return source_loader, target_train_loader, target_test_loader
if __name__ == '__main__':
torch.manual_seed(0)
source_name = "amazon"
target_name = "webcam"
print('Src: %s, Tar: %s' % (source_name, target_name))
source_loader, target_train_loader, target_test_loader = load_data(
source_name, target_name, CFG['data_path'])
model = models.Transfer_Net(CFG['n_class'],
transfer_loss='mmd',
base_net='resnet50').to(DEVICE)
optimizer = torch.optim.SGD([
{
'params': model.base_network.parameters()
},
{
'params': model.bottleneck_layer.parameters(),
'lr': 10 * CFG['lr']
},
{
'params': model.classifier_layer.parameters(),
'lr': 10 * CFG['lr']
},
],
lr=CFG['lr'],
# make_dot was moved to https://github.com/szagoruyko/pytorchviz
from torchviz import make_dot
if __name__ == "__main__":
torch.manual_seed(0)
source_name = "cmf"
target_name = "casia"
# print('Src: %s, Tar: %s' % (source_name, target_name))
# source_loader, target_train_loader, target_test_loader = load_data(
# source_name, target_name, CFG['data_path'])
model = models.Transfer_Net(CFG["n_class"],
transfer_loss="mmd",
base_net="alexnet").to(DEVICE)
optimizer = torch.optim.SGD(
[
{
"params": model.base_network.parameters()
},
{
"params": model.bottleneck_layer.parameters(),
"lr": 10 * CFG["lr"]
},
{
"params": model.classifier_layer.parameters(),
"lr": 10 * CFG["lr"]
},
],
def main(loaders, final_test_loaders, excel_files, im):
excel, sheet_src, sheet_tar = excel_files # 将excel的参数包解压出来
# 参数包准备
model = models.Transfer_Net(CFG).to(DEVICE) # 创建model，并且将其拷贝到指定的device中
params = spparams_creator(model, CFG) # 生成用于模型训练的参数
if ('RevGrad' in CFG['tranmodel']): # 对于RevGrad模型，需要单独设计优化器与衰减器
optimizer0 = torch.optim.SGD(params[0],
lr=CFG['lr'],
momentum=CFG['momentum'],
weight_decay=CFG['l2_decay']) # 主干的参数
optimizer1 = torch.optim.SGD(params[1],
lr=CFG['lr'],
momentum=CFG['momentum'],
weight_decay=CFG['l2_decay']) # domclf的参数
scheduler0 = StepLR(optimizer0, step_size=1, gamma=0.95) # 学习速率衰减方式
scheduler1 = StepLR(optimizer1, step_size=1, gamma=0.95) # 学习速率衰减方式
optimizer = [optimizer0, optimizer1] # 将两个优化器放在list中便于参数传递
scheduler = [scheduler0, scheduler1] #
else: # 对于DDC优化器与迭代器分别只有1个
optimizer = torch.optim.SGD(params,
lr=CFG['lr'],
momentum=CFG['momentum'],
weight_decay=CFG['l2_decay'])
scheduler = StepLR(optimizer, step_size=1, gamma=0.95) # 学习速率衰减方式
# 开始训练
result_trace = train(loaders, model, optimizer, scheduler,
CFG) # 用train_utils文件中的train函数进行训练
# 绘制训练曲线
acc_train, acc_test = np.array(result_trace[0]), np.array(
result_trace[3]) # index=0是
acc_train_src, acc_train_tar, acc_test_src, acc_test_tar = acc_train[:,
0], acc_train[:,
6], acc_test[:,
0], acc_test[:,
6] # 第0列是源域最终投票正确率，第6列是目标域最终投票正确率
x = [i for i in range(len(acc_train_src))]
y = [acc_train_src, acc_train_tar, acc_test_src, acc_test_tar] # x，y，用于画图
plot_curve(x,
y,
path_name_curve + '/' + CFG['tranmodel'] + '_ACC_end.png',
xlabel='Epoch',
ylabel='ACC',
title=CFG['tranmodel'] + 'ACC',
legend=[
'acc_train_src', 'acc_train_tar', 'acc_test_src',
'acc_test_tar'
]) #调用自定义函数进行画图
# 测试结果
acc_train_val_src = [acc_train_src[-1],
acc_test_src[-1]] # 训练集和验证集的收敛正确率，源域和目标域
acc_train_val_tar = [acc_train_tar[-1], acc_test_tar[-1]]
acc_test_srcs, acc_test_tars = final_test(
model, CFG, final_test_loaders) # 对多个平行测试集进行测试的正确率
acc_4excel_src = acc_train_val_src + [
0, 0
] + acc_test_srcs # 将训练集验证集的正确率，[0，0](用于占位)，平行测试集的正确率 一起拼接成list，以便于写入excel
acc_4excel_tar = acc_train_val_tar + [0, 0] + acc_test_tars
save_excel_4final_test(acc_4excel_src, acc_4excel_tar, excel, sheet_src,
sheet_tar, im, CFG)
excel.save(dir_now + '/测试结果.xls')
# pt文件保存
if ('RevGrad' in CFG['tranmodel']): # 对于RevGrad需要保存两个优化器与两个衰减器
pt = {
'model': model.state_dict(),
'optimizer0': optimizer[0].state_dict(),
'optimizer1': optimizer[1].state_dict(),
'scheduler0': scheduler[0].state_dict(),
'scheduler1': scheduler[1].state_dict(),
'CFG': CFG,
'result_trace': result_trace,
'test_result': [acc_test_srcs, acc_test_tars]
}
else: # 对于DDC优化器与迭代器分别只有1个
pt = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'CFG': CFG,
'result_trace': result_trace,
'test_result': [acc_test_srcs, acc_test_tars]
}
torch.save(pt, dir_now + '/' + CFG['tranmodel'] + '_end.pt')
return source_loader, target_train_loader, target_test_loader
if __name__ == '__main__':
torch.manual_seed(0)
source_name = "amazon"
target_name = "webcam"
print('Src: %s, Tar: %s' % (source_name, target_name))
source_loader, target_train_loader, target_test_loader = load_data(
source_name, target_name, args.data)
model = models.Transfer_Net(args.n_class,
transfer_loss=args.trans_loss,
base_net=args.model).to(DEVICE)
optimizer = torch.optim.SGD([
{
'params': model.base_network.parameters()
},
{
'params': model.bottleneck_layer.parameters(),
'lr': 10 * args.lr
},
{
'params': model.classifier_layer.parameters(),
'lr': 10 * args.lr
},
],
lr=args.lr,
CFG['backbone'] = args.backbone
CFG['lambda'] = np.logspace(args.lambda_initial, args.lambda_final,
CFG['epoch'])
torch.manual_seed(0)
source_name = args.source
target_name = "RealWorld"
print('Src: %s, Tar: %s' % (source_name, target_name))
source_loader, target_train_loader, target_test_loader = load_data(
source_name, target_name, CFG['data_path'])
model = models.Transfer_Net(CFG['n_class'],
transfer_loss='coral',
base_net=CFG['backbone']).to(DEVICE)
optimizer = torch.optim.SGD([
{
'params': model.base_network.parameters()
},
{
'params': model.bottleneck_layer.parameters(),
'lr': 10 * CFG['lr']
},
{
'params': model.classifier_layer.parameters(),
'lr': 10 * CFG['lr']
},
],
lr=CFG['lr'],