def _CUB200():
"""
CUB200数据集
:return:
"""
# 定义模型 定义评价 优化器等
lr = 1e-4
class_num = 200
torch.manual_seed(0)
torch.cuda.manual_seed_all(0)
print("cuda:0,1,2,3")
device = torch.device("cuda:2" if torch.cuda.is_available() else "cpu")
model = Net(resnet_for_pmg.resnet50(pretrained=True), class_num)
device_ids = [0, 1, 2, 3]
model = nn.DataParallel(model, device_ids=device_ids).cuda(0)
model.to(f'cuda:{model.device_ids[0]}')
criterion = torch.nn.CrossEntropyLoss()
optimzer = torch.optim.SGD(
[
{
'params': model.module.pmg.features.parameters(),
'lr': lr * 1
},
{
'params': model.module.pmg.classifier_concat.parameters(),
'lr': lr * 10
},
{
'params': model.module.pmg.classifier1.parameters(),
'lr': lr * 10
},
{
'params': model.module.pmg.classifier2.parameters(),
'lr': lr * 10
},
{
'params': model.module.pmg.classifier3.parameters(),
'lr': lr * 10
},
{
'params': model.module.pmg.conv_block1.parameters(),
'lr': lr * 10
},
{
'params': model.module.pmg.conv_block2.parameters(),
'lr': lr * 10
},
{
'params': model.module.pmg.conv_block3.parameters(),
'lr': lr * 10
},
# {'params': model.module.pmg.map1.parameters(), 'lr': lr * 10},
# {'params': model.module.pmg.map2.parameters(), 'lr': lr * 10},
# {'params': model.module.pmg.fc.parameters(), 'lr': lr * 10},
],
lr=lr,
momentum=0.9,
weight_decay=5e-4)
# torch.optim.lr_scheduler.StepLR(optimzer, 10, gamma=0.94, last_epoch=-1)
torch.optim.lr_scheduler.CosineAnnealingLR(optimzer, T_max=10)
epochs = 200
batchSize = 32
worker = 4
modelConfig = {
'model': model,
'criterion': criterion,
'optimzer': optimzer,
'epochs': epochs,
'device': device
}
from torchvision import transforms as T
# 自定义数据增强方式
trainTransforms = T.Compose([
T.Resize(550),
T.RandomCrop(448, padding=8),
T.RandomHorizontalFlip(),
T.ToTensor(),
T.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
# T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
testTransforms = T.Compose([
T.Resize(550),
T.CenterCrop(448),
T.ToTensor(),
T.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
# T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
trainLoader, testLoader, validLoader, trainLength, testLength, validLength = chooseData(
'CUB200', batchSize, worker, trainTransforms, testTransforms)
# 没有验证集，所以使用测试集来做验证集
dataConfig = {
'trainLoader': trainLoader,
'validLoader': testLoader,
'trainLength': trainLength,
'validLength': testLength
}
modelPath = os.path.join(os.getcwd(), 'checkpoints', '_CUB200.pth')
lastModelPath = os.path.join(os.getcwd(), 'checkpoints',
'_CUB200_last.pth')
historyPath = os.path.join(os.getcwd(), 'historys', '_CUB200.npy')
logPath = os.path.join(os.getcwd(), 'logs', '_CUB200.txt')
logConfig = {
'modelPath': modelPath,
'historyPath': historyPath,
'logPath': logPath,
'lastModelPath': lastModelPath
}
train(modelConfig, dataConfig, logConfig)
def _stanfordDogs():
"""
StanfordDogs数据集
:return:
"""
# 定义模型 定义评价 优化器等
lr = 1e-4
class_num = 120
print("cuda:2")
device = torch.device("cuda:2" if torch.cuda.is_available() else "cpu")
model = Net(resnet_for_pmg.resnet50(pretrained=True), class_num)
device_ids = [0]
model = nn.DataParallel(model, device_ids=device_ids).cuda(0)
model.to(f'cuda:{model.device_ids[0]}')
criterion = torch.nn.CrossEntropyLoss()
optimzer = torch.optim.SGD([
{
'params': model.module.pmg.features.parameters(),
'lr': lr * 1
},
{
'params': model.module.pmg.classifier1.parameters(),
'lr': lr * 10
},
{
'params': model.module.pmg.conv_block1.parameters(),
'lr': lr * 10
},
{
'params': model.module.pmg.conv_block2.parameters(),
'lr': lr * 10
},
{
'params': model.module.pmg.conv_block3.parameters(),
'lr': lr * 10
},
{
'params': model.module.pmg.map1.parameters(),
'lr': lr * 10
},
{
'params': model.module.pmg.map2.parameters(),
'lr': lr * 10
},
],
lr=lr,
momentum=0.9,
weight_decay=5e-4)
# torch.optim.lr_scheduler.StepLR(optimzer, 10, gamma=0.94, last_epoch=-1)
torch.optim.lr_scheduler.CosineAnnealingLR(optimzer, T_max=10)
epochs = 200
batchSize = 15
worker = 2
modelConfig = {
'model': model,
'criterion': criterion,
'optimzer': optimzer,
'epochs': epochs,
'device': device
}
from torchvision import transforms as T
# 自定义数据增强方式
# normalize 加快收敛
# normalize = T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
trainTransforms = T.Compose([
T.Resize(256),
T.RandomRotation(15),
# T.RandomResizedCrop(224,scale=(0.85,1.15)),
T.RandomCrop(224),
T.ToTensor(),
T.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
])
testTransforms = T.Compose([
T.Resize(256),
T.CenterCrop(224),
T.ToTensor(),
T.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
])
trainLoader, testLoader, validLoader, trainLength, testLength, validLength = chooseData(
'STANFORDDOGS', batchSize, worker, trainTransforms, testTransforms)
# 没有验证集，所以使用测试集来做验证集
dataConfig = {
'trainLoader': trainLoader,
'validLoader': testLoader,
'trainLength': trainLength,
'validLength': testLength
}
modelPath = os.path.join(os.getcwd(), 'checkpoints', '_stanforddogs.pth')
lastModelPath = os.path.join(os.getcwd(), 'checkpoints',
'_stanforddogs_last.pth')
historyPath = os.path.join(os.getcwd(), 'historys', '_stanforddogs.npy')
logPath = os.path.join(os.getcwd(), 'logs', '_stanforddogs.txt')
logConfig = {
'modelPath': modelPath,
'historyPath': historyPath,
'logPath': logPath,
'lastModelPath': lastModelPath
}
train(modelConfig, dataConfig, logConfig)