
ReLU-306 [-1, 256, 2, 2] 0
Conv2d-307 [-1, 1024, 2, 2] 262,144
BatchNorm2d-308 [-1, 1024, 2, 2] 2,048
ReLU-309 [-1, 1024, 2, 2] 0
Bottleneck-310 [-1, 1024, 2, 2] 0
Conv2d-311 [-1, 256, 2, 2] 262,144
BatchNorm2d-312 [-1, 256, 2, 2] 512
ReLU-313 [-1, 256, 2, 2] 0
Conv2d-314 [-1, 256, 2, 2] 589,824
BatchNorm2d-315 [-1, 256, 2, 2] 512
ReLU-316 [-1, 256, 2, 2] 0
Conv2d-317 [-1, 1024, 2, 2] 262,144
BatchNorm2d-318 [-1, 1024, 2, 2] 2,048
ReLU-319 [-1, 1024, 2, 2] 0
Bottleneck-320 [-1, 1024, 2, 2] 0
Conv2d-321 [-1, 256, 2, 2] 262,144
BatchNorm2d-322 [-1, 256, 2, 2] 512
ReLU-323 [-1, 256, 2, 2] 0
Conv2d-324 [-1, 256, 2, 2] 589,824
BatchNorm2d-325 [-1, 256, 2, 2] 512
ReLU-326 [-1, 256, 2, 2] 0
Conv2d-327 [-1, 1024, 2, 2] 262,144
BatchNorm2d-328 [-1, 1024, 2, 2] 2,048
ReLU-329 [-1, 1024, 2, 2] 0
Bottleneck-330 [-1, 1024, 2, 2] 0
Conv2d-331 [-1, 256, 2, 2] 262,144
BatchNorm2d-332 [-1, 256, 2, 2] 512
ReLU-333 [-1, 256, 2, 2] 0
Conv2d-334 [-1, 256, 2, 2] 589,824
BatchNorm2d-335 [-1, 256, 2, 2] 512
ReLU-336 [-1, 256, 2, 2] 0
Conv2d-337 [-1, 1024, 2, 2] 262,144
BatchNorm2d-338 [-1, 1024, 2, 2] 2,048
ReLU-339 [-1, 1024, 2, 2] 0
Bottleneck-340 [-1, 1024, 2, 2] 0
Conv2d-341 [-1, 256, 2, 2] 262,144
BatchNorm2d-342 [-1, 256, 2, 2] 512
ReLU-343 [-1, 256, 2, 2] 0
Conv2d-344 [-1, 256, 2, 2] 589,824
BatchNorm2d-345 [-1, 256, 2, 2] 512
ReLU-346 [-1, 256, 2, 2] 0
Conv2d-347 [-1, 1024, 2, 2] 262,144
BatchNorm2d-348 [-1, 1024, 2, 2] 2,048
ReLU-349 [-1, 1024, 2, 2] 0
Bottleneck-350 [-1, 1024, 2, 2] 0
Conv2d-351 [-1, 256, 2, 2] 262,144
BatchNorm2d-352 [-1, 256, 2, 2] 512
ReLU-353 [-1, 256, 2, 2] 0
Conv2d-354 [-1, 256, 2, 2] 589,824
BatchNorm2d-355 [-1, 256, 2, 2] 512
ReLU-356 [-1, 256, 2, 2] 0
Conv2d-357 [-1, 1024, 2, 2] 262,144
BatchNorm2d-358 [-1, 1024, 2, 2] 2,048
ReLU-359 [-1, 1024, 2, 2] 0
Bottleneck-360 [-1, 1024, 2, 2] 0
Conv2d-361 [-1, 256, 2, 2] 262,144
BatchNorm2d-362 [-1, 256, 2, 2] 512
ReLU-363 [-1, 256, 2, 2] 0
Conv2d-364 [-1, 256, 2, 2] 589,824
BatchNorm2d-365 [-1, 256, 2, 2] 512
ReLU-366 [-1, 256, 2, 2] 0
Conv2d-367 [-1, 1024, 2, 2] 262,144
BatchNorm2d-368 [-1, 1024, 2, 2] 2,048
ReLU-369 [-1, 1024, 2, 2] 0
Bottleneck-370 [-1, 1024, 2, 2] 0
Conv2d-371 [-1, 256, 2, 2] 262,144
BatchNorm2d-372 [-1, 256, 2, 2] 512
ReLU-373 [-1, 256, 2, 2] 0
Conv2d-374 [-1, 256, 2, 2] 589,824
BatchNorm2d-375 [-1, 256, 2, 2] 512
ReLU-376 [-1, 256, 2, 2] 0
Conv2d-377 [-1, 1024, 2, 2] 262,144
BatchNorm2d-378 [-1, 1024, 2, 2] 2,048
ReLU-379 [-1, 1024, 2, 2] 0
Bottleneck-380 [-1, 1024, 2, 2] 0
Conv2d-381 [-1, 256, 2, 2] 262,144
BatchNorm2d-382 [-1, 256, 2, 2] 512
ReLU-383 [-1, 256, 2, 2] 0
Conv2d-384 [-1, 256, 2, 2] 589,824
BatchNorm2d-385 [-1, 256, 2, 2] 512
ReLU-386 [-1, 256, 2, 2] 0
Conv2d-387 [-1, 1024, 2, 2] 262,144
BatchNorm2d-388 [-1, 1024, 2, 2] 2,048
ReLU-389 [-1, 1024, 2, 2] 0
Bottleneck-390 [-1, 1024, 2, 2] 0
Conv2d-391 [-1, 256, 2, 2] 262,144
BatchNorm2d-392 [-1, 256, 2, 2] 512
ReLU-393 [-1, 256, 2, 2] 0
Conv2d-394 [-1, 256, 2, 2] 589,824
BatchNorm2d-395 [-1, 256, 2, 2] 512
ReLU-396 [-1, 256, 2, 2] 0
Conv2d-397 [-1, 1024, 2, 2] 262,144
BatchNorm2d-398 [-1, 1024, 2, 2] 2,048
ReLU-399 [-1, 1024, 2, 2] 0
Bottleneck-400 [-1, 1024, 2, 2] 0
Conv2d-401 [-1, 256, 2, 2] 262,144
BatchNorm2d-402 [-1, 256, 2, 2] 512
ReLU-403 [-1, 256, 2, 2] 0
Conv2d-404 [-1, 256, 2, 2] 589,824
BatchNorm2d-405 [-1, 256, 2, 2] 512
ReLU-406 [-1, 256, 2, 2] 0
Conv2d-407 [-1, 1024, 2, 2] 262,144
BatchNorm2d-408 [-1, 1024, 2, 2] 2,048
ReLU-409 [-1, 1024, 2, 2] 0
Bottleneck-410 [-1, 1024, 2, 2] 0
Conv2d-411 [-1, 256, 2, 2] 262,144
BatchNorm2d-412 [-1, 256, 2, 2] 512
ReLU-413 [-1, 256, 2, 2] 0
Conv2d-414 [-1, 256, 2, 2] 589,824
BatchNorm2d-415 [-1, 256, 2, 2] 512
ReLU-416 [-1, 256, 2, 2] 0
Conv2d-417 [-1, 1024, 2, 2] 262,144
BatchNorm2d-418 [-1, 1024, 2, 2] 2,048
ReLU-419 [-1, 1024, 2, 2] 0
Bottleneck-420 [-1, 1024, 2, 2] 0
Conv2d-421 [-1, 256, 2, 2] 262,144
BatchNorm2d-422 [-1, 256, 2, 2] 512
ReLU-423 [-1, 256, 2, 2] 0
Conv2d-424 [-1, 256, 2, 2] 589,824
BatchNorm2d-425 [-1, 256, 2, 2] 512
ReLU-426 [-1, 256, 2, 2] 0
Conv2d-427 [-1, 1024, 2, 2] 262,144
BatchNorm2d-428 [-1, 1024, 2, 2] 2,048
ReLU-429 [-1, 1024, 2, 2] 0
Bottleneck-430 [-1, 1024, 2, 2] 0
Conv2d-431 [-1, 256, 2, 2] 262,144
BatchNorm2d-432 [-1, 256, 2, 2] 512
ReLU-433 [-1, 256, 2, 2] 0
Conv2d-434 [-1, 256, 2, 2] 589,824
BatchNorm2d-435 [-1, 256, 2, 2] 512
ReLU-436 [-1, 256, 2, 2] 0
Conv2d-437 [-1, 1024, 2, 2] 262,144
BatchNorm2d-438 [-1, 1024, 2, 2] 2,048
ReLU-439 [-1, 1024, 2, 2] 0
Bottleneck-440 [-1, 1024, 2, 2] 0
Conv2d-441 [-1, 256, 2, 2] 262,144
BatchNorm2d-442 [-1, 256, 2, 2] 512
ReLU-443 [-1, 256, 2, 2] 0
Conv2d-444 [-1, 256, 2, 2] 589,824
BatchNorm2d-445 [-1, 256, 2, 2] 512
ReLU-446 [-1, 256, 2, 2] 0
Conv2d-447 [-1, 1024, 2, 2] 262,144
BatchNorm2d-448 [-1, 1024, 2, 2] 2,048
ReLU-449 [-1, 1024, 2, 2] 0
Bottleneck-450 [-1, 1024, 2, 2] 0
Conv2d-451 [-1, 256, 2, 2] 262,144
BatchNorm2d-452 [-1, 256, 2, 2] 512
ReLU-453 [-1, 256, 2, 2] 0
Conv2d-454 [-1, 256, 2, 2] 589,824
BatchNorm2d-455 [-1, 256, 2, 2] 512
ReLU-456 [-1, 256, 2, 2] 0
Conv2d-457 [-1, 1024, 2, 2] 262,144
BatchNorm2d-458 [-1, 1024, 2, 2] 2,048
ReLU-459 [-1, 1024, 2, 2] 0
Bottleneck-460 [-1, 1024, 2, 2] 0
Conv2d-461 [-1, 256, 2, 2] 262,144
BatchNorm2d-462 [-1, 256, 2, 2] 512
ReLU-463 [-1, 256, 2, 2] 0
Conv2d-464 [-1, 256, 2, 2] 589,824
BatchNorm2d-465 [-1, 256, 2, 2] 512
ReLU-466 [-1, 256, 2, 2] 0
Conv2d-467 [-1, 1024, 2, 2] 262,144
BatchNorm2d-468 [-1, 1024, 2, 2] 2,048
ReLU-469 [-1, 1024, 2, 2] 0
Bottleneck-470 [-1, 1024, 2, 2] 0
Conv2d-471 [-1, 256, 2, 2] 262,144
BatchNorm2d-472 [-1, 256, 2, 2] 512
ReLU-473 [-1, 256, 2, 2] 0
Conv2d-474 [-1, 256, 2, 2] 589,824
BatchNorm2d-475 [-1, 256, 2, 2] 512
ReLU-476 [-1, 256, 2, 2] 0
Conv2d-477 [-1, 1024, 2, 2] 262,144
BatchNorm2d-478 [-1, 1024, 2, 2] 2,048
ReLU-479 [-1, 1024, 2, 2] 0
Bottleneck-480 [-1, 1024, 2, 2] 0
Conv2d-481 [-1, 512, 2, 2] 524,288
BatchNorm2d-482 [-1, 512, 2, 2] 1,024
ReLU-483 [-1, 512, 2, 2] 0
Conv2d-484 [-1, 512, 1, 1] 2,359,296
BatchNorm2d-485 [-1, 512, 1, 1] 1,024
ReLU-486 [-1, 512, 1, 1] 0
Conv2d-487 [-1, 2048, 1, 1] 1,048,576
BatchNorm2d-488 [-1, 2048, 1, 1] 4,096
Conv2d-489 [-1, 2048, 1, 1] 2,097,152
BatchNorm2d-490 [-1, 2048, 1, 1] 4,096
ReLU-491 [-1, 2048, 1, 1] 0
Bottleneck-492 [-1, 2048, 1, 1] 0
Conv2d-493 [-1, 512, 1, 1] 1,048,576
BatchNorm2d-494 [-1, 512, 1, 1] 1,024
ReLU-495 [-1, 512, 1, 1] 0
Conv2d-496 [-1, 512, 1, 1] 2,359,296
BatchNorm2d-497 [-1, 512, 1, 1] 1,024
ReLU-498 [-1, 512, 1, 1] 0
Conv2d-499 [-1, 2048, 1, 1] 1,048,576
BatchNorm2d-500 [-1, 2048, 1, 1] 4,096
ReLU-501 [-1, 2048, 1, 1] 0
Bottleneck-502 [-1, 2048, 1, 1] 0
Conv2d-503 [-1, 512, 1, 1] 1,048,576
BatchNorm2d-504 [-1, 512, 1, 1] 1,024
ReLU-505 [-1, 512, 1, 1] 0
Conv2d-506 [-1, 512, 1, 1] 2,359,296
BatchNorm2d-507 [-1, 512, 1, 1] 1,024
ReLU-508 [-1, 512, 1, 1] 0
Conv2d-509 [-1, 2048, 1, 1] 1,048,576
BatchNorm2d-510 [-1, 2048, 1, 1] 4,096
ReLU-511 [-1, 2048, 1, 1] 0
Bottleneck-512 [-1, 2048, 1, 1] 0
AdaptiveAvgPool2d-513 [-1, 2048, 1, 1] 0
Linear-514 [-1, 100] 204,900
LogSoftmax-515 [-1, 100] 0
================================================================
Total params: 58,348,708
Trainable params: 204,900
Non-trainable params: 58,143,808
----------------------------------------------------------------
Input size (MB): 0.01
Forward/backward pass size (MB): 12.40
Params size (MB): 222.58
Estimated Total Size (MB): 234.99
----------------------------------------------------------------
None
Params to learn
fc.0.weight
fc.0.bias
Files already downloaded and verified
Files already downloaded and verified
Epoch 0/9
----------
Time elapsed 0m 21s
train Loss: 7.5111 Acc: 0.1484
Time elapsed 0m 26s
valid Loss: 3.7821 Acc: 0.2493
/usr/local/lib/python3.7/dist-packages/torch/optim/lr_scheduler.py:154: UserWarning: The epoch parameter in `scheduler.step()` was not necessary and is being deprecated where possible. Please use `scheduler.step()` to step the scheduler. During the deprecation, if epoch is different from None, the closed form is used instead of the new chainable form, where available. Please open an issue if you are unable to replicate your use case: https://github.com/pytorch/pytorch/issues/new/choose.
warnings.warn(EPOCH_DEPRECATION_WARNING, UserWarning)
Optimizer learning rate: 0.0100000

Epoch 1/9
----------
Time elapsed 0m 47s
train Loss: 2.9405 Acc: 0.3109
Time elapsed 0m 52s
valid Loss: 3.2014 Acc: 0.2739
Optimizer learning rate: 0.0100000

Epoch 2/9
----------
Time elapsed 1m 12s
train Loss: 2.5866 Acc: 0.3622
Time elapsed 1m 17s
valid Loss: 3.2239 Acc: 0.2787
Optimizer learning rate: 0.0100000

Epoch 3/9
----------
Time elapsed 1m 38s
train Loss: 2.4077 Acc: 0.3969
Time elapsed 1m 43s
valid Loss: 3.2608 Acc: 0.2811
Optimizer learning rate: 0.0100000

Epoch 4/9
----------
Time elapsed 2m 4s
train Loss: 2.2742 Acc: 0.4263
Time elapsed 2m 9s
valid Loss: 3.4260 Acc: 0.2689
Optimizer learning rate: 0.0100000

Epoch 5/9
----------
Time elapsed 2m 29s
train Loss: 2.1942 Acc: 0.4434
Time elapsed 2m 34s
valid Loss: 3.4697 Acc: 0.2760
Optimizer learning rate: 0.0100000

Epoch 6/9
----------
Time elapsed 2m 54s
train Loss: 2.1369 Acc: 0.4583
Time elapsed 2m 59s
valid Loss: 3.5391 Acc: 0.2744
Optimizer learning rate: 0.0100000

Epoch 7/9
----------
Time elapsed 3m 20s
train Loss: 2.0382 Acc: 0.4771
Time elapsed 3m 24s
valid Loss: 3.5992 Acc: 0.2721
Optimizer learning rate: 0.0100000

Epoch 8/9
----------
Time elapsed 3m 45s
train Loss: 1.9776 Acc: 0.4939
Time elapsed 3m 50s
valid Loss: 3.7533 Acc: 0.2685
Optimizer learning rate: 0.0100000

Epoch 9/9
----------
Time elapsed 4m 11s
train Loss: 1.9309 Acc: 0.5035
Time elapsed 4m 16s
valid Loss: 3.9663 Acc: 0.2558
Optimizer learning rate: 0.0100000

Training complete in 4m 16s
Best val Acc: 0.281100