def __init__(self, args):
self.cfg = DeployConfig(args.cfg)
self.args = args
self.compose = T.Compose(self.cfg.transforms)
resize_h, resize_w = args.input_shape
self.disflow = cv2.DISOpticalFlow_create(
cv2.DISOPTICAL_FLOW_PRESET_ULTRAFAST)
self.prev_gray = np.zeros((resize_h, resize_w), np.uint8)
self.prev_cfd = np.zeros((resize_h, resize_w), np.float32)
self.is_init = True
pred_cfg = PredictConfig(self.cfg.model, self.cfg.params)
pred_cfg.disable_glog_info()
if self.args.use_gpu:
pred_cfg.enable_use_gpu(100, 0)
self.predictor = create_predictor(pred_cfg)
if self.args.test_speed:
self.cost_averager = TimeAverager()
def train(model,
train_dataset,
val_dataset=None,
optimizer=None,
save_dir='output',
iters=10000,
batch_size=2,
resume_model=None,
save_interval=1000,
log_iters=10,
num_workers=0,
use_vdl=False,
losses=None,
keep_checkpoint_max=5,
eval_begin_iters=None):
"""
Launch training.
Args:
model（nn.Layer): A matting model.
train_dataset (paddle.io.Dataset): Used to read and process training datasets.
val_dataset (paddle.io.Dataset, optional): Used to read and process validation datasets.
optimizer (paddle.optimizer.Optimizer): The optimizer.
save_dir (str, optional): The directory for saving the model snapshot. Default: 'output'.
iters (int, optional): How may iters to train the model. Defualt: 10000.
batch_size (int, optional): Mini batch size of one gpu or cpu. Default: 2.
resume_model (str, optional): The path of resume model.
save_interval (int, optional): How many iters to save a model snapshot once during training. Default: 1000.
log_iters (int, optional): Display logging information at every log_iters. Default: 10.
num_workers (int, optional): Num workers for data loader. Default: 0.
use_vdl (bool, optional): Whether to record the data to VisualDL during training. Default: False.
losses (dict, optional): A dict of loss, refer to the loss function of the model for details. Default: None.
keep_checkpoint_max (int, optional): Maximum number of checkpoints to save. Default: 5.
eval_begin_iters (int): The iters begin evaluation. It will evaluate at iters/2 if it is None. Defalust: None.
"""
model.train()
nranks = paddle.distributed.ParallelEnv().nranks
local_rank = paddle.distributed.ParallelEnv().local_rank
start_iter = 0
if resume_model is not None:
start_iter = resume(model, optimizer, resume_model)
if not os.path.isdir(save_dir):
if os.path.exists(save_dir):
os.remove(save_dir)
os.makedirs(save_dir)
if nranks > 1:
# Initialize parallel environment if not done.
if not paddle.distributed.parallel.parallel_helper._is_parallel_ctx_initialized(
):
paddle.distributed.init_parallel_env()
ddp_model = paddle.DataParallel(model)
else:
ddp_model = paddle.DataParallel(model)
batch_sampler = paddle.io.DistributedBatchSampler(train_dataset,
batch_size=batch_size,
shuffle=True,
drop_last=True)
loader = paddle.io.DataLoader(
train_dataset,
batch_sampler=batch_sampler,
num_workers=num_workers,
return_list=True,
)
if use_vdl:
from visualdl import LogWriter
log_writer = LogWriter(save_dir)
avg_loss = defaultdict(float)
iters_per_epoch = len(batch_sampler)
best_sad = np.inf
best_model_iter = -1
reader_cost_averager = TimeAverager()
batch_cost_averager = TimeAverager()
save_models = deque()
batch_start = time.time()
iter = start_iter
while iter < iters:
for data in loader:
iter += 1
if iter > iters:
break
reader_cost_averager.record(time.time() - batch_start)
# model input
if nranks > 1:
logit_dict = ddp_model(data)
else:
logit_dict = model(data)
loss_dict = model.loss(logit_dict, data, losses)
loss_dict['all'].backward()
optimizer.step()
lr = optimizer.get_lr()
if isinstance(optimizer._learning_rate,
paddle.optimizer.lr.LRScheduler):
optimizer._learning_rate.step()
model.clear_gradients()
for key, value in loss_dict.items():
avg_loss[key] += value.numpy()[0]
batch_cost_averager.record(time.time() - batch_start,
num_samples=batch_size)
if (iter) % log_iters == 0 and local_rank == 0:
for key, value in avg_loss.items():
avg_loss[key] = value / log_iters
remain_iters = iters - iter
avg_train_batch_cost = batch_cost_averager.get_average()
avg_train_reader_cost = reader_cost_averager.get_average()
eta = calculate_eta(remain_iters, avg_train_batch_cost)
logger.info(
"[TRAIN] epoch={}, iter={}/{}, loss={:.4f}, lr={:.6f}, batch_cost={:.4f}, reader_cost={:.5f}, ips={:.4f} samples/sec | ETA {}"
.format((iter - 1) // iters_per_epoch + 1, iter, iters,
avg_loss['all'], lr, avg_train_batch_cost,
avg_train_reader_cost,
batch_cost_averager.get_ips_average(), eta))
# print loss
loss_str = '[TRAIN] [LOSS] '
loss_str = loss_str + 'all={:.4f}'.format(avg_loss['all'])
for key, value in avg_loss.items():
if key != 'all':
loss_str = loss_str + ' ' + key + '={:.4f}'.format(
value)
logger.info(loss_str)
if use_vdl:
for key, value in avg_loss.items():
log_tag = 'Train/' + key
log_writer.add_scalar(log_tag, value, iter)
log_writer.add_scalar('Train/lr', lr, iter)
log_writer.add_scalar('Train/batch_cost',
avg_train_batch_cost, iter)
log_writer.add_scalar('Train/reader_cost',
avg_train_reader_cost, iter)
for key in avg_loss.keys():
avg_loss[key] = 0.
reader_cost_averager.reset()
batch_cost_averager.reset()
# save model
if (iter % save_interval == 0
or iter == iters) and local_rank == 0:
current_save_dir = os.path.join(save_dir,
"iter_{}".format(iter))
if not os.path.isdir(current_save_dir):
os.makedirs(current_save_dir)
paddle.save(model.state_dict(),
os.path.join(current_save_dir, 'model.pdparams'))
paddle.save(optimizer.state_dict(),
os.path.join(current_save_dir, 'model.pdopt'))
save_models.append(current_save_dir)
if len(save_models) > keep_checkpoint_max > 0:
model_to_remove = save_models.popleft()
shutil.rmtree(model_to_remove)
# eval model
if eval_begin_iters is None:
eval_begin_iters = iters // 2
if (iter % save_interval == 0 or iter == iters) and (
val_dataset is not None
) and local_rank == 0 and iter >= eval_begin_iters:
num_workers = 1 if num_workers > 0 else 0
sad, mse = evaluate(model,
val_dataset,
num_workers=0,
print_detail=True,
save_results=False)
model.train()
# save best model and add evaluation results to vdl
if (iter % save_interval == 0
or iter == iters) and local_rank == 0:
if val_dataset is not None and iter >= eval_begin_iters:
if sad < best_sad:
best_sad = sad
best_model_iter = iter
best_model_dir = os.path.join(save_dir, "best_model")
paddle.save(
model.state_dict(),
os.path.join(best_model_dir, 'model.pdparams'))
logger.info(
'[EVAL] The model with the best validation sad ({:.4f}) was saved at iter {}.'
.format(best_sad, best_model_iter))
if use_vdl:
log_writer.add_scalar('Evaluate/SAD', sad, iter)
log_writer.add_scalar('Evaluate/MSE', mse, iter)
batch_start = time.time()
# Sleep for half a second to let dataloader release resources.
time.sleep(0.5)
if use_vdl:
log_writer.close()
def predict(model,
model_path,
transforms,
image_list,
image_dir=None,
trimap_list=None,
save_dir='output'):
"""
predict and visualize the image_list.
Args:
model (nn.Layer): Used to predict for input image.
model_path (str): The path of pretrained model.
transforms (transforms.Compose): Preprocess for input image.
image_list (list): A list of image path to be predicted.
image_dir (str, optional): The root directory of the images predicted. Default: None.
trimap_list (list, optional): A list of trimap of image_list. Default: None.
save_dir (str, optional): The directory to save the visualized results. Default: 'output'.
"""
utils.utils.load_entire_model(model, model_path)
model.eval()
nranks = paddle.distributed.get_world_size()
local_rank = paddle.distributed.get_rank()
if nranks > 1:
img_lists = partition_list(image_list, nranks)
trimap_lists = partition_list(
trimap_list, nranks) if trimap_list is not None else None
else:
img_lists = [image_list]
trimap_lists = [trimap_list] if trimap_list is not None else None
logger.info("Start to predict...")
progbar_pred = progbar.Progbar(target=len(img_lists[0]), verbose=1)
preprocess_cost_averager = TimeAverager()
infer_cost_averager = TimeAverager()
postprocess_cost_averager = TimeAverager()
batch_start = time.time()
with paddle.no_grad():
for i, im_path in enumerate(img_lists[local_rank]):
preprocess_start = time.time()
trimap = trimap_lists[local_rank][
i] if trimap_list is not None else None
data = preprocess(img=im_path, transforms=transforms, trimap=trimap)
preprocess_cost_averager.record(time.time() - preprocess_start)
infer_start = time.time()
alpha_pred = model(data)
infer_cost_averager.record(time.time() - infer_start)
postprocess_start = time.time()
alpha_pred = reverse_transform(alpha_pred, data['trans_info'])
alpha_pred = (alpha_pred.numpy()).squeeze()
alpha_pred = (alpha_pred * 255).astype('uint8')
# get the saved name
if image_dir is not None:
im_file = im_path.replace(image_dir, '')
else:
im_file = os.path.basename(im_path)
if im_file[0] == '/' or im_file[0] == '\\':
im_file = im_file[1:]
save_path = os.path.join(save_dir, im_file)
mkdir(save_path)
save_alpha_pred(alpha_pred, save_path, trimap=trimap)
postprocess_cost_averager.record(time.time() - postprocess_start)
preprocess_cost = preprocess_cost_averager.get_average()
infer_cost = infer_cost_averager.get_average()
postprocess_cost = postprocess_cost_averager.get_average()
if local_rank == 0:
progbar_pred.update(i + 1,
[('preprocess_cost', preprocess_cost),
('infer_cost cost', infer_cost),
('postprocess_cost', postprocess_cost)])
preprocess_cost_averager.reset()
infer_cost_averager.reset()
postprocess_cost_averager.reset()
return alpha_pred
def evaluate(model,
eval_dataset,
aug_eval=False,
scales=1.0,
flip_horizontal=True,
flip_vertical=False,
is_slide=False,
stride=None,
crop_size=None,
num_workers=0,
print_detail=True):
"""
Launch evalution.
Args:
model（nn.Layer): A sementic segmentation model.
eval_dataset (paddle.io.Dataset): Used to read and process validation datasets.
aug_eval (bool, optional): Whether to use mulit-scales and flip augment for evaluation. Default: False.
scales (list|float, optional): Scales for augment. It is valid when `aug_eval` is True. Default: 1.0.
flip_horizontal (bool, optional): Whether to use flip horizontally augment. It is valid when `aug_eval` is True. Default: True.
flip_vertical (bool, optional): Whether to use flip vertically augment. It is valid when `aug_eval` is True. Default: False.
is_slide (bool, optional): Whether to evaluate by sliding window. Default: False.
stride (tuple|list, optional): The stride of sliding window, the first is width and the second is height.
It should be provided when `is_slide` is True.
crop_size (tuple|list, optional): The crop size of sliding window, the first is width and the second is height.
It should be provided when `is_slide` is True.
num_workers (int, optional): Num workers for data loader. Default: 0.
print_detail (bool, optional): Whether to print detailed information about the evaluation process. Default: True.
Returns:
float: The mIoU of validation datasets.
float: The accuracy of validation datasets.
"""
model.eval()
nranks = paddle.distributed.ParallelEnv().nranks
local_rank = paddle.distributed.ParallelEnv().local_rank
if nranks > 1:
# Initialize parallel environment if not done.
if not paddle.distributed.parallel.parallel_helper._is_parallel_ctx_initialized(
):
paddle.distributed.init_parallel_env()
batch_sampler = paddle.io.DistributedBatchSampler(eval_dataset,
batch_size=1,
shuffle=False,
drop_last=False)
loader = paddle.io.DataLoader(
eval_dataset,
batch_sampler=batch_sampler,
num_workers=num_workers,
return_list=True,
)
total_iters = len(loader)
intersect_area_all = 0
pred_area_all = 0
label_area_all = 0
if print_detail:
logger.info(
"Start evaluating (total_samples={}, total_iters={})...".format(
len(eval_dataset), total_iters))
progbar_val = progbar.Progbar(target=total_iters, verbose=1)
reader_cost_averager = TimeAverager()
batch_cost_averager = TimeAverager()
batch_start = time.time()
with paddle.no_grad():
for iter, (im, label) in enumerate(loader):
reader_cost_averager.record(time.time() - batch_start)
label = label.astype('int64')
ori_shape = label.shape[-2:]
if aug_eval:
pred = infer.aug_inference(
model,
im,
ori_shape=ori_shape,
transforms=eval_dataset.transforms.transforms,
scales=scales,
flip_horizontal=flip_horizontal,
flip_vertical=flip_vertical,
is_slide=is_slide,
stride=stride,
crop_size=crop_size)
else:
pred = infer.inference(
model,
im,
ori_shape=ori_shape,
transforms=eval_dataset.transforms.transforms,
is_slide=is_slide,
stride=stride,
crop_size=crop_size)
intersect_area, pred_area, label_area = metrics.calculate_area(
pred,
label,
eval_dataset.num_classes,
ignore_index=eval_dataset.ignore_index)
# Gather from all ranks
if nranks > 1:
intersect_area_list = []
pred_area_list = []
label_area_list = []
paddle.distributed.all_gather(intersect_area_list,
intersect_area)
paddle.distributed.all_gather(pred_area_list, pred_area)
paddle.distributed.all_gather(label_area_list, label_area)
# Some image has been evaluated and should be eliminated in last iter
if (iter + 1) * nranks > len(eval_dataset):
valid = len(eval_dataset) - iter * nranks
intersect_area_list = intersect_area_list[:valid]
pred_area_list = pred_area_list[:valid]
label_area_list = label_area_list[:valid]
for i in range(len(intersect_area_list)):
intersect_area_all = intersect_area_all + intersect_area_list[
i]
pred_area_all = pred_area_all + pred_area_list[i]
label_area_all = label_area_all + label_area_list[i]
else:
intersect_area_all = intersect_area_all + intersect_area
pred_area_all = pred_area_all + pred_area
label_area_all = label_area_all + label_area
batch_cost_averager.record(time.time() - batch_start,
num_samples=len(label))
batch_cost = batch_cost_averager.get_average()
reader_cost = reader_cost_averager.get_average()
if local_rank == 0 and print_detail:
progbar_val.update(iter + 1, [('batch_cost', batch_cost),
('reader cost', reader_cost)])
reader_cost_averager.reset()
batch_cost_averager.reset()
batch_start = time.time()
class_iou, miou = metrics.mean_iou(intersect_area_all, pred_area_all,
label_area_all)
class_acc, acc = metrics.accuracy(intersect_area_all, pred_area_all)
kappa = metrics.kappa(intersect_area_all, pred_area_all, label_area_all)
if print_detail:
logger.info(
"[EVAL] #Images={} mIoU={:.4f} Acc={:.4f} Kappa={:.4f} ".format(
len(eval_dataset), miou, acc, kappa))
logger.info("[EVAL] Class IoU: \n" + str(np.round(class_iou, 4)))
logger.info("[EVAL] Class Acc: \n" + str(np.round(class_acc, 4)))
return miou, acc
def train(model,
train_dataset,
val_dataset=None,
optimizer=None,
save_dir='output',
iters=10000,
batch_size=2,
resume_model=None,
save_interval=1000,
log_iters=10,
num_workers=0,
use_vdl=False,
losses=None,
keep_checkpoint_max=5,
test_config=None,
fp16=False,
profiler_options=None):
"""
Launch training.
Args:
model（nn.Layer): A sementic segmentation model.
train_dataset (paddle.io.Dataset): Used to read and process training datasets.
val_dataset (paddle.io.Dataset, optional): Used to read and process validation datasets.
optimizer (paddle.optimizer.Optimizer): The optimizer.
save_dir (str, optional): The directory for saving the model snapshot. Default: 'output'.
iters (int, optional): How may iters to train the model. Defualt: 10000.
batch_size (int, optional): Mini batch size of one gpu or cpu. Default: 2.
resume_model (str, optional): The path of resume model.
save_interval (int, optional): How many iters to save a model snapshot once during training. Default: 1000.
log_iters (int, optional): Display logging information at every log_iters. Default: 10.
num_workers (int, optional): Num workers for data loader. Default: 0.
use_vdl (bool, optional): Whether to record the data to VisualDL during training. Default: False.
losses (dict, optional): A dict including 'types' and 'coef'. The length of coef should equal to 1 or len(losses['types']).
The 'types' item is a list of object of paddleseg.models.losses while the 'coef' item is a list of the relevant coefficient.
keep_checkpoint_max (int, optional): Maximum number of checkpoints to save. Default: 5.
test_config(dict, optional): Evaluation config.
fp16 (bool, optional): Whether to use amp.
profiler_options (str, optional): The option of train profiler.
"""
model.train()
nranks = paddle.distributed.ParallelEnv().nranks
local_rank = paddle.distributed.ParallelEnv().local_rank
start_iter = 0
if resume_model is not None:
start_iter = resume(model, optimizer, resume_model)
if not os.path.isdir(save_dir):
if os.path.exists(save_dir):
os.remove(save_dir)
os.makedirs(save_dir)
if nranks > 1:
paddle.distributed.fleet.init(is_collective=True)
optimizer = paddle.distributed.fleet.distributed_optimizer(
optimizer) # The return is Fleet object
ddp_model = paddle.distributed.fleet.distributed_model(model)
batch_sampler = paddle.io.DistributedBatchSampler(train_dataset,
batch_size=batch_size,
shuffle=True,
drop_last=True)
loader = paddle.io.DataLoader(
train_dataset,
batch_sampler=batch_sampler,
num_workers=num_workers,
return_list=True,
worker_init_fn=worker_init_fn,
)
# use amp
if fp16:
logger.info('use amp to train')
scaler = paddle.amp.GradScaler(init_loss_scaling=1024)
if use_vdl:
from visualdl import LogWriter
log_writer = LogWriter(save_dir)
avg_loss = 0.0
avg_loss_list = []
iters_per_epoch = len(batch_sampler)
best_acc = -1.0
best_model_iter = -1
reader_cost_averager = TimeAverager()
batch_cost_averager = TimeAverager()
save_models = deque()
batch_start = time.time()
iter = start_iter
while iter < iters:
for data in loader:
iter += 1
if iter > iters:
version = paddle.__version__
if version == '2.1.2':
continue
else:
break
reader_cost_averager.record(time.time() - batch_start)
images = data[0]
labels = data[1].astype('int64')
edges = None
if len(data) == 3:
edges = data[2].astype('int64')
if hasattr(model, 'data_format') and model.data_format == 'NHWC':
images = images.transpose((0, 2, 3, 1))
if fp16:
with paddle.amp.auto_cast(
enable=True,
custom_white_list={
"elementwise_add", "batch_norm", "sync_batch_norm"
},
custom_black_list={'bilinear_interp_v2'}):
if nranks > 1:
logits_list = ddp_model(images)
else:
logits_list = model(images)
loss_list = loss_computation(logits_list=logits_list,
labels=labels,
losses=losses,
edges=edges)
loss = sum(loss_list)
scaled = scaler.scale(loss) # scale the loss
scaled.backward() # do backward
if isinstance(optimizer, paddle.distributed.fleet.Fleet):
scaler.minimize(optimizer.user_defined_optimizer, scaled)
else:
scaler.minimize(optimizer, scaled) # update parameters
else:
if nranks > 1:
logits_list = ddp_model(images)
else:
logits_list = model(images)
loss_list = loss_computation(logits_list=logits_list,
labels=labels,
losses=losses,
edges=edges)
loss = sum(loss_list)
loss.backward()
optimizer.step()
lr = optimizer.get_lr()
# update lr
if isinstance(optimizer, paddle.distributed.fleet.Fleet):
lr_sche = optimizer.user_defined_optimizer._learning_rate
else:
lr_sche = optimizer._learning_rate
if isinstance(lr_sche, paddle.optimizer.lr.LRScheduler):
lr_sche.step()
train_profiler.add_profiler_step(profiler_options)
model.clear_gradients()
avg_loss += loss.numpy()[0]
if not avg_loss_list:
avg_loss_list = [l.numpy() for l in loss_list]
else:
for i in range(len(loss_list)):
avg_loss_list[i] += loss_list[i].numpy()
batch_cost_averager.record(time.time() - batch_start,
num_samples=batch_size)
if (iter) % log_iters == 0 and local_rank == 0:
avg_loss /= log_iters
avg_loss_list = [l[0] / log_iters for l in avg_loss_list]
remain_iters = iters - iter
avg_train_batch_cost = batch_cost_averager.get_average()
avg_train_reader_cost = reader_cost_averager.get_average()
eta = calculate_eta(remain_iters, avg_train_batch_cost)
logger.info(
"[TRAIN] epoch: {}, iter: {}/{}, loss: {:.4f}, lr: {:.6f}, batch_cost: {:.4f}, reader_cost: {:.5f}, ips: {:.4f} samples/sec | ETA {}"
.format((iter - 1) // iters_per_epoch + 1, iter, iters,
avg_loss, lr, avg_train_batch_cost,
avg_train_reader_cost,
batch_cost_averager.get_ips_average(), eta))
if use_vdl:
log_writer.add_scalar('Train/loss', avg_loss, iter)
# Record all losses if there are more than 2 losses.
if len(avg_loss_list) > 1:
avg_loss_dict = {}
for i, value in enumerate(avg_loss_list):
avg_loss_dict['loss_' + str(i)] = value
for key, value in avg_loss_dict.items():
log_tag = 'Train/' + key
log_writer.add_scalar(log_tag, value, iter)
log_writer.add_scalar('Train/lr', lr, iter)
log_writer.add_scalar('Train/batch_cost',
avg_train_batch_cost, iter)
log_writer.add_scalar('Train/reader_cost',
avg_train_reader_cost, iter)
avg_loss = 0.0
avg_loss_list = []
reader_cost_averager.reset()
batch_cost_averager.reset()
if (iter % save_interval == 0 or iter == iters) and (val_dataset
is not None):
num_workers = 1 if num_workers > 0 else 0
if test_config is None:
test_config = {}
acc, fp, fn = evaluate(model,
val_dataset,
num_workers=num_workers,
save_dir=save_dir,
**test_config)
model.train()
if (iter % save_interval == 0
or iter == iters) and local_rank == 0:
current_save_dir = os.path.join(save_dir,
"iter_{}".format(iter))
if not os.path.isdir(current_save_dir):
os.makedirs(current_save_dir)
paddle.save(model.state_dict(),
os.path.join(current_save_dir, 'model.pdparams'))
paddle.save(optimizer.state_dict(),
os.path.join(current_save_dir, 'model.pdopt'))
save_models.append(current_save_dir)
if len(save_models) > keep_checkpoint_max > 0:
model_to_remove = save_models.popleft()
shutil.rmtree(model_to_remove)
if val_dataset is not None:
if acc > best_acc:
best_acc = acc
best_model_iter = iter
best_model_dir = os.path.join(save_dir, "best_model")
paddle.save(
model.state_dict(),
os.path.join(best_model_dir, 'model.pdparams'))
logger.info(
'[EVAL] The model with the best validation Acc ({:.4f}) was saved at iter {}.'
.format(best_acc, best_model_iter))
if use_vdl:
log_writer.add_scalar('Evaluate/Acc', acc, iter)
log_writer.add_scalar('Evaluate/Fp', fp, iter)
log_writer.add_scalar('Evaluate/Fn', fn, iter)
batch_start = time.time()
# Calculate flops.
if local_rank == 0:
_, c, h, w = images.shape
_ = paddle.flops(
model, [1, c, h, w],
custom_ops={paddle.nn.SyncBatchNorm: op_flops_funs.count_syncbn})
# Sleep for half a second to let dataloader release resources.
time.sleep(0.5)
if use_vdl:
log_writer.close()
def train(model,
train_dataset,
val_dataset=None,
optimizer=None,
save_dir='output',
iters=10000,
batch_size=2,
resume_model=None,
save_interval=1000,
log_iters=10,
num_workers=0,
use_vdl=False,
losses=None,
keep_checkpoint_max=5,
threshold=0.1,
nms_kernel=7,
top_k=200):
"""
Launch training.
Args:
model（nn.Layer): A sementic segmentation model.
train_dataset (paddle.io.Dataset): Used to read and process training datasets.
val_dataset (paddle.io.Dataset, optional): Used to read and process validation datasets.
optimizer (paddle.optimizer.Optimizer): The optimizer.
save_dir (str, optional): The directory for saving the model snapshot. Default: 'output'.
iters (int, optional): How may iters to train the model. Defualt: 10000.
batch_size (int, optional): Mini batch size of one gpu or cpu. Default: 2.
resume_model (str, optional): The path of resume model.
save_interval (int, optional): How many iters to save a model snapshot once during training. Default: 1000.
log_iters (int, optional): Display logging information at every log_iters. Default: 10.
num_workers (int, optional): Num workers for data loader. Default: 0.
use_vdl (bool, optional): Whether to record the data to VisualDL during training. Default: False.
losses (dict): A dict including 'types' and 'coef'. The length of coef should equal to 1 or len(losses['types']).
The 'types' item is a list of object of paddleseg.models.losses while the 'coef' item is a list of the relevant coefficient.
keep_checkpoint_max (int, optional): Maximum number of checkpoints to save. Default: 5.
threshold (float, optional): A Float, threshold applied to center heatmap score. Default: 0.1.
nms_kernel (int, optional): An Integer, NMS max pooling kernel size. Default: 7.
top_k (int, optional): An Integer, top k centers to keep. Default: 200.
"""
model.train()
nranks = paddle.distributed.ParallelEnv().nranks
local_rank = paddle.distributed.ParallelEnv().local_rank
start_iter = 0
if resume_model is not None:
start_iter = resume(model, optimizer, resume_model)
if not os.path.isdir(save_dir):
if os.path.exists(save_dir):
os.remove(save_dir)
os.makedirs(save_dir)
if nranks > 1:
# Initialize parallel environment if not done.
if not paddle.distributed.parallel.parallel_helper._is_parallel_ctx_initialized(
):
paddle.distributed.init_parallel_env()
ddp_model = paddle.DataParallel(model)
else:
ddp_model = paddle.DataParallel(model)
batch_sampler = paddle.io.DistributedBatchSampler(train_dataset,
batch_size=batch_size,
shuffle=True,
drop_last=True)
loader = paddle.io.DataLoader(
train_dataset,
batch_sampler=batch_sampler,
num_workers=num_workers,
return_list=True,
)
if use_vdl:
from visualdl import LogWriter
log_writer = LogWriter(save_dir)
avg_loss = 0.0
avg_loss_list = []
iters_per_epoch = len(batch_sampler)
best_pq = -1.0
best_model_iter = -1
reader_cost_averager = TimeAverager()
batch_cost_averager = TimeAverager()
save_models = deque()
batch_start = time.time()
iter = start_iter
while iter < iters:
for data in loader:
iter += 1
if iter > iters:
break
reader_cost_averager.record(time.time() - batch_start)
images = data[0]
semantic = data[1]
semantic_weights = data[2]
center = data[3]
center_weights = data[4]
offset = data[5]
offset_weights = data[6]
foreground = data[7]
if nranks > 1:
logits_list = ddp_model(images)
else:
logits_list = model(images)
loss_list = loss_computation(logits_list=logits_list,
losses=losses,
semantic=semantic,
semantic_weights=semantic_weights,
center=center,
center_weights=center_weights,
offset=offset,
offset_weights=offset_weights)
loss = sum(loss_list)
loss.backward()
optimizer.step()
lr = optimizer.get_lr()
if isinstance(optimizer._learning_rate,
paddle.optimizer.lr.LRScheduler):
optimizer._learning_rate.step()
model.clear_gradients()
avg_loss += loss.numpy()[0]
if not avg_loss_list:
avg_loss_list = [l.numpy() for l in loss_list]
else:
for i in range(len(loss_list)):
avg_loss_list[i] += loss_list[i].numpy()
batch_cost_averager.record(time.time() - batch_start,
num_samples=batch_size)
if (iter) % log_iters == 0 and local_rank == 0:
avg_loss /= log_iters
avg_loss_list = [l[0] / log_iters for l in avg_loss_list]
remain_iters = iters - iter
avg_train_batch_cost = batch_cost_averager.get_average()
avg_train_reader_cost = reader_cost_averager.get_average()
eta = calculate_eta(remain_iters, avg_train_batch_cost)
logger.info(
"[TRAIN] epoch={}, iter={}/{}, loss={:.4f}, lr={:.6f}, batch_cost={:.4f}, reader_cost={:.5f}, ips={:.4f} samples/sec | ETA {}"
.format((iter - 1) // iters_per_epoch + 1, iter, iters,
avg_loss, lr, avg_train_batch_cost,
avg_train_reader_cost,
batch_cost_averager.get_ips_average(), eta))
logger.info(
"[LOSS] loss={:.4f}, semantic_loss={:.4f}, center_loss={:.4f}, offset_loss={:.4f}"
.format(avg_loss, avg_loss_list[0], avg_loss_list[1],
avg_loss_list[2]))
if use_vdl:
log_writer.add_scalar('Train/loss', avg_loss, iter)
# Record all losses if there are more than 2 losses.
if len(avg_loss_list) > 1:
avg_loss_dict = {}
for i, value in enumerate(avg_loss_list):
avg_loss_dict['loss_' + str(i)] = value
for key, value in avg_loss_dict.items():
log_tag = 'Train/' + key
log_writer.add_scalar(log_tag, value, iter)
log_writer.add_scalar('Train/lr', lr, iter)
log_writer.add_scalar('Train/batch_cost',
avg_train_batch_cost, iter)
log_writer.add_scalar('Train/reader_cost',
avg_train_reader_cost, iter)
avg_loss = 0.0
avg_loss_list = []
reader_cost_averager.reset()
batch_cost_averager.reset()
# save model
if (iter % save_interval == 0
or iter == iters) and local_rank == 0:
current_save_dir = os.path.join(save_dir,
"iter_{}".format(iter))
if not os.path.isdir(current_save_dir):
os.makedirs(current_save_dir)
paddle.save(model.state_dict(),
os.path.join(current_save_dir, 'model.pdparams'))
paddle.save(optimizer.state_dict(),
os.path.join(current_save_dir, 'model.pdopt'))
save_models.append(current_save_dir)
if len(save_models) > keep_checkpoint_max > 0:
model_to_remove = save_models.popleft()
shutil.rmtree(model_to_remove)
# eval model
if (iter % save_interval == 0 or iter == iters) and (
val_dataset
is not None) and local_rank == 0 and iter > iters // 2:
num_workers = 1 if num_workers > 0 else 0
panoptic_results, semantic_results, instance_results = evaluate(
model,
val_dataset,
threshold=threshold,
nms_kernel=nms_kernel,
top_k=top_k,
num_workers=num_workers,
print_detail=False)
pq = panoptic_results['pan_seg']['All']['pq']
miou = semantic_results['sem_seg']['mIoU']
map = instance_results['ins_seg']['mAP']
map50 = instance_results['ins_seg']['mAP50']
logger.info(
"[EVAL] PQ: {:.4f}, mIoU: {:.4f}, mAP: {:.4f}, mAP50: {:.4f}"
.format(pq, miou, map, map50))
model.train()
# save best model and add evaluate results to vdl
if (iter % save_interval == 0
or iter == iters) and local_rank == 0:
if val_dataset is not None and iter > iters // 2:
if pq > best_pq:
best_pq = pq
best_model_iter = iter
best_model_dir = os.path.join(save_dir, "best_model")
paddle.save(
model.state_dict(),
os.path.join(best_model_dir, 'model.pdparams'))
logger.info(
'[EVAL] The model with the best validation pq ({:.4f}) was saved at iter {}.'
.format(best_pq, best_model_iter))
if use_vdl:
log_writer.add_scalar('Evaluate/PQ', pq, iter)
log_writer.add_scalar('Evaluate/mIoU', miou, iter)
log_writer.add_scalar('Evaluate/mAP', map, iter)
log_writer.add_scalar('Evaluate/mAP50', map50, iter)
batch_start = time.time()
# Calculate flops.
if local_rank == 0:
def count_syncbn(m, x, y):
x = x[0]
nelements = x.numel()
m.total_ops += int(2 * nelements)
_, c, h, w = images.shape
flops = paddle.flops(
model, [1, c, h, w],
custom_ops={paddle.nn.SyncBatchNorm: count_syncbn})
# Sleep for half a second to let dataloader release resources.
time.sleep(0.5)
if use_vdl:
log_writer.close()
def evaluate(model,
eval_dataset,
num_workers=0,
print_detail=True,
save_img=True):
"""
Launch evalution.
Args:
model（nn.Layer): A sementic segmentation model.
eval_dataset (paddle.io.Dataset): Used to read and process validation datasets.
num_workers (int, optional): Num workers for data loader. Default: 0.
print_detail (bool, optional): Whether to print detailed information about the evaluation process. Default: True.
Returns:
float: The mIoU of validation datasets.
float: The accuracy of validation datasets.
"""
logger.info('Validating')
evaluator = Eval(eval_dataset.NUM_CLASSES)
evaluator.reset()
model.eval()
nranks = paddle.distributed.ParallelEnv().nranks
local_rank = paddle.distributed.ParallelEnv().local_rank
if nranks > 1:
# Initialize parallel environment if not done.
if not paddle.distributed.parallel.parallel_helper._is_parallel_ctx_initialized(
):
paddle.distributed.init_parallel_env()
batch_sampler = paddle.io.DistributedBatchSampler(
eval_dataset, batch_size=1, shuffle=False, drop_last=True)
loader = paddle.io.DataLoader(
eval_dataset,
batch_sampler=batch_sampler,
num_workers=num_workers,
return_list=True,
)
progbar_val = progbar.Progbar(
target=len(loader), verbose=0 if nranks < 2 else 2)
reader_cost_averager = TimeAverager()
batch_cost_averager = TimeAverager()
batch_start = time.time()
with paddle.no_grad():
for idx, (x, y, _, item) in enumerate(loader):
reader_cost_averager.record(time.time() - batch_start)
# Forward
y = y.astype('int64')
pred = model(x) # 1, c, h, w
if len(pred) > 1:
pred = pred[0]
# Convert to numpy
label = y.squeeze(axis=1).numpy() #
argpred = np.argmax(pred.numpy(), axis=1) # 1, 1, H, W
if save_img:
save_imgs(argpred, item, './output/')
# Add to evaluator
evaluator.add_batch(label, argpred)
batch_cost_averager.record(
time.time() - batch_start, num_samples=len(label))
batch_cost = batch_cost_averager.get_average()
reader_cost = reader_cost_averager.get_average()
if local_rank == 0 and print_detail and idx % 10 == 0:
progbar_val.update(idx + 1, [('batch_cost', batch_cost),
('reader cost', reader_cost)])
reader_cost_averager.reset()
batch_cost_averager.reset()
batch_start = time.time()
PA = evaluator.pixel_accuracy()
MPA = evaluator.mean_pixel_accuracy()
MIoU = evaluator.mean_iou()
FWIoU = evaluator.fwiou()
PC = evaluator.mean_precision()
logger.info(
'PA1:{:.3f}, MPA1:{:.3f}, MIoU1:{:.3f}, FWIoU1:{:.3f}, PC:{:.3f}'.
format(PA, MPA, MIoU, FWIoU, PC))
return PA, MPA, MIoU, FWIoU
def train(model,
train_dataset,
val_dataset=None,
optimizer=None,
save_dir='output',
iters=10000,
batch_size=2,
resume_model=None,
save_interval=1000,
log_iters=10,
num_workers=0,
use_vdl=False,
losses=None,
keep_checkpoint_max=5):
"""
Launch training.
Args:
model（nn.Layer): A sementic segmentation model.
train_dataset (paddle.io.Dataset): Used to read and process training datasets.
val_dataset (paddle.io.Dataset, optional): Used to read and process validation datasets.
optimizer (paddle.optimizer.Optimizer): The optimizer.
save_dir (str, optional): The directory for saving the model snapshot. Default: 'output'.
iters (int, optional): How may iters to train the model. Defualt: 10000.
batch_size (int, optional): Mini batch size of one gpu or cpu. Default: 2.
resume_model (str, optional): The path of resume model.
save_interval (int, optional): How many iters to save a model snapshot once during training. Default: 1000.
log_iters (int, optional): Display logging information at every log_iters. Default: 10.
num_workers (int, optional): Num workers for data loader. Default: 0.
use_vdl (bool, optional): Whether to record the data to VisualDL during training. Default: False.
losses (dict): A dict including 'types' and 'coef'. The length of coef should equal to 1 or len(losses['types']).
The 'types' item is a list of object of paddleseg.models.losses while the 'coef' item is a list of the relevant coefficient.
keep_checkpoint_max (int, optional): Maximum number of checkpoints to save. Default: 5.
"""
model.train()
nranks = paddle.distributed.ParallelEnv().nranks
local_rank = paddle.distributed.ParallelEnv().local_rank
start_iter = 0
if resume_model is not None:
start_iter = resume(model, optimizer, resume_model)
if not os.path.isdir(save_dir):
if os.path.exists(save_dir):
os.remove(save_dir)
os.makedirs(save_dir)
if nranks > 1:
# Initialize parallel environment if not done.
if not paddle.distributed.parallel.parallel_helper._is_parallel_ctx_initialized(
):
paddle.distributed.init_parallel_env()
ddp_model = paddle.DataParallel(model)
else:
ddp_model = paddle.DataParallel(model)
batch_sampler = paddle.io.DistributedBatchSampler(train_dataset,
batch_size=batch_size,
shuffle=True,
drop_last=True)
loader = paddle.io.DataLoader(
train_dataset,
batch_sampler=batch_sampler,
num_workers=num_workers,
return_list=True,
)
if use_vdl:
from visualdl import LogWriter
log_writer = LogWriter(save_dir)
avg_loss = 0.0
avg_loss_list = []
iters_per_epoch = len(batch_sampler)
best_mean_iou = -1.0
best_model_iter = -1
reader_cost_averager = TimeAverager()
batch_cost_averager = TimeAverager()
save_models = deque()
batch_start = time.time()
iter = start_iter
while iter < iters:
for data in loader:
iter += 1
if iter > iters:
break
reader_cost_averager.record(time.time() - batch_start)
images = data[0]
labels = data[1].astype('int64')
edges = None
if len(data) == 3:
edges = data[2].astype('int64')
if nranks > 1:
logits_list = ddp_model(images)
else:
logits_list = model(images)
loss_list = loss_computation(logits_list=logits_list,
labels=labels,
losses=losses,
edges=edges)
loss = sum(loss_list)
loss.backward()
optimizer.step()
lr = optimizer.get_lr()
if isinstance(optimizer._learning_rate,
paddle.optimizer.lr.LRScheduler):
optimizer._learning_rate.step()
model.clear_gradients()
avg_loss += loss.numpy()[0]
if not avg_loss_list:
avg_loss_list = [l.numpy() for l in loss_list]
else:
for i in range(len(loss_list)):
avg_loss_list[i] += loss_list[i].numpy()
batch_cost_averager.record(time.time() - batch_start,
num_samples=batch_size)
if (iter) % log_iters == 0 and local_rank == 0:
avg_loss /= log_iters
avg_loss_list = [l[0] / log_iters for l in avg_loss_list]
remain_iters = iters - iter
avg_train_batch_cost = batch_cost_averager.get_average()
avg_train_reader_cost = reader_cost_averager.get_average()
eta = calculate_eta(remain_iters, avg_train_batch_cost)
logger.info(
"[TRAIN] epoch: {}, iter: {}/{}, loss: {:.4f}, lr: {:.6f}, batch_cost: {:.4f}, reader_cost: {:.5f}, ips: {:.4f} samples/sec | ETA {}"
.format((iter - 1) // iters_per_epoch + 1, iter, iters,
avg_loss, lr, avg_train_batch_cost,
avg_train_reader_cost,
batch_cost_averager.get_ips_average(), eta))
if use_vdl:
log_writer.add_scalar('Train/loss', avg_loss, iter)
# Record all losses if there are more than 2 losses.
if len(avg_loss_list) > 1:
avg_loss_dict = {}
for i, value in enumerate(avg_loss_list):
avg_loss_dict['loss_' + str(i)] = value
for key, value in avg_loss_dict.items():
log_tag = 'Train/' + key
log_writer.add_scalar(log_tag, value, iter)
log_writer.add_scalar('Train/lr', lr, iter)
log_writer.add_scalar('Train/batch_cost',
avg_train_batch_cost, iter)
log_writer.add_scalar('Train/reader_cost',
avg_train_reader_cost, iter)
avg_loss = 0.0
avg_loss_list = []
reader_cost_averager.reset()
batch_cost_averager.reset()
if (iter % save_interval == 0 or iter == iters) and (val_dataset
is not None):
num_workers = 1 if num_workers > 0 else 0
mean_iou, acc, class_iou, _, _ = evaluate(
model, val_dataset, num_workers=num_workers)
model.train()
if (iter % save_interval == 0
or iter == iters) and local_rank == 0:
current_save_dir = os.path.join(save_dir,
"iter_{}".format(iter))
if not os.path.isdir(current_save_dir):
os.makedirs(current_save_dir)
paddle.save(model.state_dict(),
os.path.join(current_save_dir, 'model.pdparams'))
paddle.save(optimizer.state_dict(),
os.path.join(current_save_dir, 'model.pdopt'))
save_models.append(current_save_dir)
if len(save_models) > keep_checkpoint_max > 0:
model_to_remove = save_models.popleft()
shutil.rmtree(model_to_remove)
if val_dataset is not None:
if mean_iou > best_mean_iou:
best_mean_iou = mean_iou
best_model_iter = iter
best_model_dir = os.path.join(save_dir, "best_model")
paddle.save(
model.state_dict(),
os.path.join(best_model_dir, 'model.pdparams'))
logger.info(
'[EVAL] The model with the best validation mIoU ({:.4f}) was saved at iter {}.'
.format(best_mean_iou, best_model_iter))
if use_vdl:
log_writer.add_scalar('Evaluate/mIoU', mean_iou, iter)
for i, iou in enumerate(class_iou):
log_writer.add_scalar('Evaluate/IoU {}'.format(i),
float(iou), iter)
log_writer.add_scalar('Evaluate/Acc', acc, iter)
batch_start = time.time()
# Calculate flops.
if local_rank == 0:
def count_syncbn(m, x, y):
x = x[0]
nelements = x.numel()
m.total_ops += int(2 * nelements)
_, c, h, w = images.shape
flops = paddle.flops(
model, [1, c, h, w],
custom_ops={paddle.nn.SyncBatchNorm: count_syncbn})
# Sleep for half a second to let dataloader release resources.
time.sleep(0.5)
if use_vdl:
log_writer.close()
def train(self,
train_dataset_src,
train_dataset_tgt,
val_dataset_tgt=None,
val_dataset_src=None,
optimizer=None,
save_dir='output',
iters=10000,
batch_size=2,
resume_model=None,
save_interval=1000,
log_iters=10,
num_workers=0,
use_vdl=False,
keep_checkpoint_max=5,
test_config=None):
"""
Launch training.
Args:
train_dataset (paddle.io.Dataset): Used to read and process training datasets.
val_dataset_tgt (paddle.io.Dataset, optional): Used to read and process validation datasets.
optimizer (paddle.optimizer.Optimizer): The optimizer.
save_dir (str, optional): The directory for saving the model snapshot. Default: 'output'.
iters (int, optional): How may iters to train the model. Defualt: 10000.
batch_size (int, optional): Mini batch size of one gpu or cpu. Default: 2.
resume_model (str, optional): The path of resume model.
save_interval (int, optional): How many iters to save a model snapshot once during training. Default: 1000.
log_iters (int, optional): Display logging information at every log_iters. Default: 10.
num_workers (int, optional): Num workers for data loader. Default: 0.
use_vdl (bool, optional): Whether to record the data to VisualDL during training. Default: False.
keep_checkpoint_max (int, optional): Maximum number of checkpoints to save. Default: 5.
test_config(dict, optional): Evaluation config.
"""
start_iter = 0
self.model.train()
nranks = paddle.distributed.ParallelEnv().nranks
local_rank = paddle.distributed.ParallelEnv().local_rank
if resume_model is not None:
logger.info(resume_model)
start_iter = resume(self.model, optimizer, resume_model)
load_ema_model(self.model, self.resume_ema)
if not os.path.isdir(save_dir):
if os.path.exists(save_dir):
os.remove(save_dir)
os.makedirs(save_dir)
if nranks > 1:
paddle.distributed.fleet.init(is_collective=True)
optimizer = paddle.distributed.fleet.distributed_optimizer(
optimizer) # The return is Fleet object
ddp_model = paddle.distributed.fleet.distributed_model(self.model)
batch_sampler_src = paddle.io.DistributedBatchSampler(
train_dataset_src,
batch_size=batch_size,
shuffle=True,
drop_last=True)
loader_src = paddle.io.DataLoader(
train_dataset_src,
batch_sampler=batch_sampler_src,
num_workers=num_workers,
return_list=True,
worker_init_fn=worker_init_fn,
)
batch_sampler_tgt = paddle.io.DistributedBatchSampler(
train_dataset_tgt,
batch_size=batch_size,
shuffle=True,
drop_last=True)
loader_tgt = paddle.io.DataLoader(
train_dataset_tgt,
batch_sampler=batch_sampler_tgt,
num_workers=num_workers,
return_list=True,
worker_init_fn=worker_init_fn,
)
if use_vdl:
from visualdl import LogWriter
log_writer = LogWriter(save_dir)
iters_per_epoch = len(batch_sampler_tgt)
best_mean_iou = -1.0
best_model_iter = -1
reader_cost_averager = TimeAverager()
batch_cost_averager = TimeAverager()
save_models = deque()
batch_start = time.time()
iter = start_iter
while iter < iters:
for _, (data_src,
data_tgt) in enumerate(zip(loader_src, loader_tgt)):
reader_cost_averager.record(time.time() - batch_start)
loss_dict = {}
#### training #####
images_tgt = data_tgt[0]
labels_tgt = data_tgt[1].astype('int64')
images_src = data_src[0]
labels_src = data_src[1].astype('int64')
edges_src = data_src[2].astype('int64')
edges_tgt = data_tgt[2].astype('int64')
if nranks > 1:
logits_list_src = ddp_model(images_src)
else:
logits_list_src = self.model(images_src)
##### source seg & edge loss ####
loss_src_seg_main = self.celoss(logits_list_src[0], labels_src)
loss_src_seg_aux = 0.1 * self.celoss(logits_list_src[1],
labels_src)
loss_src_seg = loss_src_seg_main + loss_src_seg_aux
loss_dict["source_main"] = loss_src_seg_main.numpy()[0]
loss_dict["source_aux"] = loss_src_seg_aux.numpy()[0]
loss = loss_src_seg
del loss_src_seg, loss_src_seg_aux, loss_src_seg_main
#### generate target pseudo label ####
with paddle.no_grad():
if nranks > 1:
logits_list_tgt = ddp_model(images_tgt)
else:
logits_list_tgt = self.model(images_tgt)
pred_P_1 = F.softmax(logits_list_tgt[0], axis=1)
labels_tgt_psu = paddle.argmax(pred_P_1.detach(), axis=1)
# aux label
pred_P_2 = F.softmax(logits_list_tgt[1], axis=1)
pred_c = (pred_P_1 + pred_P_2) / 2
labels_tgt_psu_aux = paddle.argmax(pred_c.detach(), axis=1)
if self.edgeconstrain:
loss_src_edge = self.bceloss_src(
logits_list_src[2], edges_src) # 1, 2 640, 1280
src_edge = paddle.argmax(
logits_list_src[2].detach().clone(),
axis=1) # 1, 1, 640,1280
src_edge_acc = ((src_edge == edges_src).numpy().sum().astype('float32')\
/functools.reduce(lambda a, b: a * b, src_edge.shape))*100
if (not self.src_only) and (iter > 200000):
#### target seg & edge loss ####
logger.info("Add target edege loss")
edges_tgt = Func.mask_to_binary_edge(
labels_tgt_psu.detach().clone().numpy(),
radius=2,
num_classes=train_dataset_tgt.NUM_CLASSES)
edges_tgt = paddle.to_tensor(edges_tgt, dtype='int64')
loss_tgt_edge = self.bceloss_tgt(
logits_list_tgt[2], edges_tgt)
loss_edge = loss_tgt_edge + loss_src_edge
else:
loss_tgt_edge = paddle.zeros([1])
loss_edge = loss_src_edge
loss += loss_edge
loss_dict['target_edge'] = loss_tgt_edge.numpy()[0]
loss_dict['source_edge'] = loss_src_edge.numpy()[0]
del loss_edge, loss_tgt_edge, loss_src_edge
#### target aug loss #######
augs = augmentation.get_augmentation()
images_tgt_aug, labels_tgt_aug = augmentation.augment(
images=images_tgt.cpu(),
labels=labels_tgt_psu.detach().cpu(),
aug=augs,
iters="{}_1".format(iter))
images_tgt_aug = images_tgt_aug.cuda()
labels_tgt_aug = labels_tgt_aug.cuda()
_, labels_tgt_aug_aux = augmentation.augment(
images=images_tgt.cpu(),
labels=labels_tgt_psu_aux.detach().cpu(),
aug=augs,
iters="{}_2".format(iter))
labels_tgt_aug_aux = labels_tgt_aug_aux.cuda()
if nranks > 1:
logits_list_tgt_aug = ddp_model(images_tgt_aug)
else:
logits_list_tgt_aug = self.model(images_tgt_aug)
loss_tgt_aug_main = 0.1 * (self.celoss(logits_list_tgt_aug[0],
labels_tgt_aug))
loss_tgt_aug_aux = 0.1 * (0.1 * self.celoss(
logits_list_tgt_aug[1], labels_tgt_aug_aux))
loss_tgt_aug = loss_tgt_aug_aux + loss_tgt_aug_main
loss += loss_tgt_aug
loss_dict['target_aug_main'] = loss_tgt_aug_main.numpy()[0]
loss_dict['target_aug_aux'] = loss_tgt_aug_aux.numpy()[0]
del images_tgt_aug, labels_tgt_aug_aux, images_tgt, \
loss_tgt_aug, loss_tgt_aug_aux, loss_tgt_aug_main
#### edge input seg; src & tgt edge pull in ######
if self.edgepullin:
src_edge_logit = logits_list_src[2]
feat_src = paddle.concat(
[logits_list_src[0], src_edge_logit], axis=1).detach()
out_src = self.model.fusion(feat_src)
loss_src_edge_rec = self.celoss(out_src, labels_src)
tgt_edge_logit = logits_list_tgt_aug[2]
# tgt_edge_logit = paddle.to_tensor(
# Func.mask_to_onehot(edges_tgt.squeeze().numpy(), 2)
# ).unsqueeze(0).astype('float32')
feat_tgt = paddle.concat(
[logits_list_tgt[0], tgt_edge_logit], axis=1).detach()
out_tgt = self.model.fusion(feat_tgt)
loss_tgt_edge_rec = self.celoss(out_tgt, labels_tgt)
loss_edge_rec = loss_tgt_edge_rec + loss_src_edge_rec
loss += loss_edge_rec
loss_dict['src_edge_rec'] = loss_src_edge_rec.numpy()[0]
loss_dict['tgt_edge_rec'] = loss_tgt_edge_rec.numpy()[0]
del loss_tgt_edge_rec, loss_src_edge_rec
#### mask input feature & pullin ######
if self.featurepullin:
# inner-class loss
feat_src = logits_list_src[0]
feat_tgt = logits_list_tgt_aug[0]
center_src_s, center_tgt_s = [], []
total_pixs = logits_list_src[0].shape[2] * \
logits_list_src[0].shape[3]
for i in range(train_dataset_tgt.NUM_CLASSES):
pred = paddle.argmax(
logits_list_src[0].detach().clone(),
axis=1).unsqueeze(0) # 1, 1, 640, 1280
sel_num = paddle.sum((pred == i).astype('float32'))
# ignore tensor that do not have features in this img
if sel_num > 0:
feat_sel_src = paddle.where(
(pred == i).expand_as(feat_src), feat_src,
paddle.zeros(feat_src.shape))
center_src = paddle.mean(feat_sel_src, axis=[
2, 3
]) / (sel_num / total_pixs) # 1, C
self.src_centers[i] = 0.99 * self.src_centers[
i] + (1 - 0.99) * center_src
pred = labels_tgt_aug.unsqueeze(0) # 1, 1, 512, 512
sel_num = paddle.sum((pred == i).astype('float32'))
if sel_num > 0:
feat_sel_tgt = paddle.where(
(pred == i).expand_as(feat_tgt), feat_tgt,
paddle.zeros(feat_tgt.shape))
center_tgt = paddle.mean(feat_sel_tgt, axis=[
2, 3
]) / (sel_num / total_pixs)
self.tgt_centers[i] = 0.99 * self.tgt_centers[
i] + (1 - 0.99) * center_tgt
center_src_s.append(center_src)
center_tgt_s.append(center_tgt)
if iter >= 3000: # average center structure alignment
src_centers = paddle.concat(self.src_centers, axis=0)
tgt_centers = paddle.concat(self.tgt_centers,
axis=0) # 19， 2048
relatmat_src = paddle.matmul(src_centers,
src_centers,
transpose_y=True) # 19，19
relatmat_tgt = paddle.matmul(tgt_centers,
tgt_centers,
transpose_y=True)
loss_intra_relate = self.klloss(relatmat_src, (relatmat_tgt+relatmat_src)/2) \
+ self.klloss(relatmat_tgt, (relatmat_tgt+relatmat_src)/2)
loss_pix_align_src = self.mseloss(
paddle.to_tensor(center_src_s),
paddle.to_tensor(
self.src_centers).detach().clone())
loss_pix_align_tgt = self.mseloss(
paddle.to_tensor(center_tgt_s),
paddle.to_tensor(
self.tgt_centers).detach().clone())
loss_feat_align = loss_pix_align_src + loss_pix_align_tgt + loss_intra_relate
loss += loss_feat_align
loss_dict['loss_pix_align_src'] = \
loss_pix_align_src.numpy()[0]
loss_dict['loss_pix_align_tgt'] = \
loss_pix_align_tgt.numpy()[0]
loss_dict['loss_intra_relate'] = \
loss_intra_relate.numpy()[0]
del loss_pix_align_tgt, loss_pix_align_src, loss_intra_relate,
self.tgt_centers = [
item.detach().clone() for item in self.tgt_centers
]
self.src_centers = [
item.detach().clone() for item in self.src_centers
]
loss.backward()
del loss
loss = sum(loss_dict.values())
optimizer.step()
self.ema.update_params()
with paddle.no_grad():
##### log & save #####
lr = optimizer.get_lr()
# update lr
if isinstance(optimizer, paddle.distributed.fleet.Fleet):
lr_sche = optimizer.user_defined_optimizer._learning_rate
else:
lr_sche = optimizer._learning_rate
if isinstance(lr_sche, paddle.optimizer.lr.LRScheduler):
lr_sche.step()
if self.cfg['save_edge']:
tgt_edge = paddle.argmax(
logits_list_tgt_aug[2].detach().clone(),
axis=1) # 1, 1, 640,1280
src_feed_gt = paddle.argmax(
src_edge_logit.astype('float32'), axis=1)
tgt_feed_gt = paddle.argmax(
tgt_edge_logit.astype('float32'), axis=1)
logger.info('src_feed_gt_{}_{}_{}'.format(
src_feed_gt.shape, src_feed_gt.max(),
src_feed_gt.min()))
logger.info('tgt_feed_gt_{}_{}_{}'.format(
tgt_feed_gt.shape, max(tgt_feed_gt),
min(tgt_feed_gt)))
save_edge(src_feed_gt, 'src_feed_gt_{}'.format(iter))
save_edge(tgt_feed_gt, 'tgt_feed_gt_{}'.format(iter))
save_edge(tgt_edge, 'tgt_pred_{}'.format(iter))
save_edge(src_edge,
'src_pred_{}_{}'.format(iter, src_edge_acc))
save_edge(edges_src, 'src_gt_{}'.format(iter))
save_edge(edges_tgt, 'tgt_gt_{}'.format(iter))
self.model.clear_gradients()
batch_cost_averager.record(time.time() - batch_start,
num_samples=batch_size)
iter += 1
if (iter) % log_iters == 0 and local_rank == 0:
label_tgt_acc = ((labels_tgt == labels_tgt_psu).numpy().sum().astype('float32')\
/functools.reduce(lambda a, b: a * b, labels_tgt_psu.shape))*100
remain_iters = iters - iter
avg_train_batch_cost = batch_cost_averager.get_average(
)
avg_train_reader_cost = reader_cost_averager.get_average(
)
eta = calculate_eta(remain_iters, avg_train_batch_cost)
logger.info(
"[TRAIN] epoch: {}, iter: {}/{}, loss: {:.4f}, tgt_pix_acc: {:.4f}, lr: {:.6f}, batch_cost: {:.4f}, reader_cost: {:.5f}, ips: {:.4f} samples/sec | ETA {}"
.format(
(iter - 1) // iters_per_epoch + 1, iter, iters,
loss, label_tgt_acc, lr, avg_train_batch_cost,
avg_train_reader_cost,
batch_cost_averager.get_ips_average(), eta))
if use_vdl:
log_writer.add_scalar('Train/loss', loss, iter)
# Record all losses if there are more than 2 losses.
if len(loss_dict) > 1:
for name, loss in loss_dict.items():
log_writer.add_scalar(
'Train/loss_' + name, loss, iter)
log_writer.add_scalar('Train/lr', lr, iter)
log_writer.add_scalar('Train/batch_cost',
avg_train_batch_cost, iter)
log_writer.add_scalar('Train/reader_cost',
avg_train_reader_cost, iter)
log_writer.add_scalar('Train/tgt_label_acc',
label_tgt_acc, iter)
reader_cost_averager.reset()
batch_cost_averager.reset()
if (iter % save_interval == 0 or iter
== iters) and (val_dataset_tgt is not None):
num_workers = 4 if num_workers > 0 else 0 # adjust num_worker=4
if test_config is None:
test_config = {}
self.ema.apply_shadow()
self.ema.model.eval()
PA_tgt, _, MIoU_tgt, _ = val.evaluate(
self.model,
val_dataset_tgt,
num_workers=num_workers,
**test_config)
if (iter % (save_interval * 30)) == 0 \
and self.cfg['eval_src']: # add evaluate on src
PA_src, _, MIoU_src, _ = val.evaluate(
self.model,
val_dataset_src,
num_workers=num_workers,
**test_config)
logger.info(
'[EVAL] The source mIoU is ({:.4f}) at iter {}.'
.format(MIoU_src, iter))
self.ema.restore()
self.model.train()
if (iter % save_interval == 0
or iter == iters) and local_rank == 0:
current_save_dir = os.path.join(
save_dir, "iter_{}".format(iter))
if not os.path.isdir(current_save_dir):
os.makedirs(current_save_dir)
paddle.save(
self.model.state_dict(),
os.path.join(current_save_dir, 'model.pdparams'))
paddle.save(
self.ema.shadow,
os.path.join(current_save_dir,
'model_ema.pdparams'))
paddle.save(
optimizer.state_dict(),
os.path.join(current_save_dir, 'model.pdopt'))
save_models.append(current_save_dir)
if len(save_models) > keep_checkpoint_max > 0:
model_to_remove = save_models.popleft()
shutil.rmtree(model_to_remove)
if val_dataset_tgt is not None:
if MIoU_tgt > best_mean_iou:
best_mean_iou = MIoU_tgt
best_model_iter = iter
best_model_dir = os.path.join(
save_dir, "best_model")
paddle.save(
self.model.state_dict(),
os.path.join(best_model_dir,
'model.pdparams'))
logger.info(
'[EVAL] The model with the best validation mIoU ({:.4f}) was saved at iter {}.'
.format(best_mean_iou, best_model_iter))
if use_vdl:
log_writer.add_scalar('Evaluate/mIoU', MIoU_tgt,
iter)
log_writer.add_scalar('Evaluate/PA', PA_tgt, iter)
if self.cfg['eval_src']:
log_writer.add_scalar('Evaluate/mIoU_src',
MIoU_src, iter)
log_writer.add_scalar('Evaluate/PA_src',
PA_src, iter)
batch_start = time.time()
self.ema.update_buffer()
# # Calculate flops.
if local_rank == 0:
def count_syncbn(m, x, y):
x = x[0]
nelements = x.numel()
m.total_ops += int(2 * nelements)
_, c, h, w = images_src.shape
flops = paddle.flops(
self.model, [1, c, h, w],
custom_ops={paddle.nn.SyncBatchNorm: count_syncbn})
# Sleep for half a second to let dataloader release resources.
time.sleep(0.5)
if use_vdl:
log_writer.close()
def evaluate(model,
eval_dataset,
num_workers=0,
is_view=False,
save_dir='output',
print_detail=True):
"""
Launch evalution.
Args:
model（nn.Layer): A sementic segmentation model.
eval_dataset (paddle.io.Dataset): Used to read and process validation datasets.
num_workers (int, optional): Num workers for data loader. Default: 0.
is_view (bool, optional): Whether to visualize results. Default: False.
save_dir (str, optional): The directory to save the json or visualized results. Default: 'output'.
print_detail (bool, optional): Whether to print detailed information about the evaluation process. Default: True.
Returns:
float: The acc of validation datasets.
float: The fp of validation datasets.
float: The fn of validation datasets.
"""
model.eval()
local_rank = paddle.distributed.ParallelEnv().local_rank
loader = paddle.io.DataLoader(
eval_dataset,
batch_size=4,
drop_last=False,
num_workers=num_workers,
return_list=True,
)
postprocessor = tusimple_processor.TusimpleProcessor(
num_classes=eval_dataset.num_classes,
cut_height=eval_dataset.cut_height,
test_gt_json=eval_dataset.test_gt_json,
save_dir=save_dir,
)
total_iters = len(loader)
if print_detail:
logger.info(
"Start evaluating (total_samples: {}, total_iters: {})...".format(
len(eval_dataset), total_iters))
progbar_val = progbar.Progbar(target=total_iters, verbose=1)
reader_cost_averager = TimeAverager()
batch_cost_averager = TimeAverager()
batch_start = time.time()
with paddle.no_grad():
for iter, (im, label, im_path) in enumerate(loader):
reader_cost_averager.record(time.time() - batch_start)
label = label.astype('int64')
ori_shape = None
time_start = time.time()
pred = infer.inference(
model,
im,
ori_shape=ori_shape,
transforms=eval_dataset.transforms.transforms)
time_end = time.time()
postprocessor.dump_data_to_json(pred[1],
im_path,
run_time=time_end - time_start,
is_dump_json=True,
is_view=is_view)
batch_cost_averager.record(time.time() - batch_start,
num_samples=len(label))
batch_cost = batch_cost_averager.get_average()
reader_cost = reader_cost_averager.get_average()
if local_rank == 0 and print_detail:
progbar_val.update(iter + 1, [('batch_cost', batch_cost),
('reader cost', reader_cost)])
reader_cost_averager.reset()
batch_cost_averager.reset()
batch_start = time.time()
acc, fp, fn, eval_result = postprocessor.bench_one_submit(local_rank)
if print_detail:
logger.info(eval_result)
return acc, fp, fn
class Predictor:
def __init__(self, args):
self.cfg = DeployConfig(args.cfg)
self.args = args
self.compose = T.Compose(self.cfg.transforms)
resize_h, resize_w = args.input_shape
self.disflow = cv2.DISOpticalFlow_create(
cv2.DISOPTICAL_FLOW_PRESET_ULTRAFAST)
self.prev_gray = np.zeros((resize_h, resize_w), np.uint8)
self.prev_cfd = np.zeros((resize_h, resize_w), np.float32)
self.is_init = True
pred_cfg = PredictConfig(self.cfg.model, self.cfg.params)
pred_cfg.disable_glog_info()
if self.args.use_gpu:
pred_cfg.enable_use_gpu(100, 0)
self.predictor = create_predictor(pred_cfg)
if self.args.test_speed:
self.cost_averager = TimeAverager()
def preprocess(self, img):
ori_shapes = []
processed_imgs = []
processed_img = self.compose(img)[0]
processed_imgs.append(processed_img)
ori_shapes.append(img.shape)
return processed_imgs, ori_shapes
def run(self, img, bg):
input_names = self.predictor.get_input_names()
input_handle = self.predictor.get_input_handle(input_names[0])
processed_imgs, ori_shapes = self.preprocess(img)
data = np.array(processed_imgs)
input_handle.reshape(data.shape)
input_handle.copy_from_cpu(data)
if self.args.test_speed:
start = time.time()
self.predictor.run()
if self.args.test_speed:
self.cost_averager.record(time.time() - start)
output_names = self.predictor.get_output_names()
output_handle = self.predictor.get_output_handle(output_names[0])
output = output_handle.copy_to_cpu()
return self.postprocess(output, img, ori_shapes[0], bg)
def postprocess(self, pred, img, ori_shape, bg):
if not os.path.exists(self.args.save_dir):
os.makedirs(self.args.save_dir)
resize_w = pred.shape[-1]
resize_h = pred.shape[-2]
if self.args.soft_predict:
if self.args.use_optic_flow:
score_map = pred[:, 1, :, :].squeeze(0)
score_map = 255 * score_map
cur_gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
cur_gray = cv2.resize(cur_gray, (resize_w, resize_h))
optflow_map = optic_flow_process(cur_gray, score_map, self.prev_gray, self.prev_cfd, \
self.disflow, self.is_init)
self.prev_gray = cur_gray.copy()
self.prev_cfd = optflow_map.copy()
self.is_init = False
score_map = np.repeat(optflow_map[:, :, np.newaxis], 3, axis=2)
score_map = np.transpose(score_map, [2, 0, 1])[np.newaxis, ...]
score_map = reverse_transform(paddle.to_tensor(score_map),
ori_shape,
self.cfg.transforms,
mode='bilinear')
alpha = np.transpose(score_map.numpy().squeeze(0),
[1, 2, 0]) / 255
else:
score_map = pred[:, 1, :, :]
score_map = score_map[np.newaxis, ...]
score_map = reverse_transform(paddle.to_tensor(score_map),
ori_shape,
self.cfg.transforms,
mode='bilinear')
alpha = np.transpose(score_map.numpy().squeeze(0), [1, 2, 0])
else:
if pred.ndim == 3:
pred = pred[:, np.newaxis, ...]
result = reverse_transform(paddle.to_tensor(pred, dtype='float32'),
ori_shape,
self.cfg.transforms,
mode='bilinear')
result = np.array(result)
if self.args.add_argmax:
result = np.argmax(result, axis=1)
else:
result = result.squeeze(1)
alpha = np.transpose(result, [1, 2, 0])
# background replace
h, w, _ = img.shape
bg = cv2.resize(bg, (w, h))
if bg.ndim == 2:
bg = bg[..., np.newaxis]
comb = (alpha * img + (1 - alpha) * bg).astype(np.uint8)
return comb
def evaluate(model,
eval_dataset,
threshold=0.1,
nms_kernel=7,
top_k=200,
num_workers=0,
print_detail=True):
"""
Launch evaluation.
Args:
model（nn.Layer): A sementic segmentation model.
eval_dataset (paddle.io.Dataset): Used to read and process validation datasets.
threshold (float, optional): Threshold applied to center heatmap score. Defalut: 0.1.
nms_kernel (int, optional): NMS max pooling kernel size. Default: 7.
top_k (int, optional): Top k centers to keep. Default: 200.
num_workers (int, optional): Num workers for data loader. Default: 0.
print_detail (bool, optional): Whether to print detailed information about the evaluation process. Default: True.
Returns:
dict: Panoptic evaluation results which includes PQ, RQ, SQ for all, each class, Things and stuff.
dict: Semantic evaluation results which includes mIoU, fwIoU, mACC and pACC.
dict: Instance evaluation results which includes mAP and mAP50, and also AP and AP50 for each class.
"""
model.eval()
nranks = paddle.distributed.ParallelEnv().nranks
local_rank = paddle.distributed.ParallelEnv().local_rank
if nranks > 1:
# Initialize parallel environment if not done.
if not paddle.distributed.parallel.parallel_helper._is_parallel_ctx_initialized(
):
paddle.distributed.init_parallel_env()
batch_sampler = paddle.io.DistributedBatchSampler(eval_dataset,
batch_size=1,
shuffle=False,
drop_last=False)
loader = paddle.io.DataLoader(
eval_dataset,
batch_sampler=batch_sampler,
num_workers=num_workers,
return_list=True,
)
total_iters = len(loader)
semantic_metric = SemanticEvaluator(eval_dataset.num_classes,
ignore_index=eval_dataset.ignore_index)
instance_metric_AP50 = InstanceEvaluator(
eval_dataset.num_classes,
overlaps=0.5,
thing_list=eval_dataset.thing_list)
instance_metric_AP = InstanceEvaluator(eval_dataset.num_classes,
overlaps=list(
np.arange(0.5, 1.0, 0.05)),
thing_list=eval_dataset.thing_list)
panoptic_metric = PanopticEvaluator(
num_classes=eval_dataset.num_classes,
thing_list=eval_dataset.thing_list,
ignore_index=eval_dataset.ignore_index,
label_divisor=eval_dataset.label_divisor)
if print_detail:
logger.info(
"Start evaluating (total_samples={}, total_iters={})...".format(
len(eval_dataset), total_iters))
progbar_val = progbar.Progbar(target=total_iters, verbose=1)
reader_cost_averager = TimeAverager()
batch_cost_averager = TimeAverager()
batch_start = time.time()
with paddle.no_grad():
for iter, data in enumerate(loader):
reader_cost_averager.record(time.time() - batch_start)
im = data[0]
raw_semantic_label = data[1] # raw semantic label.
raw_instance_label = data[2]
raw_panoptic_label = data[3]
ori_shape = raw_semantic_label.shape[-2:]
semantic, semantic_softmax, instance, panoptic, ctr_hmp = infer.inference(
model=model,
im=im,
transforms=eval_dataset.transforms.transforms,
thing_list=eval_dataset.thing_list,
label_divisor=eval_dataset.label_divisor,
stuff_area=eval_dataset.stuff_area,
ignore_index=eval_dataset.ignore_index,
threshold=threshold,
nms_kernel=nms_kernel,
top_k=top_k,
ori_shape=ori_shape)
semantic = semantic.squeeze().numpy()
semantic_softmax = semantic_softmax.squeeze().numpy()
instance = instance.squeeze().numpy()
panoptic = panoptic.squeeze().numpy()
ctr_hmp = ctr_hmp.squeeze().numpy()
raw_semantic_label = raw_semantic_label.squeeze().numpy()
raw_instance_label = raw_instance_label.squeeze().numpy()
raw_panoptic_label = raw_panoptic_label.squeeze().numpy()
# update metric for semantic, instance, panoptic
semantic_metric.update(semantic, raw_semantic_label)
gts = instance_metric_AP.convert_gt_map(raw_semantic_label,
raw_instance_label)
# print([i[0] for i in gts])
preds = instance_metric_AP.convert_pred_map(
semantic_softmax, panoptic)
# print([(i[0], i[1]) for i in preds ])
ignore_mask = raw_semantic_label == eval_dataset.ignore_index
instance_metric_AP.update(preds, gts, ignore_mask=ignore_mask)
instance_metric_AP50.update(preds, gts, ignore_mask=ignore_mask)
panoptic_metric.update(panoptic, raw_panoptic_label)
batch_cost_averager.record(time.time() - batch_start,
num_samples=len(im))
batch_cost = batch_cost_averager.get_average()
reader_cost = reader_cost_averager.get_average()
if local_rank == 0:
progbar_val.update(iter + 1, [('batch_cost', batch_cost),
('reader cost', reader_cost)])
reader_cost_averager.reset()
batch_cost_averager.reset()
batch_start = time.time()
semantic_results = semantic_metric.evaluate()
panoptic_results = panoptic_metric.evaluate()
instance_results = OrderedDict()
ins_ap = instance_metric_AP.evaluate()
ins_ap50 = instance_metric_AP50.evaluate()
instance_results['ins_seg'] = OrderedDict()
instance_results['ins_seg']['mAP'] = ins_ap['ins_seg']['mAP']
instance_results['ins_seg']['AP'] = ins_ap['ins_seg']['AP']
instance_results['ins_seg']['mAP50'] = ins_ap50['ins_seg']['mAP']
instance_results['ins_seg']['AP50'] = ins_ap50['ins_seg']['AP']
if print_detail:
logger.info(panoptic_results)
print()
logger.info(semantic_results)
print()
logger.info(instance_results)
print()
pq = panoptic_results['pan_seg']['All']['pq']
miou = semantic_results['sem_seg']['mIoU']
map = instance_results['ins_seg']['mAP']
map50 = instance_results['ins_seg']['mAP50']
logger.info(
"PQ: {:.4f}, mIoU: {:.4f}, mAP: {:.4f}, mAP50: {:.4f}".format(
pq, miou, map, map50))
return panoptic_results, semantic_results, instance_results