def train(epoch = 400):
# 创建指标计算对象
evaluator = Evaluator(4)
# 定义最好指标miou数值，初始化为0
best_pred = 0.0
writer = SummaryWriter('tblog/deeplabv3_v0507')
# 指定第二块gpu
device = torch.device('cuda:1' if torch.cuda.is_available() else 'cpu')
# 模型建立
#deeplabv3_model = resnet152()
#deeplabv3plus_model = DeepLab(backbone='resnet', output_stride=16)
#deeplabv3_model = torch.load('checkpoints/deeplabv3_model_90.pt')
ccnet_model = resnet152()
ccnet_model = ccnet_model.to(device)
#deeplabv3_model = deeplabv3_model.to(device)
#deeplabv3plus_model = deeplabv3plus_model.to(device)
# 损失函数和优化器
criterion = nn.CrossEntropyLoss().to(device) # CrossEntropyLoss适用多分类
optimizer = optim.Adam(ccnet_model.parameters(), lr=1e-3)
for epo in range(epoch):
# 每个epoch都要记录5次交叉验证的train_loss和val_loss，最后除5
train_loss = 0
val_loss = 0
val_acc = 0
val_miou = 0
for i in range(5):
# 训练部分
ccnet_model.train()
#deeplabv3_model.train()
for index, (image, label) in enumerate(train_dataloader[i]):
image = image.to(device)
label = label.to(device)
optimizer.zero_grad()
#output = deeplabv3_model(image)
#output = deeplabv3plus_model(image)
output = ccnet_model(image)
loss = criterion(output, label)
loss.backward()
iter_loss = loss.item() # 取出数值
train_loss += iter_loss
optimizer.step()
if np.mod(index, 24) == 0:
line = "epoch {}_{}, {}/{},train loss is {}".format(epo, i, index, len(train_dataloader[i]), iter_loss)
print(line)
# 写到日志文件
with open('log/logs_deeplabv3_v0507.txt', 'a') as f :
f.write(line)
f.write('\r\n')
# 验证部分
ccnet_model.eval()
with torch.no_grad():
for index, (image, label) in enumerate(val_dataloader[i]):
image = image.to(device)
#label = label.reshape(-1, 5)
label = label.to(device)
optimizer.zero_grad()
#output = deeplabv3_model(image)
#output = deeplabv3plus_model(image)
output = ccnet_model(image)
loss = criterion(output, label)
iter_loss = loss.item()
#all_val_iter_loss.append(iter_loss)
val_loss += iter_loss
# 记录相关指标数据
pred = output.cpu().numpy()
label = label.cpu().numpy()
pred = np.argmax(pred, axis=1)
evaluator.add_batch(label, pred)
Acc = evaluator.Pixel_Accuracy()
mIoU = evaluator.Mean_Intersection_over_Union()
val_acc += Acc
val_miou += mIoU
evaluator.reset() # 该5次求指标，每次求之前先清零
line_epoch = "epoch train loss = %.3f, epoch val loss = %.3f" % (train_loss/len(train_dataloader[i])/5, val_loss/len(val_dataloader[i])/5)
print(line_epoch)
with open('log/logs_deeplabv3_v0507.txt', 'a') as f :
f.write(line_epoch)
f.write('\r\n')
#Acc = evaluator.Pixel_Accuracy()
#Acc_class = evaluator.Pixel_Accuracy_Class()
#mIoU = evaluator.Mean_Intersection_over_Union()
# tensorboard记录
writer.add_scalar('train_loss', train_loss/len(train_dataloader[i])/5, epo)
writer.add_scalar('val_loss', val_loss/len(val_dataloader[i])/5, epo)
writer.add_scalar('val_Acc', val_acc/5, epo)
#writer.add_scalar('Acc_class', Acc_class, epo)
writer.add_scalar('val_mIoU', val_miou/5, epo)
# 每次验证，根据新得出的miou指标来保存模型
#global best_pred
new_pred = val_miou/5
if new_pred > best_pred:
best_pred = new_pred
torch.save(ccnet_model.state_dict(), 'models_ccnet_v0403/ccnet_{}.pth'.format(epo))
#torch.save(deeplabv3_model.state_dict(), 'models_v0304_pre/deeplabv3_{}.pth'.format(epo))
#torch.save(deeplabv3plus_model, 'checkpoints_v3p_v0316/deeplabv3plus_model_{}.pt'.format(epo))
'''
def train(epoch=400):
# 创建指标计算对象
evaluator = Evaluator(4)
# 定义最好指标miou数值，初始化为0
best_pred = 0.0
writer = SummaryWriter('tblog/deeplabv3_v0304')
# 指定第二块gpu
device = torch.device('cuda:1' if torch.cuda.is_available() else 'cpu')
# 模型建立
deeplabv3_model = resnet50()
#deeplabv3_model = torch.load('checkpoints/deeplabv3_model_90.pt')
deeplabv3_model = deeplabv3_model.to(device)
# 损失函数和优化器
criterion = nn.CrossEntropyLoss().to(device) # CrossEntropyLoss适用多分类
optimizer = optim.Adam(deeplabv3_model.parameters(), lr=1e-3)
#all_train_iter_loss = []
#all_val_iter_loss = []
for epo in range(epoch):
# 训练部分
train_loss = 0
deeplabv3_model.train()
for index, (image, label) in enumerate(train_dataloader):
image = image.to(device)
label = label.to(device)
optimizer.zero_grad()
output = deeplabv3_model(image)
loss = criterion(output, label)
#print(loss.shape)
loss.backward()
iter_loss = loss.item() # 取出数值
#all_train_iter_loss.append(iter_loss)
train_loss += iter_loss
optimizer.step()
if np.mod(index, 8) == 0:
line = "epoch {}, {}/{},train loss is {}".format(
epo, index, len(train_dataloader), iter_loss)
print(line)
# 写到日志文件
with open('log/logs_v0304.txt', 'a') as f:
f.write(line)
f.write('\r\n')
# 验证部分
val_loss = 0
deeplabv3_model.eval()
with torch.no_grad():
for index, (image, label) in enumerate(val_dataloader):
image = image.to(device)
#label = label.reshape(-1, 5)
label = label.to(device)
optimizer.zero_grad()
output = deeplabv3_model(image)
loss = criterion(output, label)
iter_loss = loss.item()
#all_val_iter_loss.append(iter_loss)
val_loss += iter_loss
# 记录相关指标数据
pred = output.cpu().numpy()
label = label.cpu().numpy()
pred = np.argmax(pred, axis=1)
evaluator.add_batch(label, pred)
line_epoch = "epoch train loss = %.3f, epoch val loss = %.3f" % (
train_loss / len(train_dataloader), val_loss / len(val_dataloader))
print(line_epoch)
with open('log/logs_v0304.txt', 'a') as f:
f.write(line_epoch)
f.write('\r\n')
Acc = evaluator.Pixel_Accuracy()
#Acc_class = evaluator.Pixel_Accuracy_Class()
mIoU = evaluator.Mean_Intersection_over_Union()
# tensorboard记录
writer.add_scalar('train_loss', train_loss / len(train_dataloader),
epo)
writer.add_scalar('val_loss', val_loss / len(val_dataloader), epo)
writer.add_scalar('Acc', Acc, epo)
#writer.add_scalar('Acc_class', Acc_class, epo)
writer.add_scalar('mIoU', mIoU, epo)
# 每次验证，根据新得出的miou指标来保存模型
#global best_pred
new_pred = mIoU
if new_pred > best_pred:
best_pred = new_pred
torch.save(deeplabv3_model.state_dict(),
'models_v0304/deeplabv3_{}.pth'.format(epo))
'''
def train(epoch=45):
# 创建指标计算对象
evaluator = Evaluator(4)
# 定义最好指标miou数值，初始化为0
best_pred = 0.0
writer = SummaryWriter('tblog/acnet_v0928')
# 指定第二块gpu
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# 模型建立
ac_model = ACNet()
ac_model = ac_model.to(device)
# 加入权重，减轻样本不平衡情况
weight = torch.from_numpy(np.array([0.25, 0.88, 0.90, 0.97])).float()
# 损失函数和优化器
criterion = nn.CrossEntropyLoss(weight=weight).to(
device) # CrossEntropyLoss适用多分类
optimizer = optim.Adam(ac_model.parameters(), lr=1e-3)
for epo in range(epoch):
# 每个epoch都要记录5次交叉验证的train_loss和val_loss，最后除5
train_loss = 0
val_loss = 0
val_acc = 0
val_miou = 0
for i in range(5):
# 训练部分
ac_model.train()
for index, (image, dsm_data,
label) in enumerate(train_dataloader[i]):
image = image.to(device)
dsm_data = dsm_data.to(device)
label = label.to(device)
# 此句提前
optimizer.zero_grad()
output = ac_model(image, dsm_data)
loss = criterion(output, label)
loss.backward()
iter_loss = loss.item() # 取出数值
train_loss += iter_loss
optimizer.step()
if np.mod(index, 18) == 0:
line = "epoch {}_{}, {}/{},train loss is {}".format(
epo, i, index, len(train_dataloader[i]), iter_loss)
print(line)
# 写到日志文件
with open('log/logs_acnet_0928.txt', 'a') as f:
f.write(line)
f.write('\r\n')
# 验证部分
ac_model.eval()
with torch.no_grad():
for index, (image, dsm_data,
label) in enumerate(val_dataloader[i]):
image = image.to(device)
dsm_data = dsm_data.to(device)
label = label.to(device)
output = ac_model(image, dsm_data)
loss = criterion(output, label)
optimizer.zero_grad()
iter_loss = loss.item()
val_loss += iter_loss
# 记录相关指标数据
pred = output.cpu().numpy()
label = label.cpu().numpy()
pred = np.argmax(pred, axis=1)
evaluator.add_batch(label, pred)
Acc = evaluator.Pixel_Accuracy()
mIoU = evaluator.Mean_Intersection_over_Union()
val_acc += Acc
val_miou += mIoU
evaluator.reset() # 该5次求指标，每次求之前先清零
line_epoch = "epoch train loss = %.3f, epoch val loss = %.3f" % (
train_loss / len(train_dataloader[i]) / 5,
val_loss / len(val_dataloader[i]) / 5)
print(line_epoch)
with open('log/logs_acnet_0928.txt', 'a') as f:
f.write(line_epoch)
f.write('\r\n')
# tensorboard记录
writer.add_scalar('train_loss',
train_loss / len(train_dataloader[i]) / 5, epo)
writer.add_scalar('val_loss', val_loss / len(val_dataloader[i]) / 5,
epo)
writer.add_scalar('val_Acc', val_acc / 5, epo)
#writer.add_scalar('Acc_class', Acc_class, epo)
writer.add_scalar('val_mIoU', val_miou / 5, epo)
# 每次验证，根据新得出的miou指标来保存模型
#global best_pred
new_pred = val_miou / 5
if new_pred > best_pred:
best_pred = new_pred
torch.save(ac_model.state_dict(),
'models_acnet_v0928/acnet_0928_{}.pth'.format(epo))
print("Training completed ")