def train(train_img_path, pths_path, batch_size, lr, decay, num_workers,
epoch_iter, interval, pretained):
file_num = len(os.listdir(train_img_path))
trainset = custom_dataset(train_img_path)
train_loader = data.DataLoader(trainset, batch_size=batch_size, \
shuffle=True, num_workers=num_workers, drop_last=True)
criterion = Loss()
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = EAST()
# TODO 可能是bug
if os.path.exists(pretained):
model.load_state_dict(torch.load(pretained))
data_parallel = False
if torch.cuda.device_count() > 1:
model = nn.DataParallel(model)
data_parallel = True
model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=lr, weight_decay=decay)
# scheduler = lr_scheduler.StepLR(optimizer, step_size=10000, gamma=0.94)
for epoch in range(epoch_iter):
model.train()
optimizer.step()
epoch_loss = 0
epoch_time = time.time()
for i, (img, gt_map) in enumerate(train_loader):
start_time = time.time()
img, gt_map = img.to(device), gt_map.to(device)
east_detect = model(img)
inside_score_loss, side_vertex_code_loss, side_vertex_coord_loss = criterion(
gt_map, east_detect)
loss = inside_score_loss + side_vertex_code_loss + side_vertex_coord_loss
epoch_loss += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
if i % 10 == 0:
print('Epoch is [{}/{}], mini-batch is [{}/{}], time consumption is {:.8f}, batch_loss is {:.8f}'.format( \
epoch + 1, epoch_iter, i + 1, int(file_num / batch_size), time.time() - start_time, loss.item()))
print(
"inside_score_loss: %f | side_vertex_code_loss: %f | side_vertex_coord_loss: %f"
% (inside_score_loss, side_vertex_code_loss,
side_vertex_coord_loss))
print('epoch_loss is {:.8f}, epoch_time is {:.8f}'.format(
epoch_loss / int(file_num / batch_size),
time.time() - epoch_time))
print(time.asctime(time.localtime(time.time())))
# print('=' * 50)
if (epoch + 1) % interval == 0:
state_dict = model.module.state_dict(
) if data_parallel else model.state_dict()
torch.save(
state_dict,
os.path.join(
pths_path, cfg.train_task_id +
'_model_epoch_{}.pth'.format(epoch + 1)))
def train(train_img_path, train_gt_path, pths_path, batch_size, lr, num_workers, epoch_iter, interval, output_dir):
# 为CPU设置种子用于生成随机数，以使得结果是确定的
torch.manual_seed(970201) # 为CPU设置随机种子
torch.cuda.manual_seed(970201) # 为当前GPU设置随机种子
logger = setup_logger("east_matrix", output_dir, get_rank())
file_num = len(os.listdir(train_img_path)) # 图片数量
trainset = custom_dataset(train_img_path, train_gt_path) # 训练集进行处理 ？？？ ***
# 加载数据，组合一个数据集和一个采样器，并在给定的数据集上提供一个可迭代的。
train_loader = data.DataLoader(trainset, batch_size=batch_size, \
shuffle=True, num_workers=num_workers, drop_last=True)
criterion = Loss() # 损失函数 ？？？ ***
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = EAST() # 网络模型 ？？？ ***
# 是否多gpu
data_parallel = False
if torch.cuda.device_count() > 1:
model = nn.DataParallel(model)
data_parallel = True
# 分配模型到gpu或cpu，根据device决定
model.to(device)
#优化器
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
# 学习率衰减策略，一半的时候衰减为十分之一
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[epoch_iter//2], gamma=0.1)
def train(train_root_path, pths_path, batch_size, lr, num_workers, epoch_iter, interval):
trainset = custom_dataset(train_root_path)
file_num = trainset.__len__()
train_loader = data.DataLoader(trainset, batch_size=batch_size, shuffle=True, num_workers=num_workers, drop_last=True)
criterion = Loss()
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = EAST(pretrained=False)
model.load_state_dict(torch.load('/home/chen-ubuntu/Desktop/checks_dataset/pths/model_epoch_stamp_8.pth'))
model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
optimizer.zero_grad()
# scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[epoch_iter//2], gamma=0.1)
for epoch in range(epoch_iter):
model.train()
epoch_loss = 0
loss_plot = []
bx = []
'''
for i, (img, gt_score, gt_geo, ignored_map) in enumerate(train_loader):
start_time = time.time()
img, gt_score, gt_geo, ignored_map = img.to(device), gt_score.to(device), gt_geo.to(device), ignored_map.to(
device)
pred_score, pred_geo = model(img)
loss = criterion(gt_score, pred_score, gt_geo, pred_geo, ignored_map)
epoch_loss += loss.item()
loss.backward()
if (i + 1) % 8 == 0:
optimizer.step()
optimizer.zero_grad()
if (i + 1) % 100 == 0:
print(
'Epoch is [{}/{}], mini-batch is [{}/{}], time consumption is {:.8f}, batch_loss is {:.8f}'.format(
epoch + 1, epoch_iter, i + 1, int(file_num / batch_size), time.time() - start_time,
loss.item()))
if (i + 1) % 30 == 0:
loss_plot.append(loss.item())
bx.append(i + epoch * int(file_num / batch_size))
plt.plot(bx, loss_plot, label='loss_mean', linewidth=1, color='b', marker='o',
markerfacecolor='green', markersize=2)
plt.savefig(os.path.abspath('./labeled2.jpg'))
print('epoch_loss is {:.8f}, epoch_time is {:.8f}'.format(epoch_loss / int(file_num / batch_size),
time.time() - epoch_time))
print(time.asctime(time.localtime(time.time())))
print('=' * 50)'''
if epoch % interval == 0:
validloss, validacc = valid(train_loader, model, criterion, device)
state_dict = model.module.state_dict() if data_parallel else model.state_dict()
def train(train_img_path, train_gt_path, pths_path, batch_size, lr,
num_workers, epoch_iter, interval):
file_num = len(os.listdir(train_img_path))
trainset = custom_dataset(train_img_path, train_gt_path)
train_loader = data.DataLoader(trainset, batch_size=batch_size, \
shuffle=True, num_workers=num_workers, drop_last=True)
criterion = Loss()
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = EAST()
data_parallel = False
if torch.cuda.device_count() > 1:
model = nn.DataParallel(model)
data_parallel = True
model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=[epoch_iter // 2],
gamma=0.1)
for epoch in range(epoch_iter):
model.train()
scheduler.step()
epoch_loss = 0
epoch_time = time.time()
for i, (img, gt_score, gt_geo, ignored_map) in enumerate(train_loader):
start_time = time.time()
img, gt_score, gt_geo, ignored_map = img.to(device), gt_score.to(
device), gt_geo.to(device), ignored_map.to(device)
pred_score, pred_geo = model(img)
loss = criterion(gt_score, pred_score, gt_geo, pred_geo,
ignored_map)
epoch_loss += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
print('Epoch is [{}/{}], mini-batch is [{}/{}], time consumption is {:.8f}, batch_loss is {:.8f}'.format(\
epoch+1, epoch_iter, i+1, int(file_num/batch_size), time.time()-start_time, loss.item()))
print('epoch_loss is {:.8f}, epoch_time is {:.8f}'.format(
epoch_loss / int(file_num / batch_size),
time.time() - epoch_time))
print(time.asctime(time.localtime(time.time())))
print('=' * 50)
if (epoch + 1) % interval == 0:
state_dict = model.module.state_dict(
) if data_parallel else model.state_dict()
torch.save(
state_dict,
os.path.join(pths_path,
'model_epoch_{}.pth'.format(epoch + 1)))
def main():
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = EAST()
model = model.eval()
model = model.to(device)
model.load_state_dict(torch.load(args.trained_model))
if os.path.exists(args.save_folder):
shutil.rmtree(args.save_folder)
os.mkdir(args.save_folder)
test_process = tqdm(os.listdir(args.img_path), ascii=True)
for img_file in test_process:
test_process.set_description("Processing")
img = Image.open(os.path.join(args.img_path, img_file))
boxes = detect(img, model, device)
#绘制boxes到图片上
plot_img = plot_boxes(img, boxes)
plot_img.save(os.path.join(args.save_folder, img_file))
if args.show_image:
plot_img.show()
def train(train_ds_path,
val_ds_path,
pths_path,
results_path,
batch_size,
lr,
num_workers,
train_iter,
interval,
opt_level=0,
checkpoint_path=None,
val_freq=10):
torch.cuda.set_device(rank)
tensorboard_dir = os.path.join(results_path, 'logs')
checkpoints_dir = os.path.join(results_path, 'checkpoints')
if rank == 0:
os.makedirs(tensorboard_dir, exist_ok=True)
os.makedirs(checkpoints_dir, exist_ok=True)
barrier()
try:
logger.info('Importing AutoResume lib...')
from userlib.auto_resume import AutoResume as auto_resume
auto_resume.init()
logger.info('Success!')
except:
logger.info('Failed!')
auto_resume = None
trainset = custom_dataset(
os.path.join(train_ds_path, 'images'),
os.path.join(train_ds_path, 'gt'),
)
valset = custom_dataset(os.path.join(val_ds_path, 'images'),
os.path.join(val_ds_path, 'gt'),
is_val=True)
logger.info(f'World Size: {world_size}, Rank: {rank}')
if world_size > 1:
train_sampler = torch.utils.data.distributed.DistributedSampler(
trainset)
val_sampler = torch.utils.data.distributed.DistributedSampler(
valset, shuffle=False)
else:
train_sampler = None
val_sampler = None
worker_init = LoaderWorkerProcessInit(rank, 43)
train_loader = DataLoader(trainset,
batch_size=batch_size,
shuffle=train_sampler is None,
sampler=train_sampler,
num_workers=num_workers,
pin_memory=True,
drop_last=True,
worker_init_fn=worker_init)
val_loader = DataLoader(valset,
batch_size=batch_size,
shuffle=False,
sampler=val_sampler,
num_workers=num_workers,
pin_memory=True,
drop_last=True,
worker_init_fn=worker_init)
criterion = Loss()
device = torch.device(
f"cuda:{rank}" if torch.cuda.is_available() else "cpu")
model = EAST()
model.to(device)
model = apex.parallel.convert_syncbn_model(model)
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=f'O{opt_level}')
start_iter = 0
if auto_resume is not None:
auto_resume_details = auto_resume.get_resume_details()
if auto_resume_details is not None:
logger.info(
'Detected that this is a resumption of a previous job!')
checkpoint_path = auto_resume_details['CHECKPOINT_PATH']
if checkpoint_path:
logger.info(f'Loading checkpoint at path "{checkpoint_path}"...')
checkpoint = torch.load(checkpoint_path, map_location=f'cuda:{rank}')
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
amp.load_state_dict(checkpoint['amp_state'])
start_iter = checkpoint['iter']
logger.info('Done')
data_parallel = False
main_model = model
if torch.distributed.is_initialized():
logger.info(
f'DataParallel: Using {torch.cuda.device_count()} devices!')
model = DDP(model)
data_parallel = True
for param_group in optimizer.param_groups:
param_group.setdefault('initial_lr', lr)
scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=[train_iter // 2],
gamma=0.1,
last_epoch=start_iter)
# This allows us to change dataset size without affecting things such as validation frequency
steps_per_epoch = 1000 // (world_size * batch_size)
step = start_iter
start_epoch = step // steps_per_epoch
epoch_iter = int(math.ceil(train_iter / steps_per_epoch))
if rank == 0:
logger.info('Initializing Tensorboard')
writer = SummaryWriter(tensorboard_dir, purge_step=step)
loss_meters = MeterDict(reset_on_value=True)
val_loss_meters = MeterDict(reset_on_value=True)
time_meters = MeterDict(reset_on_value=True)
logger.info('Training')
model.train()
train_start_time = time.time()
best_loss = 100
train_iter = [iter(train_loader)]
def get_batch():
try:
return next(train_iter[0])
except:
train_iter[0] = iter(train_loader)
return get_batch()
for epoch in range(start_epoch, epoch_iter):
if train_sampler is not None:
train_sampler.set_epoch(epoch)
epoch_loss = 0
epoch_time = time.time()
start_time = time.time()
model.train()
for i in range(steps_per_epoch):
batch = get_batch()
optimizer.zero_grad()
batch = [b.cuda(rank, non_blocking=True) for b in batch]
img, gt_score, gt_geo, ignored_map = batch
barrier()
time_meters['batch_time'].add_sample(time.time() - start_time)
pred_score, pred_geo = model(img)
loss, details = criterion(gt_score, pred_score, gt_geo, pred_geo,
ignored_map)
epoch_loss += loss.detach().item()
with amp.scale_loss(loss, optimizer) as loss_scaled:
loss_scaled.backward()
optimizer.step()
barrier()
time_meters['step_time'].add_sample(time.time() - start_time)
details['global'] = loss.detach().item()
for k, v in details.items():
loss_meters[k].add_sample(v)
if i % 10 == 0:
logger.info(f'\tStep [{i+1}/{steps_per_epoch}]')
start_time = time.time()
step += 1
scheduler.step()
if step == train_iter:
break
term_requested = auto_resume is not None and auto_resume.termination_requested(
)
checkpoint_path = None
if rank == 0:
times = {k: m.value() for k, m in time_meters.items()}
losses = {k: m.value() for k, m in loss_meters.items()}
times['epoch'] = time.time() - epoch_time
logger.info(
f'Epoch is [{epoch+1}/{epoch_iter}], time consumption is {times}, batch_loss is {losses}'
)
for k, v in times.items():
writer.add_scalar(f'performance/{k}', v, step)
for k, v in losses.items():
writer.add_scalar(f'loss/{k}', v, step)
writer.add_scalar('learning_rate', optimizer.param_groups[0]['lr'],
step)
if term_requested or (epoch + 1) % interval == 0:
state_dict = main_model.state_dict()
optim_state = optimizer.state_dict()
checkpoint_path = os.path.join(
checkpoints_dir, 'model_epoch_{}.pth'.format(epoch + 1))
logger.info(f'Saving checkpoint to "{checkpoint_path}"...')
torch.save(
{
'model': state_dict,
'optimizer': optim_state,
'amp_state': amp.state_dict(),
'epoch': epoch + 1,
'iter': step
}, checkpoint_path)
logger.info(f'Done')
if (epoch + 1) % val_freq == 0 or step == train_iter:
logger.info(f'Validating epoch {epoch+1}...')
model.eval()
val_loader.dataset.reset_random()
with torch.no_grad():
for i, batch in enumerate(val_loader):
batch = [b.cuda(rank, non_blocking=True) for b in batch]
img, gt_score, gt_geo, ignored_map = batch
barrier()
pred_score, pred_geo = model(img)
loss, details = criterion(gt_score, pred_score, gt_geo,
pred_geo, ignored_map)
details['global'] = loss.detach().item()
barrier()
for k, v in details.items():
val_loss_meters[k].add_sample(v)
print_dict = dict()
for k, m in val_loss_meters.items():
t = torch.tensor(m.value(),
device=f'cuda:{rank}',
dtype=torch.float32)
if world_size > 1:
torch.distributed.reduce(t, 0)
t /= world_size
if rank == 0:
writer.add_scalar(f'val/loss/{k}', t.item(), step)
print_dict[k] = t.item()
logger.info(f'\tLoss: {print_dict}')
val_loss = print_dict['global']
if rank == 0 and val_loss < best_loss:
logger.info(
f'This is the best model so far. New loss: {val_loss}, previous: {best_loss}'
)
best_loss = val_loss
shutil.copyfile(checkpoint_path,
os.path.join(checkpoints_dir, 'best.pth'))
logger.info('Training')
if term_requested:
logger.warning('Termination requested! Exiting...')
if rank == 0:
auto_resume.request_resume(user_dict={
'CHECKPOINT_PATH': save_path,
'EPOCH': epoch
})
break
logger.info(
f'Finished training!!! Took {time.time()-train_start_time:0.3f} seconds!'
)
def train(img_path, gt_path, pths_path, batch_size, lr, num_workers,
epoch_iter, interval):
img_files = [
os.path.join(img_path, img_file)
for img_file in sorted(os.listdir(img_path))
]
gt_files = [
os.path.join(gt_path, gt_file)
for gt_file in sorted(os.listdir(gt_path))
]
if len(img_files) != len(gt_files):
print('dataset is wrong!')
return
np.random.seed(10)
state = np.random.get_state()
np.random.shuffle(img_files)
np.random.set_state(state)
np.random.shuffle(gt_files)
segment = len(img_files) // 10
train_img_files = img_files[:segment * 1]
train_gt_files = gt_files[:segment * 1]
val_img_files = img_files[segment * 1:]
val_gt_files = gt_files[segment * 1:]
print('trainset: ', len(train_img_files))
print('validset: ', len(val_img_files))
trainset = custom_dataset(train_img_files, train_gt_files, transform=True)
validset = custom_dataset(val_img_files, val_gt_files)
train_loader = data.DataLoader(trainset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers,
drop_last=True)
valid_loader = data.DataLoader(validset,
batch_size=1,
shuffle=True,
num_workers=num_workers,
drop_last=True)
train_num = len(train_img_files)
model = EAST(pretrained=False)
model.load_state_dict(
torch.load(
'/home/chen-ubuntu/Desktop/checks_dataset/pths/model_mode1_epoch_24.pth'
))
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)
criterion = Loss()
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
optimizer.zero_grad()
batch_cnt = 0
for epoch in range(epoch_iter):
model.train()
epoch_loss = 0
'''
for i, (img, gt_score, gt_geo, ignored_map, _) in enumerate(train_loader):
batch_cnt += 1
start_time = time.time()
img, gt_score, gt_geo, ignored_map = img.to(device), gt_score.to(device), gt_geo.to(device), ignored_map.to(
device)
pred_score, pred_geo = model(img)
loss = criterion(gt_score, pred_score, gt_geo, pred_geo, ignored_map)
epoch_loss += loss.item()
loss.backward()
if (i + 1) % 8 == 0:
optimizer.step()
optimizer.zero_grad()
if (i + 1) % 8 == 0:
print(
'Epoch is [{}/{}], mini-batch is [{}/{}], time consumption is {:.8f}, batch_loss is {:.8f}'.format(
epoch + 1, epoch_iter, i + 1, int(train_num / batch_size), time.time() - start_time,
loss.item()))
writer.add_scalar('data/train_loss', loss.item(), batch_cnt)
'''
if epoch % interval == 0:
validloss, validacc = valid(valid_loader, model, criterion, device)
#writer.add_scalar('data/valid_loss', validloss, batch_cnt)
#writer.add_scalar('data/valid_acc', validacc, batch_cnt)
#state_dict = model.state_dict()
#torch.save(state_dict, os.path.join(pths_path, 'model_epoch_{}_acc_{:.3f}.pth'.format(epoch + 1, validacc)))
print('=' * 50)
def train(config):
tb_writer = SummaryWriter(config.out)
train_dataset = ICDARDataSet(config.train_data_path)
file_num = train_dataset.get_num_of_data()
train_loader = data.DataLoader(train_dataset,
batch_size=config.train_batch_size,
shuffle=True,
num_workers=config.num_workers,
drop_last=True)
criterion = Loss()
model = EAST()
data_parallel = False
if torch.cuda.device_count() > 1:
model = nn.DataParallel(model)
data_parallel = True
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=config.lr)
# scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[config.epoch // 2, config.epoch//2 +
# config.epoch//4, config.epoch//2], gamma=0.1)
scheduler = lr_scheduler.ReduceLROnPlateau(optimizer,
mode='min',
factor=0.1,
patience=3,
verbose=True,
min_lr=1e-5)
best_hmean = 0.0
for epoch in range(config.epoch):
model.train()
epoch_loss = 0
epoch_time = time.time()
for i, (img, gt_score, gt_geo,
ignored_map) in tqdm(enumerate(train_loader),
desc='Training...'):
img = img.to(device)
gt_score, gt_geo, ignored_map = gt_score.to(device), gt_geo.to(
device), ignored_map.to(device)
pred_score, pred_geo = model(img)
total_loss, classify_loss, angle_loss, iou_loss, geo_loss = criterion(
gt_score, pred_score, gt_geo, pred_geo, ignored_map)
tb_writer.add_scalar('train/loss', total_loss,
epoch * len(train_dataset) + i)
tb_writer.add_scalar('train/classify_loss', classify_loss,
epoch * len(train_dataset) + i)
tb_writer.add_scalar('train/angle_loss', angle_loss,
epoch * len(train_dataset) + i)
tb_writer.add_scalar('train/iou_loss', iou_loss,
epoch * len(train_dataset) + i)
tb_writer.add_scalar('train/geo_loss', geo_loss,
epoch * len(train_dataset) + i)
epoch_loss += total_loss.item()
optimizer.zero_grad()
total_loss.backward()
optimizer.step()
epoch_loss = epoch_loss / int(file_num / config.train_batch_size)
print('\n {} epoch_loss is {:.8f}, epoch_time is {:.8f}'.format(
epoch, epoch_loss,
time.time() - epoch_time))
print(time.asctime(time.localtime(time.time())))
print('=' * 50)
scheduler.step(epoch_loss)
tb_writer.add_scalar('lr', get_lr(optimizer),
(epoch + 1) * len(train_dataset))
_, eval_result = evaluate_batch(model, config)
print(eval_result)
tb_writer.add_scalar('train/hmean', eval_result['hmean'],
(epoch + 1) * len(train_dataset))
tb_writer.add_scalar('train/precision', eval_result['precision'],
(epoch + 1) * len(train_dataset))
tb_writer.add_scalar('train/recall', eval_result['recall'],
(epoch + 1) * len(train_dataset))
if eval_result['hmean'] > best_hmean:
best_hmean = eval_result['hmean']
state_dict = model.module.state_dict(
) if data_parallel else model.state_dict()
torch.save(
state_dict,
os.path.join(config.out,
'model_epoch_{}.pth'.format(epoch + 1)))
def train(train_img_path, train_gt_path, pths_path, batch_size, lr,
num_workers, epoch_iter, interval):
#数据处理
#import pdb
#pdb.set_trace()
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
file_num = len(os.listdir(train_img_path))
trainset = custom_dataset(train_img_path, train_gt_path)
train_loader = data.DataLoader(trainset, batch_size=batch_size, \
shuffle=True, num_workers=num_workers, drop_last=True)
#模型实现
model = EAST()
data_parallel = False
if torch.cuda.device_count() > 1:
model = nn.DataParallel(model)
data_parallel = True
model.to(device)
#loss实现
criterion = Loss()
#[完善优化算法的调用]写出优化算法的
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
#定义学习策略
scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=[epoch_iter // 2],
gamma=0.1)
for epoch in range(epoch_iter):
model.train()
epoch_loss = 0
epoch_time = time.time()
# import pdb
# pdb.set_trace()
train_process = tqdm(train_loader)
for i, (img, gt_score, gt_geo,
ignored_map) in enumerate(train_process):
start_time = time.time()
#import pdb
# pdb.set_trace()
# print("start_time=%s"%(start_time))
img, gt_score, gt_geo, ignored_map = img.to(device), gt_score.to(
device), gt_geo.to(device), ignored_map.to(device)
# 使用模型
pred_score, pred_geo = model(img)
# 计算得到loss
loss = criterion(gt_score, pred_score, gt_geo, pred_geo,
ignored_map)
epoch_loss += loss.item()
# 利用loss求取梯度
optimizer.zero_grad()
loss.backward()
#权重更新
optimizer.step()
train_process.set_description_str("epoch:{}".format(epoch + 1))
train_process.set_postfix_str("batch_loss:{:.4f}".format(
loss.item()))
'''
print('Epoch is [{}/{}], mini-batch is [{}/{}], time consumption is {:.8f}, batch_loss is {:.8f}'.format(\
epoch+1, epoch_iter, i+1, int(file_num/batch_size), time.time()-start_time, loss.item()))
'''
scheduler.step()
with open('train.csv', 'a') as f:
f.write('epoch[{}]: epoch_loss is {:.8f}, epoch_time is {:.8f}\n'.
format(epoch + 1, epoch_loss / int(file_num / batch_size),
time.time() - epoch_time))
# print('epoch_loss is {:.8f}, epoch_time is {:.8f}'.format(epoch_loss/int(file_num/batch_size), time.time()-epoch_time))
# print(time.asctime(time.localtime(time.time())))
# print('='*50)
if (epoch + 1) % interval == 0:
state_dict = model.module.state_dict(
) if data_parallel else model.state_dict()
torch.save(
state_dict,
os.path.join(pths_path,
'model_epoch_{}.pth'.format(epoch + 1)))
def train(train_img_path, train_gt_path, pths_path, batch_size, lr,
num_workers, epoch_iter, interval):
file_num = len(os.listdir(train_img_path))
trainset = custom_dataset(train_img_path, train_gt_path)
train_loader = data.DataLoader(trainset, batch_size=batch_size, \
shuffle=True, num_workers=num_workers, drop_last=True)
test_img_path = os.path.abspath('../ICDAR_2015/test_img')
test_gt_path = os.path.abspath('../ICDAR_2015/test_gt')
file_num2 = len(os.listdir(test_img_path))
testset = custom_dataset(test_img_path, test_gt_path)
test_loader = data.DataLoader(testset, batch_size=batch_size, \
shuffle=True, num_workers=num_workers, drop_last=True)
criterion = Loss()
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = EAST()
data_parallel = False
if torch.cuda.device_count() > 1:
model = nn.DataParallel(model)
data_parallel = True
model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
try:
print("(Continue) Loading east...")
checkpoint = torch.load('./pths/east.pth')
model.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
epoch_dict = checkpoint['epoch_loss']
test_dict = checkpoint['test_loss']
total_epoch = checkpoint['epoch']
best_loss = checkpoint['best_loss']
best_acc = checkpoint['best_acc']
except FileNotFoundError:
print("(Initialize) Loading east_vgg16...")
model.load_state_dict(torch.load('./pths/east_vgg16.pth'))
epoch_dict = dict()
test_dict = dict()
total_epoch = 0
best_loss = float('inf')
best_acc = 0
print("Continue from epoch {}".format(total_epoch))
print("Epoch_dict", epoch_dict)
print("Test_dict", test_dict)
scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=[300],
gamma=0.1)
for epoch in range(epoch_iter):
model.train()
scheduler.step()
epoch_loss = 0
test_loss = 0
epoch_time = time.time()
for i, (img, gt_score, gt_geo, ignored_map) in enumerate(train_loader):
start_time = time.time()
img, gt_score, gt_geo, ignored_map = img.to(device), gt_score.to(
device), gt_geo.to(device), ignored_map.to(device)
pred_score, pred_geo = model(img)
loss = criterion(gt_score, pred_score, gt_geo, pred_geo,
ignored_map)
epoch_loss += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
print('Epoch is [{}/{}], mini-batch is [{}/{}], time consumption is {:.8f}, batch_loss is {:.8f}'.format(\
epoch+1, epoch_iter, i+1, int(file_num/batch_size), time.time()-start_time, loss.item()))
epoch_dict[total_epoch + epoch +
1] = (epoch_loss / int(file_num / batch_size), epoch_loss)
print('epoch_loss is {:.8f}, epoch_time is {:.8f}, epoch_loss: {}'.
format(epoch_loss / int(file_num / batch_size),
time.time() - epoch_time, epoch_loss))
model_state_dict = model.module.state_dict(
) if data_parallel else model.state_dict()
with torch.no_grad():
for i, (img, gt_score, gt_geo,
ignored_map) in enumerate(test_loader):
img, gt_score, gt_geo, ignored_map = img.to(
device), gt_score.to(device), gt_geo.to(
device), ignored_map.to(device)
pred_score, pred_geo = model(img)
loss = criterion(gt_score, pred_score, gt_geo, pred_geo,
ignored_map)
test_loss += loss.item()
print('Epoch (test) is [{}/{}], mini-batch is [{}/{}], time consumption is {:.8f}, batch_loss is {:.8f}'.format(\
epoch+1, epoch_iter, i+1, int(file_num2/batch_size), time.time()-start_time, loss.item()))
test_dict[total_epoch + epoch +
1] = (test_loss / int(file_num2 / batch_size), test_loss)
print(
'test_loss is {:.8f}, epoch_time is {:.8f}, test_loss: {}'.format(
test_loss / int(file_num2 / batch_size),
time.time() - epoch_time, test_loss))
print(time.asctime(time.localtime(time.time())))
print('=' * 50)
if (epoch + 1) % interval == 0:
torch.save(
{
'epoch': total_epoch + epoch + 1,
'model_state_dict': model_state_dict,
'optimizer_state_dict': optimizer.state_dict(),
'epoch_loss': epoch_dict,
'test_loss': test_dict,
'best_loss': best_loss,
'best_acc': best_acc
}, os.path.join(pths_path, 'east.pth'))
if (total_epoch + epoch + 1) % 10 == 0:
torch.save(
{
'epoch': total_epoch + epoch + 1,
'model_state_dict': model_state_dict,
'optimizer_state_dict': optimizer.state_dict(),
'epoch_loss': epoch_dict,
'test_loss': test_dict,
'best_loss': best_loss,
'best_acc': best_acc
},
os.path.join(
pths_path,
'east_epoch_{}.pth'.format(total_epoch + epoch + 1)))
if test_loss / int(file_num2 / batch_size) < best_loss:
torch.save(
{
'epoch': total_epoch + epoch + 1,
'model_state_dict': model_state_dict,
'optimizer_state_dict': optimizer.state_dict(),
'epoch_loss': epoch_dict,
'test_loss': test_dict,
'best_loss': best_loss,
'best_acc': best_acc
}, os.path.join(pths_path, 'east_best_loss.pth'))
def train(train_img_path, train_gt_path, pths_path, batch_size, lr,
num_workers, epoch_iter, interval):
file_num = len(os.listdir(train_img_path))
trainset = custom_dataset(train_img_path, train_gt_path)
train_loader = data.DataLoader(trainset, batch_size=batch_size, \
shuffle=True, num_workers=num_workers, drop_last=True)
criterion = Loss()
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = EAST(pretrained=False)
model.load_state_dict(
torch.load('/root/last_dataset/east_tmp_pths/east_model_9_0.2783.pth'))
data_parallel = False
if torch.cuda.device_count() > 1:
model = nn.DataParallel(model)
data_parallel = True
model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
optimizer.zero_grad()
#scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[epoch_iter//2], gamma=0.1)
for epoch in range(epoch_iter):
model.train()
epoch_loss = 0
epoch_time = time.time()
loss_plot = []
bx = []
for i, (img, gt_score, gt_geo, ignored_map) in enumerate(train_loader):
start_time = time.time()
img, gt_score, gt_geo, ignored_map = img.to(device), gt_score.to(
device), gt_geo.to(device), ignored_map.to(device)
pred_score, pred_geo = model(img)
loss = criterion(gt_score, pred_score, gt_geo, pred_geo,
ignored_map)
epoch_loss += loss.item()
loss.backward()
if (i + 1) % 3:
optimizer.step()
optimizer.zero_grad()
if (i + 1) % 100 == 0:
print(
'Epoch is [{}/{}], mini-batch is [{}/{}], time consumption is {:.8f}, batch_loss is {:.8f}'
.format(epoch + 1, epoch_iter, i + 1,
int(file_num / batch_size),
time.time() - start_time, loss.item()))
'''
if (i + 1) % 100 == 0:
loss_plot.append(loss.item())
bx.append(i + epoch * int(file_num / batch_size))
plt.plot(bx, loss_plot, label='loss_mean', linewidth=1, color='b', marker='o',
markerfacecolor='green', markersize=2)
plt.savefig(os.path.abspath('./labeled.jpg'))
'''
print('epoch_loss is {:.8f}, epoch_time is {:.8f}'.format(
epoch_loss / int(file_num / batch_size),
time.time() - epoch_time))
print(time.asctime(time.localtime(time.time())))
print('=' * 50)
if epoch % interval == 0:
state_dict = model.module.state_dict(
) if data_parallel else model.state_dict()
torch.save(
state_dict,
os.path.join(
pths_path, 'east_model_{}_{:.4f}.pth'.format(
epoch + 10, epoch_loss / int(file_num / batch_size))))
def train(train_img_path, train_gt_path, pths_path, batch_size, lr,
num_workers, epoch_iter, interval):
# import pdb
# pdb.set_trace()
# 加载数据
file_num = len(os.listdir(train_img_path))
trainset = custom_dataset(train_img_path, train_gt_path)
train_loader = data.DataLoader(trainset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers,
drop_last=True)
# 加载模型
model = EAST()
data_parallel = False
if torch.cuda.device_count() > 1:
model = nn.DataParallel(model)
data_parallel = True
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)
# 设置loss
criterion = Loss()
# [完善优化算法的调用]写出优化算法
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
# 定义学习策略, milestones is a list of epoch indices, and ust be increasing.
scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=[epoch_iter // 2],
gamma=.1)
for epoch in range(epoch_iter):
model.train()
# when epoch meets epoch_iter // 2,
# this scheduler will schedule learning rate
scheduler.step()
epoch_loss = 0
epoch_time = time.time()
for i, (img, gt_score, gt_geo, ignored_map) in enumerate(train_loader):
start_time = time.time()
print("start_time=%s" % start_time)
# import pdb
# pdb.set_trace()
img, gt_score, gt_geo, ignored_map = img.to(device), gt_score.to(device), \
gt_geo.to(device), ignored_map.to(device)
# 前向反馈
pred_score, pred_geo = model(img)
# 计算loss
loss = criterion(gt_score, pred_score, gt_geo, pred_geo,
ignored_map)
epoch_loss += loss.item()
# 反向传播，优化器梯度需先清零！
optimizer.zero_grad()
loss.backward()
# 模型权重更新
optimizer.step()
print(
'Epoch is [{}/{}], mini-batch is [{}/{}], time consumption is {:.8f}, batch_loss is {:.8f}'
.format(epoch + 1, epoch_iter, i + 1,
int(file_num / batch_size),
time.time() - start_time, loss.item()))
print('epoch_loss is {:.8f}, epoch_time is {:.8f}'.format(
epoch_loss / int(file_num / batch_size),
time.time() - epoch_time))
print(time.asctime(time.localtime(time.time())))
print('=' * 50)
# 每5个周期保存一下模型的权重
if (epoch + 1) % interval == 0:
state_dict = model.module.state_dict(
) if data_parallel else model.state_dict()
torch.save(
state_dict,
os.path.join(pths_path,
'model_epoch_{}.pth'.format(epoch + 1)))