def train(initial_checkpoint):
## setup -----------------
os.makedirs(CHECKPOINTS + '/checkpoint', exist_ok=True)
os.makedirs(CHECKPOINTS + '/train', exist_ok=True)
os.makedirs(CHECKPOINTS + '/backup', exist_ok=True)
log = Logger()
log.write('\n--- [START %s] %s\n\n' % (IDENTIFIER, '-' * 64))
log.write('\tSEED = %u\n' % SEED)
log.write('\tPROJECT_PATH = %s\n' % CODE)
log.write('\t__file__ = %s\n' % __file__)
log.write('\tRESULT = %s\n' % CHECKPOINTS)
log.write('\n')
log.write('\t<additional comments>\n')
log.write('\t ... \n')
log.write('\n')
## dataset ----------------------------------------
log.write('Configuring dataset...\n')
batch_size = 16
train_dataset = TGSDataset(
'list_train' + str(FOLD) + '_3600' + ne + "_balanced", train_augment,
'train')
os.makedirs(CHECKPOINTS + '/list_train' + str(FOLD) + '_3600' + ne +
"_balanced",
exist_ok=True)
train_loader = DataLoader(
train_dataset,
sampler=RandomSampler(train_dataset),
#sampler = ConstantSampler(train_dataset,[31]*batch_size*100),
batch_size=batch_size,
drop_last=True,
num_workers=8,
pin_memory=True,
collate_fn=null_collate)
valid_dataset = TGSDataset(
'list_valid' + str(FOLD) + '_400' + ne + "_balanced", valid_augment,
'train')
valid_loader = DataLoader(valid_dataset,
sampler=RandomSampler(valid_dataset),
batch_size=batch_size,
drop_last=False,
num_workers=8,
pin_memory=True,
collate_fn=null_collate)
assert (len(train_dataset) >= batch_size)
log.write('batch_size = %d\n' % (batch_size))
log.write('train_dataset.split = %s\n' % (train_dataset.split))
log.write('valid_dataset.split = %s\n' % (valid_dataset.split))
log.write('\n')
#debug
if 0: #debug ##-------------------------------
for input, truth, index, cache in train_loader:
images = input.cpu().data.numpy().squeeze()
masks = truth.cpu().data.numpy().squeeze()
batch_size = len(index)
for b in range(batch_size):
image = images[b] * 255
image = np.dstack([image, image, image])
mask = masks[b]
image_show('image', image, resize=2)
image_show_norm('mask', mask, max=1, resize=2)
overlay0 = draw_mask_overlay(mask, image, color=[0, 0, 255])
overlay0 = draw_mask_to_contour_overlay(mask,
overlay0,
2,
color=[0, 0, 255])
image_show('overlay0', overlay0, resize=2)
cv2.waitKey(0)
#--------------------------------------
## net ----------------------------------------
log.write('Configuring neural network...\n')
net = Net().cuda()
if initial_checkpoint is not None:
log.write('\tinitial_checkpoint = %s\n' % initial_checkpoint)
net.load_state_dict(
torch.load(initial_checkpoint,
map_location=lambda storage, loc: storage))
log.write("The net is an instance of {}.".format(type(net)))
log.write('\n')
## optimiser ----------------------------------
num_iters = 300 * 1000
iter_smooth = 20
iter_log = 50
iter_valid = 100
iter_save = [0, num_iters-1]\
+ list(range(0,num_iters,500))#1*1000
FREEZE = False
#------------------------------------------------------
if FREEZE: ##freeze
for p in net.feature_net.parameters():
p.requires_grad = False
#------------------------------------------------------
scheduler = lambda x: (0.01 / 2) * (np.cos(PI * (np.mod(
x - 1, 300 * 1000 / 30) / (300 * 1000 / 30))) + 1)
#log.write(scheduler(1))
#log.write(scheduler(5000))
#log.write(scheduler(10001))
optimizer = optim.SGD(filter(lambda p: p.requires_grad, net.parameters()),
lr=0.01,
momentum=0.9,
weight_decay=0.0001)
start_iter = 0
start_epoch = 0
if initial_checkpoint is not None:
checkpoint = torch.load(
initial_checkpoint.replace('_model.pth', '_optimizer.pth'))
start_iter = checkpoint['iter']
start_epoch = checkpoint['epoch']
rate = get_learning_rate(optimizer) #load all except learning rate
optimizer.load_state_dict(checkpoint['optimizer'])
adjust_learning_rate(optimizer, rate)
pass
## start training here! ##############################################
log.write('Start training...\n')
log.write(
' rate iter epoch | valid_loss | train_loss | batch_loss | time \n'
)
log.write(
'-------------------------------------------------------------------------------------------------------------------------------\n'
)
train_loss = np.zeros(6, np.float32)
valid_loss = np.zeros(6, np.float32)
batch_loss = np.zeros(6, np.float32)
rate = 0
iter = 0
i = 0
start = timer()
while iter < num_iters: # loop over the dataset multiple times
sum_train_loss = np.zeros(6, np.float32)
sum = 0
optimizer.zero_grad()
for input, truth, index, cache in train_loader:
if 0: #debug ##-------------------------------
image = input.cpu().data.numpy().squeeze()
mask = truth.cpu().data.numpy().squeeze()
batch_size = len(index)
for b in range(batch_size):
image_show_norm('image', image[b], max=1, resize=2)
image_show_norm('mask', mask[b], max=1, resize=2)
cv2.waitKey(0)
#--------------------------------------
len_train_dataset = len(train_dataset)
batch_size = len(index)
iter = i + start_iter
epoch = (iter -
start_iter) * batch_size / len_train_dataset + start_epoch
num_samples = epoch * len_train_dataset
if iter % iter_valid == 0:
net.set_mode('valid')
valid_loss = validation(net, valid_loader)
net.set_mode('train')
log.write2('\r')
log.write('%0.4f %5.1f %6.1f | %0.3f %0.3f (%0.3f) | %0.3f %0.3f | %0.3f %0.3f | %s \n' % (\
rate, iter/1000, epoch,
valid_loss[0], valid_loss[1], valid_loss[2],
train_loss[0], train_loss[1],
batch_loss[0], batch_loss[1],
time_to_str((timer() - start),'min')))
time.sleep(0.01)
if iter in iter_save:
torch.save(
net.state_dict(), CHECKPOINTS + "/" + train_dataset.split +
'/' + MODEL + OHEM + '%08d_model.pth' % (iter))
torch.save(
{
'optimizer': optimizer.state_dict(),
'iter': iter,
'epoch': epoch,
}, CHECKPOINTS + "/" + train_dataset.split + '/' + MODEL +
OHEM + '%08d_optimizer.pth' % (iter))
pass
# learning rate schduler -------------
if scheduler is not None:
#scheduler.batch_step()
lr = scheduler(iter)
if lr < 0: break
adjust_learning_rate(optimizer, lr)
rate = get_learning_rate(optimizer)
#rate = 0.01
# one iteration update -------------
#net.set_mode('train',is_freeze_bn=True)
net.set_mode('train')
input = input.cuda()
truth = truth.cuda()
logit = data_parallel(net, input) #net(input)
if OHEM == "OHEM":
loss = net.focal_loss(logit, truth, 1.0, 0.5,
0.25) + net.criterion(logit, truth)
else:
loss = net.criterion(logit, truth)
dice = net.metric(logit, truth)
loss.backward()
optimizer.step()
optimizer.zero_grad()
#torch.nn.utils.clip_grad_norm(net.parameters(), 1)
# print statistics ------------
batch_loss = np.array((
loss.item(),
dice.item(),
0,
0,
0,
0,
))
sum_train_loss += batch_loss
sum += 1
if iter % iter_smooth == 0:
train_loss = sum_train_loss / sum
sum_train_loss = np.zeros(6, np.float32)
sum = 0
log.write2('\r%0.4f %5.1f %6.1f | %0.3f %0.3f (%0.3f) | %0.3f %0.3f | %0.3f %0.3f | %s ' % (\
rate, iter/1000, epoch,
valid_loss[0], valid_loss[1], valid_loss[2],
train_loss[0], train_loss[1],
batch_loss[0], batch_loss[1],
time_to_str((timer() - start), 'min')))
i = i + 1
pass #-- end of one data loader --
pass #-- end of all iterations --
log.write('\n')
def train():
initial_checkpoint = None
#'checkpoint/00048500_model.pth'\
# None #'/root/share/project/kaggle/tgs/results/resnet34-resize128-focus/fold0-1a/checkpoint/00003500_model.pth'
## setup -----------------
os.makedirs(CHECKPOINTS + '/checkpoint', exist_ok=True)
os.makedirs(CHECKPOINTS + '/train', exist_ok=True)
os.makedirs(CHECKPOINTS + '/backup', exist_ok=True)
#backup_project_as_zip(PROJECT_PATH, RESULT +'/backup/code.train.%s.zip'%IDENTIFIER)
log = Logger()
#log.open(RESULT+'/log.train.txt',mode='a')
print('\n--- [START %s] %s\n\n' % (IDENTIFIER, '-' * 64))
print('\tSEED = %u\n' % SEED)
print('\tPROJECT_PATH = %s\n' % CODE)
print('\t__file__ = %s\n' % __file__)
print('\tRESULT = %s\n' % CHECKPOINTS)
print('\n')
print('\t<additional comments>\n')
print('\t ... \n')
print('\n')
## dataset ----------------------------------------
print('Configuring dataset...\n')
batch_size = 16
train_dataset = TGSDataset('list_train8_3600', train_augment, 'train')
os.makedirs(CHECKPOINTS + '/list_train8_3600', exist_ok=True)
train_loader = DataLoader(
train_dataset,
sampler=RandomSampler(train_dataset),
#sampler = ConstantSampler(train_dataset,[31]*batch_size*100),
batch_size=batch_size,
drop_last=True,
num_workers=8,
pin_memory=True,
collate_fn=null_collate)
valid_dataset = TGSDataset('list_valid8_400', valid_augment, 'train')
valid_loader = DataLoader(valid_dataset,
sampler=RandomSampler(valid_dataset),
batch_size=batch_size,
drop_last=False,
num_workers=8,
pin_memory=True,
collate_fn=null_collate)
assert (len(train_dataset) >= batch_size)
print('batch_size = %d\n' % (batch_size))
print('train_dataset.split = %s\n' % (train_dataset.split))
print('valid_dataset.split = %s\n' % (valid_dataset.split))
print('\n')
#debug
if 0: #debug ##-------------------------------
for input, truth, index, cache in train_loader:
images = input.cpu().data.numpy().squeeze()
masks = truth.cpu().data.numpy().squeeze()
batch_size = len(index)
for b in range(batch_size):
image = images[b] * 255
image = np.dstack([image, image, image])
mask = masks[b]
image_show('image', image, resize=2)
image_show_norm('mask', mask, max=1, resize=2)
overlay0 = draw_mask_overlay(mask, image, color=[0, 0, 255])
overlay0 = draw_mask_to_contour_overlay(mask,
overlay0,
2,
color=[0, 0, 255])
image_show('overlay0', overlay0, resize=2)
cv2.waitKey(0)
#--------------------------------------
## net ----------------------------------------
print('Configuring neural network...\n')
net = Net().cuda()
if initial_checkpoint is not None:
print('\tinitial_checkpoint = %s\n' % initial_checkpoint)
net.load_state_dict(
torch.load(initial_checkpoint,
map_location=lambda storage, loc: storage))
print("The net is an instance of {}.".format(type(net)))
print('\n')
## optimiser ----------------------------------
num_iters = 300 * 1000
iter_smooth = 20
iter_log = 50
iter_valid = 100
iter_save = [0, num_iters-1]\
+ list(range(0,num_iters,500))#1*1000
FREEZE = False
#------------------------------------------------------
if FREEZE: ##freeze
for p in net.feature_net.parameters():
p.requires_grad = False
#from cls import CyclicLR
#net.set_mode('train',is_freeze_bn=True)
#------------------------------------------------------
scheduler = lambda x: (0.01 / 2) * (np.cos(PI * (np.mod(
x - 1, 300 * 1000 / 30) / (300 * 1000 / 30))) + 1)
print(scheduler(1))
print(scheduler(5000))
print(scheduler(10001))
#scheduler = CyclicLR(base_lr=0.01, max_lr=0.01, step_size=10000, gamma=1., scale_fn=clr_fn, scale_mode='iterations')
#schduler = None #StepLR([ (0, 0.01), (200, 0.001)])
#base_params = list(map(id, net.resnet.parameters()))
#decode_params = filter(lambda p: id(p) not in base_params, net.parameters())
#params = [ {"params": decode_params, "lr": 0.01},
# {"params": net.resnet.parameters(), "lr": 0.005}, ]
#optimizer = torch.optim.SGD(params, momentum=0.9, weight_decay=0.0001)
optimizer = optim.SGD(filter(lambda p: p.requires_grad, net.parameters()),
lr=0.01,
momentum=0.9,
weight_decay=0.0001)
#scheduler = CyclicLR(optimizer,base_lr=0.01, max_lr=0.01, step_size=10000, gamma=1., scale_fn=clr_fn, scale_mode='iterations')
#scheduler= CyclicLR(optimizer, base_lr=0.001, max_lr=0.01, step_size=10000, gamma=0.99, mode='cos_anneal')
start_iter = 0
start_epoch = 0
if initial_checkpoint is not None:
checkpoint = torch.load(
initial_checkpoint.replace('_model.pth', '_optimizer.pth'))
start_iter = checkpoint['iter']
start_epoch = checkpoint['epoch']
rate = get_learning_rate(optimizer) #load all except learning rate
#optimizer.load_state_dict(checkpoint['optimizer'])
adjust_learning_rate(optimizer, rate)
pass
## start training here! ##############################################
print('Start training...\n')
#print(' samples_per_epoch = %d\n\n'%len(train_dataset))
print(
' rate iter epoch | valid_loss | train_loss | batch_loss | time \n'
)
print(
'-------------------------------------------------------------------------------------------------------------------------------\n'
)
train_loss = np.zeros(6, np.float32)
valid_loss = np.zeros(6, np.float32)
batch_loss = np.zeros(6, np.float32)
rate = 0
iter = 0
i = 0
start = timer()
while iter < num_iters: # loop over the dataset multiple times
sum_train_loss = np.zeros(6, np.float32)
sum = 0
optimizer.zero_grad()
for input, truth, index, cache in train_loader:
if 0: #debug ##-------------------------------
image = input.cpu().data.numpy().squeeze()
mask = truth.cpu().data.numpy().squeeze()
batch_size = len(index)
for b in range(batch_size):
image_show_norm('image', image[b], max=1, resize=2)
image_show_norm('mask', mask[b], max=1, resize=2)
cv2.waitKey(0)
#--------------------------------------
len_train_dataset = len(train_dataset)
batch_size = len(index)
iter = i + start_iter
epoch = (iter -
start_iter) * batch_size / len_train_dataset + start_epoch
num_samples = epoch * len_train_dataset
if iter % iter_valid == 0:
net.set_mode('valid')
valid_loss = validation(net, valid_loader)
net.set_mode('train')
print('\r', end='', flush=True)
print('%0.4f %5.1f %6.1f | %0.3f %0.3f (%0.3f) | %0.3f %0.3f | %0.3f %0.3f | %s \n' % (\
rate, iter/1000, epoch,
valid_loss[0], valid_loss[1], valid_loss[2],
train_loss[0], train_loss[1],
batch_loss[0], batch_loss[1],
time_to_str((timer() - start),'min')))
time.sleep(0.01)
#if 1:
if iter in iter_save:
torch.save(
net.state_dict(), CHECKPOINTS + "/" + train_dataset.split +
'/%08d_model.pth' % (iter))
torch.save(
{
'optimizer': optimizer.state_dict(),
'iter': iter,
'epoch': epoch,
}, CHECKPOINTS + "/" + train_dataset.split +
'/%08d_optimizer.pth' % (iter))
pass
# learning rate schduler -------------
if scheduler is not None:
#scheduler.batch_step()
lr = scheduler(iter)
if lr < 0: break
adjust_learning_rate(optimizer, lr)
rate = get_learning_rate(optimizer)
#rate = 0.01
# one iteration update -------------
#net.set_mode('train',is_freeze_bn=True)
net.set_mode('train')
input = input.cuda()
truth = truth.cuda()
logit = data_parallel(net, input) #net(input)
loss = net.criterion(logit, truth)
#loss = torch.nn.BCEWithLogitsLoss(logit,truth)
dice = net.metric(logit, truth)
loss.backward()
optimizer.step()
optimizer.zero_grad()
#torch.nn.utils.clip_grad_norm(net.parameters(), 1)
# print statistics ------------
batch_loss = np.array((
loss.item(),
dice.item(),
0,
0,
0,
0,
))
sum_train_loss += batch_loss
sum += 1
if iter % iter_smooth == 0:
train_loss = sum_train_loss / sum
sum_train_loss = np.zeros(6, np.float32)
sum = 0
print('\r%0.4f %5.1f %6.1f | %0.3f %0.3f (%0.3f) | %0.3f %0.3f | %0.3f %0.3f | %s ' % (\
rate, iter/1000, epoch,
valid_loss[0], valid_loss[1], valid_loss[2],
train_loss[0], train_loss[1],
batch_loss[0], batch_loss[1],
time_to_str((timer() - start), 'min')), end='',flush=True)
i = i + 1
#<debug> ===================================================================
if 0:
#if iter%200==0:
#voxel, aux, query, link, truth, cache = make_valid_batch(valid_dataset.dataset, batch_size=2)
net.set_mode('test') #
with torch.no_grad():
logit = net(input)
prob = F.sigmoid(logit)
loss = net.criterion(logit, truth)
dice = net.metric(logit, truth)
if 0:
loss = net.criterion(logit, truth)
accuracy, hit_rate, precision_rate = net.metric(
logit, truth)
valid_loss[0] = loss.item()
valid_loss[1] = accuracy.item()
valid_loss[2] = hit_rate.item()
valid_loss[3] = precision_rate.item()
#show only b in batch ---
b = 1
prob = prob.data.cpu().numpy()[b].squeeze()
truth = truth.data.cpu().numpy()[b].squeeze()
input = input.data.cpu().numpy()[b].squeeze()
all = np.hstack([input, truth, prob])
image_show_norm('all', all, max=1, resize=3)
cv2.waitKey(100)
net.set_mode('train')
#<debug> ===================================================================
pass #-- end of one data loader --
pass #-- end of all iterations --
if 1: #save last
torch.save(
net.state_dict(), CHECKPOINTS + '/checkpoint/' +
train_dataset.split + '/%d_model.pth' % (i))
torch.save(
{
'optimizer': optimizer.state_dict(),
'iter': i,
'epoch': epoch,
}, CHECKPOINTS + '/checkpoint/' + train_dataset.split +
'/%d_optimizer.pth' % (i))
print('\n')