def main(parser_data):
device = torch.device(
parser_data.device if torch.cuda.is_available() else "cpu")
print("Using {} device training.".format(device.type))
if not os.path.exists("save_weights"):
os.mkdir("save_weights")
data_transform = {
"train":
transform.Compose([
transform.SSDCropping(),
transform.Resize(),
transform.ColorJitter(),
transform.ToTensor(),
transform.RandomHorizontalFlip(),
transform.Normalization(),
transform.AssignGTtoDefaultBox()
]),
"val":
transform.Compose([
transform.Resize(),
transform.ToTensor(),
transform.Normalization()
])
}
VOC_root = parser_data.data_path
# check voc root
if os.path.exists(os.path.join(VOC_root, "VOCdevkit")) is False:
raise FileNotFoundError(
"VOCdevkit dose not in path:'{}'.".format(VOC_root))
train_dataset = VOC2012DataSet(VOC_root,
data_transform['train'],
train_set='train.txt')
# 注意训练时，batch_size必须大于1
batch_size = parser_data.batch_size
assert batch_size > 1, "batch size must be greater than 1"
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
print('Using %g dataloader workers' % nw)
train_data_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=nw,
collate_fn=train_dataset.collate_fn)
val_dataset = VOC2012DataSet(VOC_root,
data_transform['val'],
train_set='val.txt')
val_data_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=nw,
collate_fn=train_dataset.collate_fn)
model = create_model(num_classes=21, device=device)
model.to(device)
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.0005,
momentum=0.9,
weight_decay=0.0005)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=5,
gamma=0.3)
# 如果指定了上次训练保存的权重文件地址，则接着上次结果接着训练
if parser_data.resume != "":
checkpoint = torch.load(parser_data.resume)
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
parser_data.start_epoch = checkpoint['epoch'] + 1
print("the training process from epoch{}...".format(
parser_data.start_epoch))
train_loss = []
learning_rate = []
val_map = []
val_data = None
# 如果电脑内存充裕，可提前加载验证集数据，以免每次验证时都要重新加载一次数据，节省时间
# val_data = get_coco_api_from_dataset(val_data_loader.dataset)
for epoch in range(parser_data.start_epoch, parser_data.epochs):
utils.train_one_epoch(model=model,
optimizer=optimizer,
data_loader=train_data_loader,
device=device,
epoch=epoch,
print_freq=50,
train_loss=train_loss,
train_lr=learning_rate)
lr_scheduler.step()
utils.evaluate(model=model,
data_loader=val_data_loader,
device=device,
data_set=val_data,
mAP_list=val_map)
# save weights
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch
}
torch.save(save_files, "./save_weights/ssd300-{}.pth".format(epoch))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
def main():
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print("Using {} device training.".format(device.type))
# 用来保存coco_info的文件
results_file = "results{}.txt".format(
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
# 检查保存权重文件夹是否存在，不存在则创建
if not os.path.exists("save_weights"):
os.makedirs("save_weights")
data_transform = {
"train":
transforms.Compose(
[transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
"val":
transforms.Compose([transforms.ToTensor()])
}
COCO_root = "/data/coco2017"
# load train data set
# coco2017 -> annotations -> instances_train2017.json
train_data_set = CocoDetection(COCO_root, "train", data_transform["train"])
# 注意这里的collate_fn是自定义的，因为读取的数据包括image和targets，不能直接使用默认的方法合成batch
batch_size = 8
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
print('Using %g dataloader workers' % nw)
train_data_loader = torch.utils.data.DataLoader(
train_data_set,
batch_size=batch_size,
shuffle=True,
pin_memory=True,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
# load validation data set
# coco2017 -> annotations -> instances_val2017.json
val_data_set = CocoDetection(COCO_root, "val", data_transform["val"])
val_data_set_loader = torch.utils.data.DataLoader(
val_data_set,
batch_size=batch_size,
shuffle=False,
pin_memory=True,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
# create model num_classes equal background + 80 classes
model = create_model(num_classes=81)
# print(model)
model.to(device)
train_loss = []
learning_rate = []
val_map = []
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.001,
momentum=0.9,
weight_decay=1e-4)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer,
milestones=[16, 22],
gamma=0.1)
num_epochs = 26
for epoch in range(num_epochs):
# train for one epoch, printing every 50 iterations
mean_loss, lr = utils.train_one_epoch(model,
optimizer,
train_data_loader,
device,
epoch,
print_freq=50,
warmup=True)
train_loss.append(mean_loss.item())
learning_rate.append(lr)
# update the learning rate
lr_scheduler.step()
# evaluate on the test dataset
coco_info = utils.evaluate(model, val_data_set_loader, device=device)
# write into txt
with open(results_file, "a") as f:
# 写入的数据包括coco指标还有loss和learning rate
result_info = [
str(round(i, 4)) for i in coco_info + [mean_loss.item(), lr]
]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
val_map.append(coco_info[1]) # pascal mAP
# save weights
# 仅保存最后5个epoch的权重
if epoch in range(num_epochs)[-5:]:
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch
}
torch.save(save_files,
"./save_weights/mobile-model-{}.pth".format(epoch))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
def main(args):
print(args)
# mp.spawn(main_worker, args=(args,), nprocs=args.world_size, join=True)
init_distributed_mode(args)
device = torch.device(args.device)
results_file = "results{}.txt".format(
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
# Data loading code
print("Loading data")
data_transform = {
"train":
transform.Compose([
transform.SSDCropping(),
transform.Resize(),
transform.ColorJitter(),
transform.ToTensor(),
transform.RandomHorizontalFlip(),
transform.Normalization(),
transform.AssignGTtoDefaultBox()
]),
"val":
transform.Compose([
transform.Resize(),
transform.ToTensor(),
transform.Normalization()
])
}
VOC_root = args.data_path
# check voc root
if os.path.exists(os.path.join(VOC_root, "VOCdevkit")) is False:
raise FileNotFoundError(
"VOCdevkit dose not in path:'{}'.".format(VOC_root))
# load train data set
train_data_set = VOC2012DataSet(VOC_root,
data_transform["train"],
train_set='train.txt')
# load validation data set
val_data_set = VOC2012DataSet(VOC_root,
data_transform["val"],
train_set='val.txt')
print("Creating data loaders")
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_data_set)
test_sampler = torch.utils.data.distributed.DistributedSampler(
val_data_set)
else:
train_sampler = torch.utils.data.RandomSampler(train_data_set)
test_sampler = torch.utils.data.SequentialSampler(val_data_set)
if args.aspect_ratio_group_factor >= 0:
# 统计所有图像比例在bins区间中的位置索引
group_ids = create_aspect_ratio_groups(
train_data_set, k=args.aspect_ratio_group_factor)
train_batch_sampler = GroupedBatchSampler(train_sampler, group_ids,
args.batch_size)
else:
train_batch_sampler = torch.utils.data.BatchSampler(train_sampler,
args.batch_size,
drop_last=True)
data_loader = torch.utils.data.DataLoader(
train_data_set,
batch_sampler=train_batch_sampler,
num_workers=args.workers,
collate_fn=train_data_set.collate_fn)
data_loader_test = torch.utils.data.DataLoader(
val_data_set,
batch_size=1,
sampler=test_sampler,
num_workers=args.workers,
collate_fn=train_data_set.collate_fn)
print("Creating model")
model = create_model(num_classes=args.num_classes + 1, device=device)
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
model_without_ddp = model.module
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=args.lr_step_size,
gamma=args.lr_gamma)
# lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=args.lr_steps, gamma=args.lr_gamma)
# 如果传入resume参数，即上次训练的权重地址，则接着上次的参数训练
if args.resume:
# If map_location is missing, torch.load will first load the module to CPU
# and then copy each parameter to where it was saved,
# which would result in all processes on the same machine using the same set of devices.
checkpoint = torch.load(
args.resume, map_location='cpu') # 读取之前保存的权重文件(包括优化器以及学习率策略)
model_without_ddp.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
if args.test_only:
utils.evaluate(model, data_loader_test, device=device)
return
train_loss = []
learning_rate = []
val_map = []
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
mean_loss, lr = utils.train_one_epoch(model, optimizer, data_loader,
device, epoch, args.print_freq)
# only first process to save training info
if args.rank in [-1, 0]:
train_loss.append(mean_loss.item())
learning_rate.append(lr)
# update learning rate
lr_scheduler.step()
# evaluate after every epoch
coco_info = utils.evaluate(model, data_loader_test, device=device)
if args.rank in [-1, 0]:
# write into txt
with open(results_file, "a") as f:
result_info = [
str(round(i, 4))
for i in coco_info + [mean_loss.item(), lr]
]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
val_map.append(coco_info[1]) # pascal mAP
if args.output_dir:
# 只在主节点上执行保存权重操作
save_on_master(
{
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'args': args,
'epoch': epoch
}, os.path.join(args.output_dir, 'model_{}.pth'.format(epoch)))
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
if args.rank in [-1, 0]:
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
def main(args):
init_distributed_mode(args)
print(args)
device = torch.device(args.device)
# 用来保存coco_info的文件
results_file = "results{}.txt".format(
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
# Data loading code
print("Loading data")
data_transform = {
"train":
transforms.Compose(
[transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
"val":
transforms.Compose([transforms.ToTensor()])
}
COCO_root = args.data_path
# load train data set
# coco2017 -> annotations -> instances_train2017.json
train_data_set = CocoDetection(COCO_root, "train", data_transform["train"])
# load validation data set
# coco2017 -> annotations -> instances_val2017.json
val_data_set = CocoDetection(COCO_root, "val", data_transform["val"])
print("Creating data loaders")
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_data_set)
test_sampler = torch.utils.data.distributed.DistributedSampler(
val_data_set)
else:
train_sampler = torch.utils.data.RandomSampler(train_data_set)
test_sampler = torch.utils.data.SequentialSampler(val_data_set)
if args.aspect_ratio_group_factor >= 0:
# 统计所有图像比例在bins区间中的位置索引
group_ids = create_aspect_ratio_groups(
train_data_set, k=args.aspect_ratio_group_factor)
train_batch_sampler = GroupedBatchSampler(train_sampler, group_ids,
args.batch_size)
else:
train_batch_sampler = torch.utils.data.BatchSampler(train_sampler,
args.batch_size,
drop_last=True)
data_loader = torch.utils.data.DataLoader(
train_data_set,
batch_sampler=train_batch_sampler,
num_workers=args.workers,
collate_fn=train_data_set.collate_fn)
data_loader_test = torch.utils.data.DataLoader(
val_data_set,
batch_size=1,
sampler=test_sampler,
num_workers=args.workers,
collate_fn=train_data_set.collate_fn)
print("Creating model")
# create model num_classes equal background + 80 classes
model = create_model(num_classes=args.num_classes + 1, device=device)
model.to(device)
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
model_without_ddp = model.module
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=args.lr_step_size, gamma=args.lr_gamma)
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
optimizer, milestones=args.lr_steps, gamma=args.lr_gamma)
# 如果传入resume参数，即上次训练的权重地址，则接着上次的参数训练
if args.resume:
# If map_location is missing, torch.load will first load the module to CPU
# and then copy each parameter to where it was saved,
# which would result in all processes on the same machine using the same set of devices.
checkpoint = torch.load(
args.resume, map_location='cpu') # 读取之前保存的权重文件(包括优化器以及学习率策略)
model_without_ddp.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
train_loss = []
learning_rate = []
val_map = []
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
mean_loss, lr = utils.train_one_epoch(model,
optimizer,
data_loader,
device,
epoch,
args.print_freq,
warmup=True)
# update learning rate
lr_scheduler.step()
# evaluate after every epoch
coco_info = utils.evaluate(model, data_loader_test, device=device)
# 只在主进程上进行写操作
if args.rank in [-1, 0]:
train_loss.append(mean_loss.item())
learning_rate.append(lr)
val_map.append(coco_info[1]) # pascal mAP
# write into txt
with open(results_file, "a") as f:
# 写入的数据包括coco指标还有loss和learning rate
result_info = [
str(round(i, 4))
for i in coco_info + [mean_loss.item(), lr]
]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
if args.output_dir:
# 只在主节点上执行保存权重操作
save_on_master(
{
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'args': args,
'epoch': epoch
}, os.path.join(args.output_dir, 'model_{}.pth'.format(epoch)))
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
if args.rank in [-1, 0]:
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
def main(parser_data):
device = torch.device(
parser_data.device if torch.cuda.is_available() else "cpu")
print(device)
if not os.path.exists("save_weights"):
os.mkdir("save_weights")
data_transform = {
"train":
transform.Compose([
transform.SSDCropping(),
transform.Resize(),
transform.ColorJitter(),
transform.ToTensor(),
transform.RandomHorizontalFlip(),
transform.Normalization(),
transform.AssignGTtoDefaultBox()
]),
"val":
transform.Compose([
transform.Resize(),
transform.ToTensor(),
transform.Normalization()
])
}
night_root = parser_data.data_path
train_dataset = NightDataSet(night_root,
data_transform['train'],
train_set='train.txt')
# aa = train_dataset[1]
# 注意训练时，batch_size必须大于1
train_data_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=8,
shuffle=True,
num_workers=4,
collate_fn=utils.collate_fn)
val_dataset = NightDataSet(night_root,
data_transform['val'],
train_set='val.txt')
# bb = val_dataset[2]
val_data_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=4,
shuffle=False,
num_workers=0,
collate_fn=utils.collate_fn)
model = create_model(num_classes=3, device=device)
print(model)
model.to(device)
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.005,
momentum=0.9,
weight_decay=0.0005)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=5,
gamma=0.5)
# 如果指定了上次训练保存的权重文件地址，则接着上次结果接着训练
if parser_data.resume != "":
checkpoint = torch.load(parser_data.resume)
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
parser_data.start_epoch = checkpoint['epoch'] + 1
print("the training process from epoch{}...".format(
parser_data.start_epoch))
train_loss = []
learning_rate = []
val_map = []
train_val_map = []
val_data = None
# 如果电脑内存充裕，可提前加载验证集数据，以免每次验证时都要重新加载一次数据，节省时间
# val_data = get_coco_api_from_dataset(val_data_loader.dataset)
for epoch in range(parser_data.start_epoch, parser_data.epochs):
utils.train_one_epoch(model=model,
optimizer=optimizer,
data_loader=train_data_loader,
device=device,
epoch=epoch,
print_freq=50,
train_loss=train_loss,
train_lr=learning_rate)
lr_scheduler.step()
if epoch >= 20 or epoch == 10:
utils.evaluate(model=model,
data_loader=val_data_loader,
device=device,
data_set=val_data,
mAP_list=val_map)
# save weights
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch
}
torch.save(save_files, "./save_weights/ssd512-{}.pth".format(epoch))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
def main():
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print("Using {} device training.".format(device.type))
# 用来保存coco_info的文件
results_file = "results{}.txt".format(
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
# 检查保存权重文件夹是否存在，不存在则创建
if not os.path.exists("save_weights"):
os.makedirs("save_weights")
data_transform = {
"train":
transforms.Compose(
[transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
"val":
transforms.Compose([transforms.ToTensor()])
}
VOC_root = "./"
# check voc root
if os.path.exists(os.path.join(VOC_root, "VOCdevkit")) is False:
raise FileNotFoundError(
"VOCdevkit dose not in path:'{}'.".format(VOC_root))
# load train data set
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> train.txt
train_data_set = VOC2012DataSet(VOC_root, data_transform["train"],
"train.txt")
# 注意这里的collate_fn是自定义的，因为读取的数据包括image和targets，不能直接使用默认的方法合成batch
batch_size = 8
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
print('Using %g dataloader workers' % nw)
train_data_loader = torch.utils.data.DataLoader(
train_data_set,
batch_size=batch_size,
shuffle=True,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
# load validation data set
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> val.txt
val_data_set = VOC2012DataSet(VOC_root, data_transform["val"], "val.txt")
val_data_set_loader = torch.utils.data.DataLoader(
val_data_set,
batch_size=batch_size,
shuffle=False,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
# create model num_classes equal background + 20 classes
model = create_model(num_classes=21)
# print(model)
model.to(device)
train_loss = []
learning_rate = []
val_map = []
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #
# first frozen backbone and train 5 epochs #
# 首先冻结前置特征提取网络权重（backbone），训练rpn以及最终预测网络部分 #
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #
for param in model.backbone.parameters():
param.requires_grad = False
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.005,
momentum=0.9,
weight_decay=0.0005)
init_epochs = 5
for epoch in range(init_epochs):
# train for one epoch, printing every 10 iterations
mean_loss, lr = utils.train_one_epoch(model,
optimizer,
train_data_loader,
device,
epoch,
print_freq=50,
warmup=True)
train_loss.append(mean_loss.item())
learning_rate.append(lr)
# evaluate on the test dataset
coco_info = utils.evaluate(model, val_data_set_loader, device=device)
# write into txt
with open(results_file, "a") as f:
# 写入的数据包括coco指标还有loss和learning rate
result_info = [
str(round(i, 4)) for i in coco_info + [mean_loss.item(), lr]
]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
val_map.append(coco_info[1]) # pascal mAP
torch.save(model.state_dict(), "./save_weights/pretrain.pth")
# # # # # # # # # # # # # # # # # # # # # # # # # # # #
# second unfrozen backbone and train all network #
# 解冻前置特征提取网络权重（backbone），接着训练整个网络权重 #
# # # # # # # # # # # # # # # # # # # # # # # # # # # #
# 冻结backbone部分底层权重
for name, parameter in model.backbone.named_parameters():
split_name = name.split(".")[0]
if split_name in ["0", "1", "2", "3"]:
parameter.requires_grad = False
else:
parameter.requires_grad = True
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.005,
momentum=0.9,
weight_decay=0.0005)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=3,
gamma=0.33)
num_epochs = 20
for epoch in range(init_epochs, num_epochs + init_epochs, 1):
# train for one epoch, printing every 50 iterations
mean_loss, lr = utils.train_one_epoch(model,
optimizer,
train_data_loader,
device,
epoch,
print_freq=50)
train_loss.append(mean_loss.item())
learning_rate.append(lr)
# update the learning rate
lr_scheduler.step()
# evaluate on the test dataset
coco_info = utils.evaluate(model, val_data_set_loader, device=device)
# write into txt
with open(results_file, "a") as f:
# 写入的数据包括coco指标还有loss和learning rate
result_info = [
str(round(i, 4)) for i in coco_info + [mean_loss.item(), lr]
]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
val_map.append(coco_info[1]) # pascal mAP
# save weights
# 仅保存最后5个epoch的权重
if epoch in range(num_epochs + init_epochs)[-5:]:
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch
}
torch.save(save_files,
"./save_weights/mobile-model-{}.pth".format(epoch))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
def main():
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(device)
# 检查保存权重文件夹是否存在，不存在则创建
if not os.path.exists("save_weights"):
os.makedirs("save_weights")
data_transform = {
"train": transforms.Compose([transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
"val": transforms.Compose([transforms.ToTensor()])
}
VOC_root = "./"
# load train data set
train_data_set = VOC2012DataSet(VOC_root, data_transform["train"], True)
# 注意这里的collate_fn是自定义的，因为读取的数据包括image和targets，不能直接使用默认的方法合成batch
train_data_loader = torch.utils.data.DataLoader(train_data_set,
batch_size=8,
shuffle=True,
num_workers=0,
collate_fn=utils.collate_fn)
# load validation data set
val_data_set = VOC2012DataSet(VOC_root, data_transform["val"], False)
val_data_set_loader = torch.utils.data.DataLoader(val_data_set,
batch_size=1,
shuffle=False,
num_workers=0,
collate_fn=utils.collate_fn)
# create model num_classes equal background + 20 classes
model = create_model(num_classes=21)
print(model)
model.to(device)
train_loss = []
learning_rate = []
val_mAP = []
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #
# first frozen backbone and train 5 epochs #
# 首先冻结前置特征提取网络权重（backbone），训练rpn以及最终预测网络部分 #
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #
for param in model.backbone.parameters():
param.requires_grad = False
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params, lr=0.005,
momentum=0.9, weight_decay=0.0005)
num_epochs = 5
for epoch in range(num_epochs):
# train for one epoch, printing every 10 iterations
utils.train_one_epoch(model, optimizer, train_data_loader,
device, epoch, print_freq=50,
train_loss=train_loss, train_lr=learning_rate)
# evaluate on the test dataset
utils.evaluate(model, val_data_set_loader, device=device, mAP_list=val_mAP)
torch.save(model.state_dict(), "./save_weights/pretrain.pth")
# # # # # # # # # # # # # # # # # # # # # # # # # # # #
# second unfrozen backbone and train all network #
# 解冻前置特征提取网络权重（backbone），接着训练整个网络权重 #
# # # # # # # # # # # # # # # # # # # # # # # # # # # #
# 冻结backbone部分底层权重
for name, parameter in model.backbone.named_parameters():
split_name = name.split(".")[0]
if split_name in ["0", "1", "2", "3"]:
parameter.requires_grad = False
else:
parameter.requires_grad = True
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params, lr=0.005,
momentum=0.9, weight_decay=0.0005)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=5,
gamma=0.33)
num_epochs = 20
for epoch in range(num_epochs):
# train for one epoch, printing every 50 iterations
utils.train_one_epoch(model, optimizer, train_data_loader,
device, epoch, print_freq=50,
train_loss=train_loss, train_lr=learning_rate)
# update the learning rate
lr_scheduler.step()
# evaluate on the test dataset
utils.evaluate(model, val_data_set_loader, device=device, mAP_list=val_mAP)
# save weights
if epoch > 10:
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch}
torch.save(save_files, "./save_weights/mobile-model-{}.pth".format(epoch))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_mAP) != 0:
from plot_curve import plot_map
plot_map(val_mAP)
def main(parser_data):
device = torch.device(
parser_data.device if torch.cuda.is_available() else "cpu")
print("Using {} device training.".format(device.type))
if not os.path.exists("save_weights"):
os.mkdir("save_weights")
results_file = "results{}.txt".format(
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
data_transform = {
"train":
transforms.Compose([
transforms.SSDCropping(),
transforms.Resize(),
transforms.ColorJitter(),
transforms.ToTensor(),
transforms.RandomHorizontalFlip(),
transforms.Normalization(),
transforms.AssignGTtoDefaultBox()
]),
"val":
transforms.Compose([
transforms.Resize(),
transforms.ToTensor(),
transforms.Normalization()
])
}
VOC_root = parser_data.data_path
# check voc root
if os.path.exists(os.path.join(VOC_root, "VOCdevkit")) is False:
raise FileNotFoundError(
"VOCdevkit dose not in path:'{}'.".format(VOC_root))
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> train.txt
train_dataset = VOCDataSet(VOC_root,
"2012",
data_transform['train'],
train_set='train.txt')
# 注意训练时，batch_size必须大于1
batch_size = parser_data.batch_size
assert batch_size > 1, "batch size must be greater than 1"
# 防止最后一个batch_size=1，如果最后一个batch_size=1就舍去
drop_last = True if len(train_dataset) % batch_size == 1 else False
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
print('Using %g dataloader workers' % nw)
train_data_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=nw,
collate_fn=train_dataset.collate_fn,
drop_last=drop_last)
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> val.txt
val_dataset = VOCDataSet(VOC_root,
"2012",
data_transform['val'],
train_set='val.txt')
val_data_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=nw,
collate_fn=train_dataset.collate_fn)
model = create_model(num_classes=args.num_classes + 1)
model.to(device)
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.0005,
momentum=0.9,
weight_decay=0.0005)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=5,
gamma=0.3)
# 如果指定了上次训练保存的权重文件地址，则接着上次结果接着训练
if parser_data.resume != "":
checkpoint = torch.load(parser_data.resume, map_location='cpu')
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
parser_data.start_epoch = checkpoint['epoch'] + 1
print("the training process from epoch{}...".format(
parser_data.start_epoch))
train_loss = []
learning_rate = []
val_map = []
# 提前加载验证集数据，以免每次验证时都要重新加载一次数据，节省时间
val_data = get_coco_api_from_dataset(val_data_loader.dataset)
for epoch in range(parser_data.start_epoch, parser_data.epochs):
mean_loss, lr = utils.train_one_epoch(model=model,
optimizer=optimizer,
data_loader=train_data_loader,
device=device,
epoch=epoch,
print_freq=50)
train_loss.append(mean_loss.item())
learning_rate.append(lr)
# update learning rate
lr_scheduler.step()
coco_info = utils.evaluate(model=model,
data_loader=val_data_loader,
device=device,
data_set=val_data)
# write into txt
with open(results_file, "a") as f:
# 写入的数据包括coco指标还有loss和learning rate
result_info = [
str(round(i, 4)) for i in coco_info + [mean_loss.item()]
] + [str(round(lr, 6))]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
val_map.append(coco_info[1]) # pascal mAP
# save weights
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch
}
torch.save(save_files, "./save_weights/ssd300-{}.pth".format(epoch))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
def main(parser_data):
device = torch.device(
parser_data.device if torch.cuda.is_available() else "cpu")
print(device)
if not os.path.exists("save_weights"):
os.mkdir("save_weights")
data_transform = {
"train":
transform.Compose([
transform.SSDCropping(),
transform.Resize(),
transform.ColorJitter(),
transform.ToTensor(),
transform.RandomHorizontalFlip(),
transform.Normalization(),
transform.AssignGTtoDefaultBox()
]),
"val":
transform.Compose([
transform.Resize(),
transform.ToTensor(),
transform.Normalization()
])
}
XRay_root = parser_data.data_path
train_dataset = XRayDataset(XRay_root,
data_transform['train'],
train_set='train.txt')
# Note that the batch_size must be greater than 1
train_data_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=8,
shuffle=True,
num_workers=4,
collate_fn=utils.collate_fn)
val_dataset = XRayDataset(XRay_root,
data_transform['val'],
train_set='val.txt')
val_data_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=1,
shuffle=False,
num_workers=0,
collate_fn=utils.collate_fn)
model = create_model(num_classes=6, device=device)
model.to(device)
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.0005,
momentum=0.9,
weight_decay=0.0005)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=5,
gamma=0.3)
# If the address of the weight file saved by the last training is specified, the training continues with the last result
if parser_data.resume != "":
checkpoint = torch.load(parser_data.resume)
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
parser_data.start_epoch = checkpoint['epoch'] + 1
print("the training process from epoch{}...".format(
parser_data.start_epoch))
train_loss = []
learning_rate = []
val_map = []
val_data = None
# If your computer has sufficient memory, you can save time by loading the validation set data in advance to avoid having to reload the data each time you validate
# val_data = get_coco_api_from_dataset(val_data_loader.dataset)
for epoch in range(parser_data.start_epoch, parser_data.epochs):
utils.train_one_epoch(model=model,
optimizer=optimizer,
data_loader=train_data_loader,
device=device,
epoch=epoch,
print_freq=50,
train_loss=train_loss,
train_lr=learning_rate)
lr_scheduler.step()
utils.evaluate(model=model,
data_loader=val_data_loader,
device=device,
data_set=val_data,
mAP_list=val_map)
# save weights
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch
}
torch.save(save_files, "./save_weights/ssd300-{}.pth".format(epoch))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
cocoEval.evaluate()
cocoEval.accumulate()
cocoEval.summarize()
ap_per_category(self.cocoGt, cocoEval, epoch_index)
print_txt = cocoEval.stats
coco_mAP = print_txt[0]
voc_mAP = print_txt[1]
if isinstance(mAP_list, list):
mAP_list.append(voc_mAP)
if __name__ == '__main__':
train_loss = []
learning_rate = []
val_mAP = []
trainer = _Trainer()
for epoch_index in range(cfg.max_epoch):
trainer.train_one_epoch(epoch_index,
train_loss=train_loss,
train_lr=learning_rate)
trainer.eval(epoch_index, mAP_list=val_mAP)
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_mAP) != 0:
plot_map(val_mAP)
def main(args):
# utils.init_distributed_mode(args)
# print(args)
device = torch.device(args.device if torch.cuda.is_available() else "cpu")
print(device)
# # 用来保存coco_info的文件
results_file = "results{}.txt".format(
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
# # Data loading code
print("Loading data")
data_transform = {
"train":
transforms.Compose(
[transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
"val":
transforms.Compose([transforms.ToTensor()])
}
coco_root = args.data_path
# check root
if os.path.exists(coco_root) is False:
raise FileNotFoundError(
"coco dose not in path:'{}'.".format(coco_root))
batch_size = args.batch_size
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
0]) # number of workers
print('Using %g dataloader workers' % nw)
val_dataset = get_coco(coco_root, "val", data_transform["val"])
print("Creating data loaders")
# if args.distributed:
# train_sampler = torch.utils.data.distributed.DistributedSampler(dataset)
# test_sampler = torch.utils.data.distributed.DistributedSampler(val_dataset)
# else:
# train_sampler = torch.utils.data.RandomSampler(dataset)
# test_sampler = torch.utils.data.SequentialSampler(val_dataset)
# if args.aspect_ratio_group_factor >= 0:
# group_ids = create_aspect_ratio_groups(dataset, k=args.aspect_ratio_group_factor)
# train_batch_sampler = GroupedBatchSampler(train_sampler, group_ids, args.batch_size)
# else:
# train_batch_sampler = torch.utils.data.BatchSampler(train_sampler, args.batch_size, drop_last=True)
val_data_set_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=batch_size,
shuffle=False,
pin_memory=True,
num_workers=nw,
collate_fn=val_dataset.collate_fn)
print("Creating model")
model = get_model(num_classes=args.num_classes + 1)
model.to(device)
print(model)
model_without_ddp = model
# if args.distributed:
# model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu])
# model_without_ddp = model.module
train_loss = []
learning_rate = []
val_map = []
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=args.lr_step_size, gamma=args.lr_gamma)
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
optimizer, milestones=args.lr_steps, gamma=args.lr_gamma)
if args.resume:
checkpoint = torch.load(args.resume, map_location=device)
model_without_ddp.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
if args.test_only:
train_eval_utils.evaluate(model, val_data_set_loader, device=device)
return
for epoch in range(args.start_epoch, args.epochs):
metric_logger = train_eval_utils.train_one_epoch(
model, optimizer, val_data_set_loader, device, epoch,
args.print_freq)
mean_loss, lr = metric_logger["mloss"], metric_logger["lr"]
train_loss.append(mean_loss)
learning_rate.append(lr)
lr_scheduler.step()
# evaluate after every epoch
coco_info = train_eval_utils.evaluate(model,
val_data_set_loader,
device=device)
# write into txt
with open(results_file, "a") as f:
# 写入的数据包括coco指标还有loss和learning rate
result_info = [
str(round(i, 4)) for i in coco_info + [mean_loss, lr]
]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
val_map.append(coco_info[1]) # pascal mAP
if args.output_dir:
utils.save_on_master(
{
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'args': args,
'epoch': epoch
},
os.path.join(args.output_dir,
'resnet-fpn-model-{}.pth'.format(epoch)))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
def main():
device = torch.device(cfg['device_name'])
print("Using {} device training.".format(device.type))
if not os.path.exists(cfg['save_dir']):
os.makedirs(cfg['save_dir'])
data_transform = {
"train": Compose([ToTensor(), RandomHorizontalFlip(0.5)]),
"val": Compose([ToTensor()])
}
if not os.path.exists(cfg['data_root_dir']):
raise FileNotFoundError("dataset root dir not exist!")
# load train data set
train_data_set = coco(cfg['data_root_dir'], 'train', '2017',
data_transform["train"])
batch_size = cfg['batch_size']
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
print('Using {} dataloader workers'.format(nw))
train_data_loader = torch.utils.data.DataLoader(
train_data_set,
batch_size=batch_size,
shuffle=True,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
# load validation data set
val_data_set = coco(cfg['data_root_dir'], 'val', '2017',
data_transform["train"])
val_data_set_loader = torch.utils.data.DataLoader(
val_data_set,
batch_size=batch_size,
shuffle=False,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
# create model num_classes equal background + 20 classes
model = create_model(num_classes=cfg['num_class'])
model.to(device)
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.005,
momentum=0.9,
weight_decay=0.0005)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=5,
gamma=0.33)
# 如果指定了上次训练保存的权重文件地址，则接着上次结果接着训练
if cfg['resume'] != "":
checkpoint = torch.load(cfg['resume'])
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
cfg['start_epoch'] = checkpoint['epoch'] + 1
print("the training process from epoch{}...".format(
parser_data.start_epoch))
train_loss = []
learning_rate = []
val_mAP = []
for epoch in range(cfg['start_epoch'], cfg['num_epochs']):
# train for one epoch, printing every 10 iterations
utils.train_one_epoch(model,
optimizer,
train_data_loader,
device,
epoch,
train_loss=train_loss,
train_lr=learning_rate,
print_freq=50,
warmup=True)
# update the learning rate
lr_scheduler.step()
# evaluate on the test dataset
utils.evaluate(model,
val_data_set_loader,
device=device,
mAP_list=val_mAP)
# save weights
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch
}
model_save_dir = cfg['model_save_dir']
if not os.path.exists(model_save_dir):
os.makedirs(model_save_dir)
torch.save(
save_files,
os.path.join(model_save_dir,
"{}-model-{}.pth".format(cfg['backbone'], epoch)))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_mAP) != 0:
from plot_curve import plot_map
plot_map(val_mAP)
def main(parser_data):
device = torch.device(
parser_data.device if torch.cuda.is_available() else "cpu")
print("Using {} device training.".format(device.type))
data_transform = {
"train":
transforms.Compose(
[transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
"val":
transforms.Compose([transforms.ToTensor()])
}
Data_root = parser_data.data_path
# check voc root
if os.path.exists(os.path.join(Data_root,
"TGK_DATASET")) is False: # 修改数据集的名字
raise FileNotFoundError(
"TGK_DATASET dose not in path:'{}'.".format(Data_root))
# load train data set
train_data_set = TGKDataSet(Data_root, data_transform["train"], True)
# 注意这里的collate_fn是自定义的，因为读取的数据包括image和targets，不能直接使用默认的方法合成batch
batch_size = parser_data.batch_size
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
print('Using %g dataloader workers' % nw)
train_data_loader = torch.utils.data.DataLoader(
train_data_set,
batch_size=batch_size,
shuffle=True,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
# load validation data set
val_data_set = TGKDataSet(Data_root, data_transform["val"], False)
val_data_set_loader = torch.utils.data.DataLoader(
val_data_set,
batch_size=batch_size,
shuffle=False,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
# create model num_classes equal background + classes number
model = create_model(num_classes=2) # 修改标注类别数，需要+1
# print(model)
model.to(device)
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.0025,
momentum=0.9,
weight_decay=0.0005)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=5,
gamma=0.33)
# 如果指定了上次训练保存的权重文件地址，则接着上次结果接着训练
if parser_data.resume != "":
checkpoint = torch.load(parser_data.resume)
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
parser_data.start_epoch = checkpoint['epoch'] + 1
print("the training process from epoch{}...".format(
parser_data.start_epoch))
train_loss = []
learning_rate = []
val_mAP = []
for epoch in range(parser_data.start_epoch, parser_data.epochs):
# train for one epoch, printing every 10 iterations
utils.train_one_epoch(model,
optimizer,
train_data_loader,
device,
epoch,
train_loss=train_loss,
train_lr=learning_rate,
print_freq=10,
warmup=True)
# update the learning rate
lr_scheduler.step()
# evaluate on the test dataset
utils.evaluate(model,
val_data_set_loader,
device=device,
mAP_list=val_mAP)
# save weights
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch
}
torch.save(save_files,
"./save_weights/resNetFpn-model-{}.pth".format(epoch))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_mAP) != 0:
from plot_curve import plot_map
plot_map(val_mAP)
def main(args):
device = torch.device(args.device if torch.cuda.is_available() else "cpu")
print(device)
# 用来保存coco_info的文件
results_file = "results{}.txt".format(
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
data_transform = {
"train":
transforms.Compose(
[transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
"val":
transforms.Compose([transforms.ToTensor()])
}
VOC_root = args.data_path
# check voc root
if os.path.exists(os.path.join(VOC_root, "VOCdevkit")) is False:
raise FileNotFoundError(
"VOCdevkit dose not in path:'{}'.".format(VOC_root))
batch_size = args.batch_size
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
0]) # number of workers
print('Using %g dataloader workers' % nw)
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> train.txt
train_data_set = VOC2012DataSet(VOC_root, data_transform["train"],
"train.txt")
train_data_loader = torch.utils.data.DataLoader(
train_data_set,
batch_size=batch_size,
shuffle=True,
pin_memory=True,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
val_data_set = VOC2012DataSet(VOC_root, data_transform["val"], "val.txt")
val_data_set_loader = torch.utils.data.DataLoader(
val_data_set,
batch_size=batch_size,
shuffle=False,
pin_memory=True,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
model = get_model(num_classes=args.num_classes + 1)
print(model)
model.to(device)
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.005,
momentum=0.9,
weight_decay=0.0005)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=3,
gamma=0.33)
# 加载上次保存的权重
# 如果指定了上次训练保存的权重文件地址，则接着上次结果接着训练
if args.resume != "":
checkpoint = torch.load(args.resume, map_location=device)
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
print("the training process from epoch{}...".format(args.start_epoch))
train_loss = []
learning_rate = []
val_map = []
for epoch in range(args.start_epoch, args.epochs):
# train for one epoch, printing every 10 iterations
metric_logger = train_eval_utils.train_one_epoch(model,
optimizer,
train_data_loader,
device=device,
epoch=epoch,
print_freq=50)
mean_loss, lr = metric_logger["mloss"], metric_logger["lr"]
train_loss.append(mean_loss)
learning_rate.append(lr)
# update the learning rate
lr_scheduler.step()
# evaluate on the test dataset
coco_info = train_eval_utils.evaluate(model,
val_data_set_loader,
device=device)
# write into txt
with open(results_file, "a") as f:
# 写入的数据包括coco指标还有loss和learning rate
result_info = [
str(round(i, 4)) for i in coco_info + [mean_loss, lr]
]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
val_map.append(coco_info[1]) # pascal mAP
# save weights
# save_files = {
# 'model': model.state_dict(),
# 'optimizer': optimizer.state_dict(),
# 'lr_scheduler': lr_scheduler.state_dict(),
# 'epoch': epoch}
# torch.save(save_files, "./weights/resNetFpn-model-{}.pth".format(epoch))
if args.output_dir:
utils.save_on_master(
{
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'args': args,
'epoch': epoch
},
os.path.join(args.output_dir,
'resnet-fpn-model-{}.pth'.format(epoch)))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
def main():
GPU_count = torch.cuda.device_count()
# 默认使用最后一块 GPU，最大可能保证其空闲
device = torch.device(
"cuda:{}".format(GPU_count -
1) if torch.cuda.is_available() else "cpu")
print("Using cuda:{0}/{1} device training.".format(
GPU_count - 1, [item for item in range(GPU_count)]))
# 检查保存权重文件夹是否存在，不存在则创建
if not os.path.exists("save_weights"):
os.makedirs("save_weights")
data_transform = {
"train":
transforms.Compose(
[transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
"val":
transforms.Compose([transforms.ToTensor()])
}
VOC_root = "/home/shaolun/PYTHON/object-detection/faster-rcnn.pytorch/data/"
# check voc root
if os.path.exists(os.path.join(VOC_root, "VOCdevkit")) is False:
raise FileNotFoundError(
"VOCdevkit dose not in path:'{}'.".format(VOC_root))
# load train data set
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> train.txt
train_data_set = VOC2007DataSet(VOC_root, data_transform["train"],
"train.txt")
# 注意这里的collate_fn是自定义的，因为读取的数据包括image和targets，不能直接使用默认的方法合成batch
batch_size = 8
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
print('Using %g dataloader workers' % nw)
train_data_loader = torch.utils.data.DataLoader(
train_data_set,
batch_size=batch_size,
shuffle=True,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
# load validation data set
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> val.txt
val_data_set = VOC2007DataSet(VOC_root, data_transform["val"], "val.txt")
val_data_set_loader = torch.utils.data.DataLoader(
val_data_set,
batch_size=batch_size,
shuffle=False,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
# create model num_classes equal background + 20 classes
model = create_model(num_classes=21)
# print(model)
model.to(device)
train_loss = []
learning_rate = []
val_mAP = []
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #
# first frozen backbone and train 5 epochs #
# 首先冻结前置特征提取网络权重（backbone），训练rpn以及最终预测网络部分 #
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #
for param in model.backbone.parameters():
param.requires_grad = False
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.005,
momentum=0.9,
weight_decay=0.0005)
num_epochs = 5
# for epoch in range(num_epochs):
# # train for one epoch, printing every 10 iterations
# utils.train_one_epoch(model, optimizer, train_data_loader,
# device, epoch, print_freq=50,
# train_loss=train_loss, train_lr=learning_rate)
#
# # evaluate on the test dataset
# utils.evaluate(model, val_data_set_loader, device=device, mAP_list=val_mAP)
#
# torch.save(model.state_dict(), "./save_weights/pretrain.pth")
# # # # # # # # # # # # # # # # # # # # # # # # # # # #
# second unfrozen backbone and train all network #
# 解冻前置特征提取网络权重（backbone），接着训练整个网络权重 #
# # # # # # # # # # # # # # # # # # # # # # # # # # # #
# 冻结backbone部分底层权重
for name, parameter in model.backbone.named_parameters():
split_name = name.split(".")[0]
if split_name in ["0", "1", "2", "3"]:
parameter.requires_grad = False
else:
parameter.requires_grad = True
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.005,
momentum=0.9,
weight_decay=0.0005)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=5,
gamma=0.33)
num_epochs = 20
for epoch in range(num_epochs):
# train for one epoch, printing every 50 iterations
utils.train_one_epoch(model,
optimizer,
train_data_loader,
device,
epoch,
print_freq=50,
train_loss=train_loss,
train_lr=learning_rate)
# update the learning rate
lr_scheduler.step()
# evaluate on the test dataset
utils.evaluate(model,
val_data_set_loader,
device=device,
mAP_list=val_mAP)
# save weights
if epoch > 10:
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch
}
torch.save(save_files,
"./save_weights/mobile-model-{}.pth".format(epoch))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_mAP) != 0:
from plot_curve import plot_map
plot_map(val_mAP)
def main(parser_data):
device = torch.device(
parser_data.device if torch.cuda.is_available() else "cpu")
print("Using {} device training.".format(device.type))
results_file = "results{}.txt".format(
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
data_transform = {
"train":
transforms.Compose(
[transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
"val":
transforms.Compose([transforms.ToTensor()])
}
VOC_root = parser_data.data_path
# check voc root
if os.path.exists(os.path.join(VOC_root, "VOCdevkit")) is False:
raise FileNotFoundError(
"VOCdevkit dose not in path:'{}'.".format(VOC_root))
# load train data set
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> train.txt
train_dataset = VOCDataSet(VOC_root, "2012", data_transform["train"],
"train.txt")
train_sampler = None
# 是否按图片相似高宽比采样图片组成batch
# 使用的话能够减小训练时所需GPU显存，默认使用
if args.aspect_ratio_group_factor >= 0:
train_sampler = torch.utils.data.RandomSampler(train_dataset)
# 统计所有图像高宽比例在bins区间中的位置索引
group_ids = create_aspect_ratio_groups(
train_dataset, k=args.aspect_ratio_group_factor)
# 每个batch图片从同一高宽比例区间中取
train_batch_sampler = GroupedBatchSampler(train_sampler, group_ids,
args.batch_size)
# 注意这里的collate_fn是自定义的，因为读取的数据包括image和targets，不能直接使用默认的方法合成batch
batch_size = parser_data.batch_size
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
print('Using %g dataloader workers' % nw)
if train_sampler:
# 如果按照图片高宽比采样图片，dataloader中需要使用batch_sampler
train_data_loader = torch.utils.data.DataLoader(
train_dataset,
batch_sampler=train_batch_sampler,
pin_memory=True,
num_workers=nw,
collate_fn=train_dataset.collate_fn)
else:
train_data_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=batch_size,
shuffle=True,
pin_memory=True,
num_workers=nw,
collate_fn=train_dataset.collate_fn)
# load validation data set
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> val.txt
val_dataset = VOCDataSet(VOC_root, "2012", data_transform["val"],
"val.txt")
val_data_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=1,
shuffle=False,
pin_memory=True,
num_workers=nw,
collate_fn=val_dataset.collate_fn)
# create model
# 注意：不包含背景
model = create_model(num_classes=parser_data.num_classes)
# print(model)
model.to(device)
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.005,
momentum=0.9,
weight_decay=0.0005)
scaler = torch.cuda.amp.GradScaler() if args.amp else None
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=3,
gamma=0.33)
# 如果指定了上次训练保存的权重文件地址，则接着上次结果接着训练
if parser_data.resume != "":
checkpoint = torch.load(parser_data.resume, map_location='cpu')
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
parser_data.start_epoch = checkpoint['epoch'] + 1
if args.amp and "scaler" in checkpoint:
scaler.load_state_dict(checkpoint["scaler"])
print("the training process from epoch{}...".format(
parser_data.start_epoch))
train_loss = []
learning_rate = []
val_map = []
for epoch in range(parser_data.start_epoch, parser_data.epochs):
# train for one epoch, printing every 10 iterations
mean_loss, lr = utils.train_one_epoch(model,
optimizer,
train_data_loader,
device,
epoch,
print_freq=50,
warmup=True,
scaler=scaler)
train_loss.append(mean_loss.item())
learning_rate.append(lr)
# update the learning rate
lr_scheduler.step()
# evaluate on the test dataset
coco_info = utils.evaluate(model, val_data_loader, device=device)
# write into txt
with open(results_file, "a") as f:
# 写入的数据包括coco指标还有loss和learning rate
result_info = [
str(round(i, 4)) for i in coco_info + [mean_loss.item()]
] + [str(round(lr, 6))]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
val_map.append(coco_info[1]) # pascal map
# save weights
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch
}
if args.amp:
save_files["scaler"] = scaler.state_dict()
torch.save(save_files,
"./save_weights/resNetFpn-model-{}.pth".format(epoch))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
def main():
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(device)
if not os.path.exists("save_weights"):
os.mkdir("save_weights")
data_transform = {
"train": transform.Compose([transform.SSDCropping(),
transform.Resize(),
transform.ColorJitter(),
transform.ToTensor(),
transform.RandomHorizontalFlip(),
transform.Normalization(),
transform.AssignGTtoDefaultBox()]),
"val": transform.Compose([transform.Resize(),
transform.ToTensor(),
transform.Normalization()])
}
voc_path = "../"
train_dataset = VOC2012DataSet(voc_path, data_transform['train'], True)
# 注意训练时，batch_size必须大于1
train_data_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=8,
shuffle=True,
num_workers=0,
collate_fn=utils.collate_fn)
val_dataset = VOC2012DataSet(voc_path, data_transform['val'], False)
val_data_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=1,
shuffle=False,
num_workers=0,
collate_fn=utils.collate_fn)
model = create_model(num_classes=21, device=device)
model.to(device)
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params, lr=0.002,
momentum=0.9, weight_decay=0.0005)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=5,
gamma=0.3)
train_loss = []
learning_rate = []
val_map = []
val_data = None
# 如果电脑内存充裕，可提前加载验证集数据，以免每次验证时都要重新加载一次数据，节省时间
# val_data = get_coco_api_from_dataset(val_data_loader.dataset)
for epoch in range(20):
utils.train_one_epoch(model=model, optimizer=optimizer,
data_loader=train_data_loader,
device=device, epoch=epoch,
print_freq=50, train_loss=train_loss,
train_lr=learning_rate, warmup=True)
lr_scheduler.step()
utils.evaluate(model=model, data_loader=val_data_loader,
device=device, data_set=val_data, mAP_list=val_map)
# save weights
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch}
torch.save(save_files, "./save_weights/ssd300-{}.pth".format(epoch))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
def main():
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# 如果有可用的GPU就默认采用第0块设备，如果没有就用CPU
print("Using {} device training.".format(device.type))
# 用来保存coco_info的文件
results_file = "results{}.txt".format(datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
# 检查保存权重文件夹是否存在，不存在则创建
if not os.path.exists("save_weights"):
os.makedirs("save_weights")
data_transform = {
"train": transforms.Compose([transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
# 转化为tensor，然后随机水平翻转，GT也应该变化
"val": transforms.Compose([transforms.ToTensor()])
}
VOC_root = "./" # VOC数据集放在了根目录下
# VOC_root = os.getcwd() # VOC数据集放在了当前目录下
# check voc root
if os.path.exists(os.path.join(VOC_root, "VOCdevkit")) is False:
raise FileNotFoundError("VOCdevkit dose not in path:'{}'.".format(VOC_root))
# load train data set
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> train.txt
train_data_set = VOC2012DataSet(VOC_root, data_transform["train"], "train.txt")
# 使用VOC2012DataSet类来定义我们的数据集，然后再torch.DataLoader载入
# 注意这里的collate_fn是自定义的，因为读取的数据包括image和targets，不能直接使用默认的方法合成batch
batch_size = 8 # 根据GPU来设定
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8]) # number of workers，多线程图像读取和预处理，可以为4或8
print('Using %g dataloader workers' % nw)
train_data_loader = torch.utils.data.DataLoader(train_data_set,
batch_size=batch_size,
shuffle=True,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
# load validation data set
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> val.txt
val_data_set = VOC2012DataSet(VOC_root, data_transform["val"], "val.txt")
val_data_set_loader = torch.utils.data.DataLoader(val_data_set,
batch_size=batch_size,
shuffle=False,
pin_memory=True,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
# create model num_classes equal background + 20 classes
model = create_model(num_classes=21) # 20+1，加上了背景这个类别
# print(model)
model.to(device) # 将模型指派到设备中
train_loss = []
learning_rate = []
val_map = []
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #
# first frozen backbone and train 5 epochs #
# 首先冻结前置特征提取网络权重（backbone），训练rpn以及最终预测网络部分 #
# 因为现在只有backbone的预训练权值，用来微调rpn以及最终预测网络部分 #
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #
for param in model.backbone.parameters():
param.requires_grad = False
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params, lr=0.005,
momentum=0.9, weight_decay=0.0005)
init_epochs = 5 # 通过5个epochs进行微调
for epoch in range(init_epochs):
# train for one epoch, printing every 10 iterations
mean_loss, lr = utils.train_one_epoch(model, optimizer, train_data_loader,
device, epoch, print_freq=50, warmup=True)
train_loss.append(mean_loss.item())
learning_rate.append(lr)
# evaluate on the test dataset
coco_info = utils.evaluate(model, val_data_set_loader, device=device)
# write into txt
with open(results_file, "a") as f:
# 写入的数据包括coco指标还有loss和learning rate
result_info = [str(round(i, 4)) for i in coco_info + [mean_loss.item(), lr]]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
val_map.append(coco_info[1]) # pascal mAP
torch.save(model.state_dict(), "./save_weights/pretrain.pth") # 保存权值
# # # # # # # # # # # # # # # # # # # # # # # # # # # #
# second unfrozen backbone and train all network #
# 解冻前置特征提取网络权重（backbone），接着训练整个网络权重 #
# # # # # # # # # # # # # # # # # # # # # # # # # # # #
# 冻结backbone部分底层权重
for name, parameter in model.backbone.named_parameters():
split_name = name.split(".")[0]
if split_name in ["0", "1", "2", "3"]:
parameter.requires_grad = False
# 前面的几层都是通用的特征，所以可以冻结部分底层权重；这样不仅可以加快训练，并且效更好
# 这也是pytorch官方训练ResNet50和RPN的方法
else:
parameter.requires_grad = True
# define optimizer
params = [p for p in model.parameters() if p.requires_grad] # 遍历模型的所有权重，找出需要训练的即p.requires_grad=True
optimizer = torch.optim.SGD(params, lr=0.005,
momentum=0.9, weight_decay=0.0005) # 将参数传入SGD优化器中；初始lr，动量，decay
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, # 调整学习率的方法有很多，这里使用的是.StepLR
step_size=3,
gamma=0.33) # 设置学习率的路线，每隔step_size步，lr乘上系数gamma
num_epochs = 20 # 迭代了20个epoch
for epoch in range(init_epochs, num_epochs+init_epochs, 1):
# train for one epoch, printing every 50 iterations
mean_loss, lr = utils.train_one_epoch(model, optimizer, train_data_loader,
device, epoch, print_freq=50)
train_loss.append(mean_loss.item())
learning_rate.append(lr)
# update the learning rate
lr_scheduler.step() # 记录lr_scheduler方法已经执行一步了
# evaluate on the test dataset
coco_info = utils.evaluate(model, val_data_set_loader, device=device)
# write into txt，这些都可以没有
with open(results_file, "a") as f:
# 写入的数据包括coco指标还有loss和learning rate
result_info = [str(round(i, 4)) for i in coco_info + [mean_loss.item(), lr]]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
val_map.append(coco_info[1]) # pascal mAP
# save weights
# 仅保存最后5个epoch的权重
if epoch in range(num_epochs+init_epochs)[-5:]:
# if epoch >10： # 可以简单点，直接从第10个权重开始保存
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch}
torch.save(save_files, "./save_weights/mobile-model-{}.pth".format(epoch))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
def main(parser_data):
device = torch.device(
parser_data.device if torch.cuda.is_available() else "cpu")
print("Using {} device training.".format(device.type))
train_data_set = TrainDataset(lines[:num_train],
(input_shape[0], input_shape[1]),
mosaic=False)
# 注意这里的collate_fn是自定义的，因为读取的数据包括image和targets，不能直接使用默认的方法合成batch
batch_size = parser_data.batch_size
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
print('Using %g dataloader workers' % nw)
train_data_loader = torch.utils.data.DataLoader(
train_data_set,
batch_size=batch_size,
shuffle=True,
num_workers=nw,
drop_last=True,
collate_fn=train_dataset_collate)
# load validation data set
# val_data_set = VOC2012DataSet(VOC_root, data_transform["val"], False)
val_data_set = TrainDataset(lines[num_train:num_val + num_train],
(input_shape[0], input_shape[1]),
mosaic=False)
val_data_loader = torch.utils.data.DataLoader(
val_data_set,
batch_size=batch_size,
shuffle=False,
num_workers=nw,
drop_last=True,
collate_fn=train_dataset_collate)
# create model num_classes equal background + 1 classes
model = create_model(num_classes=2)
model.to(device)
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.005,
momentum=0.9,
weight_decay=0.0005)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=5,
gamma=0.33)
# 如果指定了上次训练保存的权重文件地址，则接着上次结果接着训练
if parser_data.resume != "":
checkpoint = torch.load(parser_data.resume)
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
parser_data.start_epoch = checkpoint['epoch'] + 1
print("the training process from epoch{}...".format(
parser_data.start_epoch))
train_loss = []
val_loss = []
learning_rate = []
val_mAP = []
for epoch in range(parser_data.start_epoch, parser_data.epochs):
# train for one epoch, printing every 10 iterations
pro_epoch_total_loss_train = []
pro_epoch_total_loss_val = []
utils.train_one_epoch(
model,
optimizer,
train_data_loader,
device,
epoch,
parser_data.epochs,
pro_epoch_total_loss_train=pro_epoch_total_loss_train,
train_loss=train_loss,
train_lr=learning_rate,
warmup=True)
# update the learning rate
lr_scheduler.step()
# evaluate on the test dataset
utils.val_one_epoch(model,
val_data_loader,
device,
epoch,
parser_data.epochs,
pro_epoch_total_loss_val=pro_epoch_total_loss_val,
val_loss=val_loss)
# save weights
# save_files = {
# 'model': model.state_dict(),
# 'optimizer': optimizer.state_dict(),
# 'lr_scheduler': lr_scheduler.state_dict(),
# 'epoch': epoch}
# lr = optimizer.param_groups[0]["lr"]
# torch.save(save_files, "./save_weights/Epoch_%03d_Loss_%.4f_lr_%.6f.pth"% ((epoch+1),
# np.mean(pro_epoch_total_loss_train),
# lr))
save_files = {
'model': model,
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch
}
lr = optimizer.param_groups[0]["lr"]
torch.save(
save_files, "./save_weights1/Epoch_%03d_Loss_%.4f_lr_%.6f.pth" %
((epoch + 1), np.mean(pro_epoch_total_loss_train), lr))
print('Loss_train: %.4f || Loss_val: %.4f ' %
(np.mean(pro_epoch_total_loss_train),
np.mean(pro_epoch_total_loss_val)))
print('Saving state, iter:', str(epoch + 1))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_mAP) != 0:
from plot_curve import plot_map
plot_map(val_mAP)
save_folder,
net,
num_classes,
dataset,
detect,
AP_stats=ap_stats)
else:
multi_scale_test_net(target_size,
save_folder,
net,
num_classes,
dataset,
detect,
AP_stats=ap_stats)
# print the best model.
max_idx = np.argmax(np.asarray(ap_stats['ap50']))
print('Best ap50: {:.4f} at epoch {}'.format(ap_stats['ap50'][max_idx],
ap_stats['epoch'][max_idx]))
res_file = os.path.join(save_folder, 'ap_stats.json')
print('Writing ap stats json to {}'.format(res_file))
with open(res_file, 'w') as fid:
json.dump(ap_stats, fid)
# plot curves
fig_name = 'ap.png'
metrics = ['ap50']
legend = ['ap50']
plot_map(save_folder, ap_stats, metrics, legend, fig_name)
txt_log = prefix + '/log.txt'
plot_loss(save_folder, txt_log)
def main(parser_data):
device = torch.device(
parser_data.device if torch.cuda.is_available() else "cpu")
print("Using {} device training.".format(device.type))
# 用来保存coco_info的文件
results_file = "results{}.txt".format(
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
data_transform = {
"train":
transforms.Compose(
[transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
"val":
transforms.Compose([transforms.ToTensor()])
}
VOC_root = parser_data.data_path
# check voc root
if os.path.exists(os.path.join(VOC_root, "VOCdevkit")) is False:
raise FileNotFoundError(
"VOCdevkit dose not in path:'{}'.".format(VOC_root))
# load train data set
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> train.txt
train_data_set = VOC2012DataSet(VOC_root, data_transform["train"],
"train.txt")
# 注意这里的collate_fn是自定义的，因为读取的数据包括image和targets，不能直接使用默认的方法合成batch
batch_size = parser_data.batch_size
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
print('Using %g dataloader workers' % nw)
train_data_loader = torch.utils.data.DataLoader(
train_data_set,
batch_size=batch_size,
shuffle=True,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
# load validation data set
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> val.txt
val_data_set = VOC2012DataSet(VOC_root, data_transform["val"], "val.txt")
val_data_set_loader = torch.utils.data.DataLoader(
val_data_set,
batch_size=batch_size,
shuffle=False,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
# create model num_classes equal background + 20 classes
model = create_model(num_classes=21, device=device)
# print(model)
model.to(device)
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.005,
momentum=0.9,
weight_decay=0.0005)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=3,
gamma=0.33)
# 如果指定了上次训练保存的权重文件地址，则接着上次结果接着训练
if parser_data.resume != "":
checkpoint = torch.load(parser_data.resume, map_location=device)
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
parser_data.start_epoch = checkpoint['epoch'] + 1
print("the training process from epoch{}...".format(
parser_data.start_epoch))
train_loss = []
learning_rate = []
val_map = []
for epoch in range(parser_data.start_epoch, parser_data.epochs):
# train for one epoch, printing every 10 iterations
mean_loss, lr = utils.train_one_epoch(model,
optimizer,
train_data_loader,
device=device,
epoch=epoch,
print_freq=50,
warmup=True)
train_loss.append(mean_loss.item())
learning_rate.append(lr)
# update the learning rate
lr_scheduler.step()
# evaluate on the test dataset
coco_info = utils.evaluate(model, val_data_set_loader, device=device)
# write into txt
with open(results_file, "a") as f:
result_info = [
str(round(i, 4)) for i in coco_info + [mean_loss.item(), lr]
]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
val_map.append(coco_info[1]) # pascal mAP
# save weights
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch
}
torch.save(save_files,
"./save_weights/resNetFpn-model-{}.pth".format(epoch))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
def main(args):
device = torch.device(args.device if torch.cuda.is_available() else "cpu")
print("Using {} device training.".format(device.type))
# 用来保存coco_info的文件
results_file = "results{}.txt".format(
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
data_transform = {
"train":
transforms.Compose(
[transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
"val":
transforms.Compose([transforms.ToTensor()])
}
COCO_root = args.data_path
# load train data set
# coco2017 -> annotations -> instances_train2017.json
train_data_set = CocoDetection(COCO_root, "train", data_transform["train"])
# 注意这里的collate_fn是自定义的，因为读取的数据包括image和targets，不能直接使用默认的方法合成batch
batch_size = args.batch_size
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
print('Using %g dataloader workers' % nw)
train_data_loader = torch.utils.data.DataLoader(
train_data_set,
batch_size=batch_size,
shuffle=True,
pin_memory=True,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
# load validation data set
# coco2017 -> annotations -> instances_val2017.json
val_data_set = CocoDetection(COCO_root, "val", data_transform["val"])
val_data_set_loader = torch.utils.data.DataLoader(
val_data_set,
batch_size=batch_size,
shuffle=False,
pin_memory=True,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
# create model num_classes equal background + 80 classes
model = create_model(num_classes=args.num_classes + 1)
# print(model)
model.to(device)
train_loss = []
learning_rate = []
val_map = []
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
optimizer, milestones=args.lr_steps, gamma=args.lr_gamma)
# 如果传入resume参数，即上次训练的权重地址，则接着上次的参数训练
if args.resume:
# If map_location is missing, torch.load will first load the module to CPU
# and then copy each parameter to where it was saved,
# which would result in all processes on the same machine using the same set of devices.
checkpoint = torch.load(
args.resume, map_location=device) # 读取之前保存的权重文件(包括优化器以及学习率策略)
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
for epoch in range(args.start_epoch, args.epochs):
# train for one epoch, printing every 50 iterations
mean_loss, lr = utils.train_one_epoch(model,
optimizer,
train_data_loader,
device,
epoch,
print_freq=50,
warmup=True)
train_loss.append(mean_loss.item())
learning_rate.append(lr)
# update the learning rate
lr_scheduler.step()
# evaluate on the test dataset
coco_info = utils.evaluate(model, val_data_set_loader, device=device)
# write into txt
with open(results_file, "a") as f:
# 写入的数据包括coco指标还有loss和learning rate
result_info = [
str(round(i, 4)) for i in coco_info + [mean_loss.item(), lr]
]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
val_map.append(coco_info[1]) # pascal mAP
# save weights
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch
}
torch.save(save_files,
"./save_weights/mobile-model-{}.pth".format(epoch))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
def main(parser_data):
device = torch.device(parser_data.device if torch.cuda.is_available() else "cpu")
print(device)
data_transform = {
"train": transforms.Compose([transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
"val": transforms.Compose([transforms.ToTensor()])
}
VOC_root = parser_data.data_path
assert os.path.exists(os.path.join(VOC_root, "VOCdevkit")), "not found VOCdevkit in path:'{}'".format(VOC_root)
# load train data set
train_data_set = VOC2012DataSet(VOC_root, data_transform["train"], True)
# 注意这里的collate_fn是自定义的，因为读取的数据包括image和targets，不能直接使用默认的方法合成batch
train_data_loader = torch.utils.data.DataLoader(train_data_set,
batch_size=1,
shuffle=False,
num_workers=0,
collate_fn=utils.collate_fn)
# load validation data set
val_data_set = VOC2012DataSet(VOC_root, data_transform["val"], False)
val_data_set_loader = torch.utils.data.DataLoader(val_data_set,
batch_size=2,
shuffle=False,
num_workers=0,
collate_fn=utils.collate_fn)
# create model num_classes equal background + 20 classes
model = create_model(num_classes=21)
# print(model)
model.to(device)
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params, lr=0.005,
momentum=0.9, weight_decay=0.0005)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=5,
gamma=0.33)
# 如果指定了上次训练保存的权重文件地址，则接着上次结果接着训练
if parser_data.resume != "":
checkpoint = torch.load(parser_data.resume)
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
parser_data.start_epoch = checkpoint['epoch'] + 1
print("the training process from epoch{}...".format(parser_data.start_epoch))
train_loss = []
learning_rate = []
val_mAP = []
for epoch in range(parser_data.start_epoch, parser_data.epochs):
# train for one epoch, printing every 10 iterations
utils.train_one_epoch(model, optimizer, train_data_loader,
device, epoch, train_loss=train_loss, train_lr=learning_rate,
print_freq=50, warmup=True)
# update the learning rate
lr_scheduler.step()
# evaluate on the test dataset
utils.evaluate(model, val_data_set_loader, device=device, mAP_list=val_mAP)
# save weights
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch}
torch.save(save_files, "./save_weights/resNetFpn-model-{}.pth".format(epoch))
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_mAP) != 0:
from plot_curve import plot_map
plot_map(val_mAP)
(epoch + 1, running_loss / step, val_accurate))
plt.plot(range(len(lr_list)), lr_list, color='r')
val_map.append(val_accurate)
df = pd.DataFrame([train_loss,lr_list,val_map],index=clm)
df.to_csv('Result.csv')
# plot loss and lr curve
if len(train_loss) != 0 and len(lr_list) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, lr_list)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
scheduler.step()
print('Finished Training')
# net.eval()
# acc = 0.0
# with torch.no_grad():
# for val_data in tqdm(validate_loader):
# val_images, val_labels = val_data
# outputs = net(val_images.to(device)) # eval model only have last output layer
# # loss = loss_function(outputs, test_labels)
# predict_y = torch.max(outputs, dim=1)[1]
# acc += (predict_y == val_labels.to(device)).sum().item()
# val_accurate = acc / val_num
# print('test_accuracy: %.3f' %(val_accurate))