def create_network(train_image, train_label, classes, network='unet_simple', image_size=(1024, 384), for_test=False):
if network == 'unet_base':
predict = unet_base(train_image, classes, image_size)
elif network == 'unet_simple':
predict = unet_simple(train_image, classes, image_size)
elif network == 'deeplabv3p':
predict = deeplabv3p(train_image, classes)
else:
raise Exception('Not support this model:', network)
print('The program will run', network)
if for_test == False:
loss, miou = create_loss(predict, train_label, classes)
return loss, miou, predict
elif for_test == True:
return predict
else:
raise Exception('Wrong Status:', for_test)
def main():
# network = 'deeplabv3p'
# save_model_path = "./model_weights/" + network + "_"
# model_path = "./model_weights/" + network + "_0_6000"
data_dir = ''
val_percent = .1
epochs = 9
kwargs = {
'num_workers': 4,
'pin_memory': True
} if torch.cuda.is_available() else {}
training_dataset = LaneDataset(
"~/workspace/myDL/CV/week8/Lane_Segmentation_pytorch/data_list/train.csv",
transform=transforms.Compose(
[ImageAug(),
DeformAug(),
ScaleAug(),
CutOut(32, 0.5),
ToTensor()]))
training_data_batch = DataLoader(training_dataset,
batch_size=2,
shuffle=True,
drop_last=True,
**kwargs)
dataset = BasicDataset(data_dir,
img_size=cfg.IMG_SIZE,
crop_offset=cfg.crop_offset)
n_val = int(len(dataset) * val_percent)
n_train = len(dataset) - n_val
train, val = random_split(dataset, [n_train, n_val])
train_loader = DataLoader(train,
batch_size=cfg.batch_size,
shuffle=True,
num_workers=8,
pin_memory=True)
val_loader = DataLoader(val,
batch_size=cfg.batch_size,
shuffle=False,
num_workers=8,
pin_memory=True)
model = unet_base(cfg.num_classes, cfg.IMG_SIZE)
model.cuda()
optimizer = torch.optim.Adam(model.parameters(),
lr=cfg.base_lr,
betas=(0.9, 0.99))
bce_criterion = nn.BCEWithLogitsLoss()
dice_criterion = MulticlassDiceLoss()
model.train()
epoch_loss = 0
dataprocess = tqdm(training_data_batch)
for batch_item in dataprocess:
image, mask = batch_item['image'], batch_item['mask']
if torch.cuda.is_available():
image, mask = image.cuda(), mask.cuda()
image = image.to(torch.float32).requires_grad_()
mask = mask.to(torch.float32).requires_grad_()
masks_pred = model(image)
masks_pred = torch.argmax(masks_pred, dim=1)
masks_pred = masks_pred.to(torch.float32)
mask = mask.to(torch.float32)
# print('mask_pred:', masks_pred)
# print('mask:', mask)
loss = bce_criterion(masks_pred, mask) + dice_criterion(
masks_pred, mask)
epoch_loss += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
def main():
# network = 'deeplabv3p'
# save_model_path = "./model_weights/" + network + "_"
# model_path = "./model_weights/" + network + "_0_6000"
data_dir = '/env/data_list_for_Lane_Segmentation/train.csv'
val_percent = .1
epochs = 9
dataset = BasicDataset(data_dir,
img_size=cfg.IMG_SIZE,
crop_offset=cfg.crop_offset)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
n_val = int(len(dataset) * val_percent)
n_train = len(dataset) - n_val
train, val = random_split(dataset, [n_train, n_val])
train_loader = DataLoader(train,
batch_size=cfg.batch_size,
shuffle=True,
num_workers=0,
pin_memory=True)
val_loader = DataLoader(val,
batch_size=cfg.batch_size,
shuffle=False,
num_workers=0,
pin_memory=True)
model = unet_base(cfg.num_classes, cfg.IMG_SIZE)
optimizer = torch.optim.Adam(model.parameters(),
lr=cfg.base_lr,
betas=(0.9, 0.99))
bce_criterion = nn.BCEWithLogitsLoss()
dice_criterion = MulticlassDiceLoss()
for epoch in range(epochs):
model.train()
epoch_loss = 0
with tqdm(total=n_train,
desc=f'Epoch {epoch + 1}/{epochs}',
unit='img') as pbar:
for batch in train_loader:
imgs = batch['image']
true_masks = batch['mask']
imgs = imgs.to(device=device, dtype=torch.float32)
mask_type = torch.float32 if cfg.num_classes == 1 else torch.long
true_masks = true_masks.to(device=device, dtype=mask_type)
masks_pred = model(imgs)
loss = bce_criterion(masks_pred, true_masks) + dice_criterion(
masks_pred, true_masks)
epoch_loss += loss.item()
pbar.set_postfix(**{'loss (batch)': loss.item()})
optimizer.zero_grad()
loss.backward()
optimizer.step()
pbar.update(imgs.shape[0])
global_step += 1
if global_step % (len(dataset) // (10 * cfg.batch_size)) == 0:
val_score = eval_net(model, val_loader, device, n_val)
print('val_score:', val_score)
def main():
batch_size = 16
# 训练ckpt和inf模型路径
param_base_dir = os.path.join(code_base_dir, "params")
param_base_dir = os.path.join(param_base_dir, args.net)
infer_param_path = os.path.join(param_base_dir, "inf")
ckpt_param_path = os.path.join(param_base_dir, "ckpt")
print(infer_param_path)
print(ckpt_param_path)
test_reader = paddle.batch(data_reader(8, 10, True), batch_size)
train_program = fluid.Program()
train_init = fluid.Program()
with fluid.program_guard(train_program, train_init):
image = fluid.layers.data(name="image",
shape=[3, 512, 512],
dtype="float32")
label = fluid.layers.data(name="label",
shape=[1, 512, 512],
dtype="int32")
if args.net == "unet_simple":
prediction = unet_simple(image, 2, [512, 512])
elif args.net == "unet_base":
prediction = unet_base(image, 2, [512, 512])
elif args.net == "deeplabv3":
prediction = deeplabv3p(image, 2)
else:
print("错误的网络类型")
sys.exit(0)
avg_loss = create_loss(prediction, label, 2)
miou = mean_iou(prediction, label, 2)
decay = paddle.fluid.regularizer.L2Decay(0.0001)
optimizer = fluid.optimizer.AdamOptimizer(learning_rate=0.003,
regularization=decay)
# optimizer = fluid.optimizer.MomentumOptimizer(learning_rate=0.006, momentum=0.8,regularization=decay)
optimizer.minimize(avg_loss)
place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(train_init)
# fluid.io.load_persistables(exe, ckpt_param_path, train_init)
exe_test = fluid.Executor(place)
test_program = train_program.clone(for_test=True)
train_program = fluid.CompiledProgram(train_program).with_data_parallel(
loss_name=avg_loss.name)
# test_program = fluid.CompiledProgram(test_program).with_data_parallel(loss_name=avg_loss.name)
train_feeder = fluid.DataFeeder(place=place, feed_list=[image, label])
test_feeder = fluid.DataFeeder(place=place, feed_list=[image, label])
step = 1
best_miou = 0
train_prompt = "Train_miou"
test_prompt = "Test_miou"
# plot_prompt = Ploter(train_prompt, test_prompt)
writer = LogWriter(logdir="../log/{}".format(datetime.now()))
train_reader = fluid.io.xmap_readers(aug_mapper, data_reader(0, 8), 8, 32,
False)
step = 0
eval_step = 1000
for pass_id in range(args.num_epochs):
batch = []
for record in train_reader():
batch.append(record)
if len(batch) == batch_size:
step += 1
avg_loss_value, miou_value = exe.run(
train_program,
feed=train_feeder.feed(batch),
fetch_list=[avg_loss, miou])
batch = []
writer.add_scalar(tag="train_loss",
step=step,
value=avg_loss_value[0])
if step % 10 == 0:
print(
"\t\tTrain pass {}, Step {}, Cost {}, Miou {}".format(
pass_id, step, avg_loss_value[0], miou_value[0]))
eval_miou = 0
if step % eval_step == 0:
auc_metric = fluid.metrics.Auc("AUC")
test_losses = []
test_mious = []
for test_data in test_reader():
_, test_loss, test_miou = exe_test.run(
test_program,
feed=test_feeder.feed(test_data),
fetch_list=[prediction, avg_loss, miou])
test_losses.append(test_loss[0])
test_mious.append(test_miou[0])
eval_miou = np.average(np.array(test_mious))
writer.add_scalar(tag="test_loss",
step=step / eval_step,
value=np.average(np.array(test_losses)))
print("Test loss: {} ,miou: {}".format(
np.average(np.array(test_losses)), eval_miou))
if math.isnan(float(avg_loss_value[0])):
sys.exit("got NaN loss, training failed.")
if step % eval_step == 0 and eval_miou > best_miou:
best_miou = eval_miou
print("Saving params of step: %d" % step)
fluid.io.save_inference_model(
infer_param_path,
feeded_var_names=["image"],
target_vars=[prediction],
executor=exe,
main_program=train_program,
)
# fluid.io.save_persistables(exe, ckpt_param_path, train_program)
print(best_miou)
def main():
batch_size = 32
num_epochs = args.num_epochs
net_type = args.net
# 训练ckpt和inf模型路径
param_base_dir = os.path.join(code_base_dir, 'params')
param_base_dir = os.path.join(param_base_dir, net_type)
infer_param_path = os.path.join(param_base_dir, "inf")
ckpt_param_path = os.path.join(param_base_dir, "ckpt")
print(infer_param_path)
print(ckpt_param_path)
train_reader = paddle.batch(
paddle.reader.shuffle(data_reader(), int(batch_size * 1.5)),
batch_size)
test_reader = paddle.batch(
paddle.reader.shuffle(data_reader(8, 10), int(batch_size * 1.5)),
batch_size)
train_program = fluid.Program()
train_init = fluid.Program()
with fluid.program_guard(train_program, train_init):
image = fluid.layers.data(name='image',
shape=[3, 512, 512],
dtype='float32')
label = fluid.layers.data(name='label',
shape=[1, 512, 512],
dtype='int32')
train_loader = fluid.io.DataLoader.from_generator(
feed_list=[image, label], capacity=batch_size)
test_loader = fluid.io.DataLoader.from_generator(
feed_list=[image, label], capacity=batch_size)
if net_type == "unet_simple":
prediction = unet_simple(image, 2, [512, 512])
elif net_type == "unet_base":
prediction = unet_base(image, 2, [512, 512])
elif net_type == "deeplabv3":
prediction = deeplabv3p(image, 2)
else:
print("错误的网络类型")
sys.exit(0)
avg_loss = create_loss(prediction, label, 2)
miou = mean_iou(prediction, label, 2)
# decay=paddle.fluid.regularizer.L2Decay(0.1)
# optimizer = fluid.optimizer.SGD(learning_rate=0.0005,regularization=decay)
# optimizer = fluid.optimizer.DecayedAdagradOptimizer(learning_rate=0.02,regularization=decay)
# optimizer = fluid.optimizer.RMSProp(learning_rate=0.1,momentum=0.8,centered=True, regularization=decay)
optimizer = fluid.optimizer.AdamOptimizer(learning_rate=0.003)
# optimizer = fluid.optimizer.MomentumOptimizer(learning_rate=0.006, momentum=0.8,regularization=decay)
optimizer.minimize(avg_loss)
place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(train_init)
# fluid.io.load_persistables(exe, ckpt_param_path, train_init)
exe_test = fluid.Executor(place)
test_program = train_program.clone(for_test=True)
# train_program=fluid.CompiledProgram(train_program).with_data_parallel(loss_name=avg_loss.name)
test_program = fluid.CompiledProgram(test_program).with_data_parallel(
loss_name=avg_loss.name)
train_loader.set_sample_list_generator(train_reader, places=place)
test_loader.set_sample_list_generator(test_reader, places=place)
feeder = fluid.DataFeeder(place=place, feed_list=[image, label])
step = 1
best_miou = 0
train_prompt = "Train_miou"
test_prompt = "Test_miou"
plot_prompt = Ploter(train_prompt, test_prompt)
for pass_id in range(num_epochs):
for data_train in train_loader():
avg_loss_value, miou_value = exe.run(train_program,
feed=data_train,
fetch_list=[avg_loss, miou])
if step % 10 == 0:
print("\t\tTrain pass %d, Step %d, Cost %f, Miou %f" %
(pass_id, step, avg_loss_value[0], miou_value[0]))
# if step % 10 ==0:
# plot_prompt.append(train_prompt, step, miou_value[0])
# plot_prompt.plot()
eval_miou = 0
if step % 100 == 0:
auc_metric = fluid.metrics.Auc("AUC")
test_losses = []
test_mious = []
for _, test_data in enumerate(test_loader()):
# print(test_data)
# input("pause")
_, test_loss, test_miou = exe_test.run(
test_program,
feed=test_data,
fetch_list=[prediction, avg_loss, miou])
test_losses.append(test_loss[0])
test_mious.append(test_miou[0])
eval_miou = np.average(np.array(test_mious))
# plot_prompt.append(test_prompt, step, eval_miou)
# plot_prompt.plot()
print("Test loss: %f ,miou: %f" %
(np.average(np.array(test_losses)), eval_miou))
if math.isnan(float(avg_loss_value[0])):
sys.exit("got NaN loss, training failed.")
if step % 100 == 0 and param_base_dir is not None and eval_miou > best_miou:
best_miou = eval_miou
print("Saving params of step: %d" % step)
fluid.io.save_inference_model(infer_param_path,
feeded_var_names=['image'],
target_vars=[prediction],
executor=exe,
main_program=train_program)
fluid.io.save_persistables(exe, ckpt_param_path, train_program)
step += 1
print(best_miou)