def qianyixuexi():
train_loader,test_loader=Img_S.Get_Dataloader()
model_ft = models.resnet18()
model_ft.load_state_dict(torch.load('D:\\Learn_Pytorch\\renet_model\\resnet18-5c106cde.pth'))
monum_ftrs = model_ft.fc.in_features
model_ft.fc = nn.Linear(monum_ftrs, 10)
model_ft = model_ft.to(device)
criterion = nn.CrossEntropyLoss()
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)
for epoch in range(epochs):
# for image,label in train_loader:
# print(image)
train_tool.train_one_epoch(model_ft,criterion,optimizer_ft,train_loader,device,epoch,print_freq=50)
exp_lr_scheduler.step()
train_tool.evaluate(model_ft,criterion,test_loader,device)
def Hyperspectral_RNN():
'''使用RNN对高光谱数据进行分类'''
opath = "D:\\ZHR_USE\\torchmodel\\RNNHyperspectral\\Results\\双向RNN_Salinas\\"
writer = SummaryWriter(opath)
epochs = 301
train_loader, test_loader = Ima_S.Get_Salinas_Dataloader_RNNdata(9)
#参数
i_size, hidden_size, num_layers, num_classes = 10, 50, 2, 16
# net=RNNModel.RNN(i_size,hidden_size,num_layers,num_classes,device)
net = RNNModel.RNN_Bidirectional(i_size, hidden_size, num_layers,
num_classes, device)
# net = RNNModel.LSTM(i_size, hidden_size, num_layers, num_classes, device)
# net = RNNModel.GRU_Bidirectional(i_size, hidden_size, num_layers, num_classes, device)
net.to(device)
dataiter = iter(train_loader)
hydata, label = dataiter.next()
writer.add_graph(net, (hydata.float(), ))
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(net.parameters(), lr=0.01)
test_iter_num = 0
#记录中间结果
train_loss = []
train_predict_label = 0
train_true_label = 0
test_loss = []
test_predict_label = 0
test_true_label = 0
test_iter_num = 0
ACC = 0
for epoch in range(epochs):
train_loss_one_epoch, train_predict_label, train_true_label = train_tool.train_one_epoch(
net, criterion, optimizer, train_loader, device, epoch, 50, writer)
train_loss.append(train_loss_one_epoch)
if epoch % 10 == 0:
test_iter_num += 1
test_loss_one_epoch, test_predict_label, test_true_label = train_tool.evaluate(
net, criterion, test_loader, device, test_iter_num, writer)
test_loss.append(test_loss_one_epoch)
ACC_one_epoch = metrics.accuracy_score(test_true_label,
test_predict_label)
if ACC_one_epoch > ACC:
Out_Result(opath, net, train_predict_label, train_true_label,
test_predict_label, test_true_label)
ACC = ACC_one_epoch
#输出最终的评价
# print(train_loss)
train_loss = np.array(train_loss)
np.savetxt(os.path.join(opath, "train_loss.csv"),
train_loss,
delimiter=',',
fmt="%.04f")
test_loss = np.array(test_loss)
np.savetxt(os.path.join(opath, "test_loss.csv"),
test_loss,
delimiter=',',
fmt="%.04f")
writer.close()
def Run():
#tensorboard 可视化
input_to_model=torch.rand([64,1,28,28])
train_loader,test_loader=Img_Split.Get_Salinas_Dataloader()
print("Load end...")
net=NET(nT).to(device)
# writer.add_graph(net, input_to_model)
# writer.close()
criterion=nn.CrossEntropyLoss()
optimizer=optim.SGD(net.parameters(),lr=0.01,momentum=0.9)
for epoch in range(epochs):
# for image,label in train_loader:
# print(image)
train_tool.train_one_epoch(net,criterion,optimizer,train_loader,device,epoch,print_freq=50)
if epoch%5==0:
print("test")
train_tool.evaluate(net,criterion,test_loader,device)
def Run():
trainloader, testloader = LY9.Get_Y9_Data()
net = NET.NET_MultiChannel(5)
criterion = nn.CrossEntropyLoss()
optimizer_ft = optim.SGD(net.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=1000,
gamma=0.1)
for epoch in range(epochs):
# for image,label in train_loader:
# print(image)
train_tool.train_one_epoch(net,
criterion,
optimizer_ft,
trainloader,
device,
epoch,
print_freq=50)
exp_lr_scheduler.step()
if epoch % 100 == 0:
train_tool.evaluate(net, criterion, testloader, device)
torch.save(net, "D:\\Net.pkl")
torch.save(net.state_dict(), "D:\\Net_param.pkl")
def RNN_All_Param_Run():
'''利用RNN对高光谱数据进行分类，所有参数一次输出'''
opath = "D:\\ZHR_USE\\torchmodel\\RNNHyperspectral\\Results\\IndianPines_AllResults\\"
epochs = 301
best_acc_dic = {} #记录过程中的所有模型的最佳精度
for s in range(3, 31, 2):
train_loader, test_loader = Ima_S.Get_IPData_RNNData(s) #读入数据
# print("train data xinxi:",len(train_loader))
# for x,y in train_loader:
# print(x.shape,y.shape)
# exit()
# train_loader, test_loader = Ima_S.Get_IPData_RNNData(s)
for hidden in range(50, 51, 1): #不同神经元遍历
models = RNNModel.Get_RNN_Models(10, hidden, 2, 16, device)
for x in models.items(): #不同模型遍历
model_name = x[0]
sub_path = os.path.join(
opath, model_name + "\\timestep_" + str(s) + "神经元个数_" +
str(hidden))
if not os.path.exists(sub_path):
os.makedirs(sub_path)
start_time = time.time()
net = x[1]
net.to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(net.parameters(), lr=0.01)
test_iter_num = 0
# 记录中间结果
train_loss = []
train_predict_label = 0
train_true_label = 0
test_loss = []
test_predict_label = 0
test_true_label = 0
test_iter_num = 0
ACC = 0
for epoch in range(epochs):
train_loss_one_epoch, train_predict_label, train_true_label = train_tool.train_one_epoch(
net, criterion, optimizer, train_loader, device, epoch,
50)
train_loss.append(train_loss_one_epoch)
if epoch % 10 == 0:
test_iter_num += 1
test_loss_one_epoch, test_predict_label, test_true_label = train_tool.evaluate(
net, criterion, test_loader, device, test_iter_num)
test_loss.append(test_loss_one_epoch)
ACC_one_epoch = metrics.accuracy_score(
test_true_label, test_predict_label)
print("精度ACC：", ACC_one_epoch)
if ACC_one_epoch > ACC:
Out_Result(sub_path, net, train_predict_label,
train_true_label, test_predict_label,
test_true_label, epoch)
ACC = ACC_one_epoch
best_acc_dic[model_name + "\ttimestep:" + str(s) + "\t神经元个数:" +
str(hidden)] = ACC
#输出花费时间
fp = open(os.path.join(sub_path, "spent_time.txt"), 'w')
print("spent time: " + str(time.time() - start_time), file=fp)
fp.close()
# 输出最终的评价
# print(train_loss)
train_loss = np.array(train_loss)
np.savetxt(os.path.join(sub_path, "train_loss.csv"),
train_loss,
delimiter=',',
fmt="%.04f")
test_loss = np.array(test_loss)
np.savetxt(os.path.join(sub_path, "test_loss.csv"),
test_loss,
delimiter=',',
fmt="%.04f")
best_acc_list = best_acc_dic.items()
best_acc = np.row_stack(best_acc_list)
np.savetxt(os.path.join(opath, "不同参数精度对比.txt"),
best_acc,
fmt='%s',
delimiter="\t")
def Run_8wdata():
opath = "D:\\ZHR_USE\\1DCNN\Results\\8wdata_1DCNN_73核3_Adam_9windowsize"
writer = SummaryWriter(opath)
window_size = 9
trainloader, testloader = L8w.Get_Data_Logging_1DCNN(window_size)
flat_len = int(((window_size * 2 + 1 - 7) / 2 - 2) / 2)
net = NET.NET_8wdata_MultiChannel_3(6, flat_len)
#tensorboard
dataiter = iter(trainloader)
logging, label = dataiter.next()
print(logging.shape)
writer.add_graph(net, (logging.float(), ))
criterion = nn.CrossEntropyLoss()
optimizer_ft = optim.Adam(net.parameters(), lr=0.001)
# decay LR by a factor of 0.1 every 7 epochs
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft,
step_size=100,
gamma=0.1)
train_loss = []
train_predict_label = 0
train_true_label = 0
test_loss = []
test_predict_label = 0
test_true_label = 0
test_iter_num = 0
ACC = 0 #初始精度
#输出每次迭代的精度
fptrainACC = open(os.path.join(opath, "train_ACC.txt"), 'w')
fptrainACC = open(os.path.join(opath, "train_ACC.txt"), 'a')
fptestACC = open(os.path.join(opath, "test_ACC.txt"), 'w')
fptestACC = open(os.path.join(opath, "test_ACC.txt"), 'a')
for epoch in range(epochs):
# for image,label in train_loader:
# print(image)
train_loss_one_epoch, train_predict_label, train_true_label = train_tool.train_one_epoch(
net,
criterion,
optimizer_ft,
trainloader,
device,
epoch,
print_freq=200,
writer=writer)
train_loss.append(train_loss_one_epoch)
# exp_lr_scheduler.step()
train_ACC = metrics.accuracy_score(train_true_label,
train_predict_label)
print(train_ACC, file=fptrainACC)
if epoch % 10 == 0:
test_loss_one_epoch, test_predict_label, test_true_label = train_tool.evaluate(
net, criterion, testloader, device, test_iter_num, writer)
ACC_one_epoch = metrics.accuracy_score(test_true_label,
test_predict_label)
print(ACC_one_epoch, file=fptestACC)
if ACC_one_epoch > ACC:
Out_Result(opath, net, train_predict_label, train_true_label,
test_predict_label, test_true_label, epoch)
ACC = ACC_one_epoch
test_iter_num += 1
test_loss.append(test_loss_one_epoch)
# print(train_loss)
train_loss = np.array(train_loss)
np.savetxt(os.path.join(opath, "train_loss.csv"),
train_loss,
delimiter=',',
fmt="%.04f")
test_loss = np.array(test_loss)
np.savetxt(os.path.join(opath, "test_loss.csv"),
test_loss,
delimiter=',',
fmt="%.04f")
writer.close()
def Run_8wdata_Diff_WindowSize():
'''此算法可一次导出所有窗口的所有结果'''
opath = "D:\\ZHR_USE\\1DCNN\\Results\\"
ACC_list = []
for i in range(7, 10, 2):
subpath = os.path.join(opath, "SGD_" + str(i) + "windowsize_epochs400")
if not os.path.exists(subpath):
os.mkdir(subpath)
trainloader, testloader = L8w.Get_Data_Logging_1DCNN(i)
flat_len = int(((i * 2 + 1 - 7) / 2 - 2) / 2)
net = NET.NET_8wdata_MultiChannel_3(6, flat_len)
net.to(device)
# tensorboard
criterion = nn.CrossEntropyLoss()
optimizer_ft = optim.SGD(net.parameters(), lr=0.001)
# decay LR by a factor of 0.1 every 7 epochs
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft,
step_size=100,
gamma=0.1)
train_ACC, test_ACC = 0, 0
test_iter_num = 0
ACC = 0
train_loss = []
test_loss = []
# 输出每次迭代的精度
fptrainACC = open(os.path.join(subpath, "train_ACC.txt"), 'w')
fptrainACC = open(os.path.join(subpath, "train_ACC.txt"), 'a')
fptestACC = open(os.path.join(subpath, "test_ACC.txt"), 'w')
fptestACC = open(os.path.join(subpath, "test_ACC.txt"), 'a')
for epoch in range(2):
# for image,label in train_loader:
# print(image)
train_loss_one_epoch, train_predict_label, train_true_label = train_tool.train_one_epoch(
net,
criterion,
optimizer_ft,
trainloader,
device,
epoch,
print_freq=200)
train_loss.append(train_loss_one_epoch)
train_ACC = metrics.accuracy_score(train_true_label,
train_predict_label)
print(train_ACC, file=fptrainACC)
test_loss_one_epoch, test_predict_label, test_true_label = train_tool.evaluate(
net, criterion, testloader, device, test_iter_num)
test_loss.append(test_loss_one_epoch)
ACC_one_epoch = metrics.accuracy_score(test_true_label,
test_predict_label)
print(ACC_one_epoch, file=fptestACC)
if ACC_one_epoch > ACC:
Out_Result(subpath, net, train_predict_label, train_true_label,
test_predict_label, test_true_label, epoch)
ACC = ACC_one_epoch
# train_ACC=metrics.accuracy_score(train_true_label, train_predict_label)
# test_ACC = metrics.accuracy_score(test_true_label, test_predict_label)
# print(loss)
train_loss = np.array(train_loss)
np.savetxt(os.path.join(subpath, "train_loss.csv"),
train_loss,
delimiter=',',
fmt="%.04f")
test_loss = np.array(test_loss)
np.savetxt(os.path.join(subpath, "test_loss.csv"),
test_loss,
delimiter=',',
fmt="%.04f")
ACC = np.array((i, train_ACC, test_ACC))
ACC_list.append(ACC)
odata = np.row_stack(ACC_list)
np.savetxt(opath + "SGD400迭代次数不同窗口的ACC.txt", odata, fmt="%.04f")