def test():
"""The function of eval."""
start_time = time.time()
args = parse_args()
# logger
args.outputs_dir = os.path.join(
args.log_path,
datetime.datetime.now().strftime('%Y-%m-%d_time_%H_%M_%S'))
rank_id = int(os.environ.get('RANK_ID'))
args.logger = get_logger(args.outputs_dir, rank_id)
context.reset_auto_parallel_context()
parallel_mode = ParallelMode.STAND_ALONE
context.set_auto_parallel_context(parallel_mode=parallel_mode,
gradients_mean=True,
device_num=1)
args.logger.info('Creating Network....')
network = YOLOV3DarkNet53(is_training=False)
config = ConfigYOLOV3DarkNet53()
if args.testing_shape:
config.test_img_shape = conver_testing_shape(args)
# convert fusion network to quantization aware network
if config.quantization_aware:
network = quant.convert_quant_network(network,
bn_fold=True,
per_channel=[True, False],
symmetric=[True, False])
args.logger.info(args.pretrained)
if os.path.isfile(args.pretrained):
param_dict = load_checkpoint(args.pretrained)
param_dict_new = {}
for key, values in param_dict.items():
if key.startswith('moments.'):
continue
elif key.startswith('yolo_network.'):
param_dict_new[key[13:]] = values
else:
param_dict_new[key] = values
load_param_into_net(network, param_dict_new)
args.logger.info('load_model {} success'.format(args.pretrained))
else:
args.logger.info('{} not exists or not a pre-trained file'.format(
args.pretrained))
assert FileNotFoundError(
'{} not exists or not a pre-trained file'.format(args.pretrained))
exit(1)
data_root = args.data_root
ann_file = args.annFile
ds, data_size = create_yolo_dataset(data_root,
ann_file,
is_training=False,
batch_size=args.per_batch_size,
max_epoch=1,
device_num=1,
rank=rank_id,
shuffle=False,
config=config)
args.logger.info('testing shape : {}'.format(config.test_img_shape))
args.logger.info('totol {} images to eval'.format(data_size))
network.set_train(False)
# init detection engine
detection = DetectionEngine(args)
input_shape = Tensor(tuple(config.test_img_shape), ms.float32)
args.logger.info('Start inference....')
for i, data in enumerate(ds.create_dict_iterator()):
image = data["image"]
image_shape = data["image_shape"]
image_id = data["img_id"]
prediction = network(image, input_shape)
output_big, output_me, output_small = prediction
output_big = output_big.asnumpy()
output_me = output_me.asnumpy()
output_small = output_small.asnumpy()
image_id = image_id.asnumpy()
image_shape = image_shape.asnumpy()
detection.detect([output_small, output_me, output_big],
args.per_batch_size, image_shape, image_id)
if i % 1000 == 0:
args.logger.info('Processing... {:.2f}% '.format(
i * args.per_batch_size / data_size * 100))
args.logger.info('Calculating mAP...')
detection.do_nms_for_results()
result_file_path = detection.write_result()
args.logger.info('result file path: {}'.format(result_file_path))
eval_result = detection.get_eval_result()
cost_time = time.time() - start_time
args.logger.info('\n=============coco eval reulst=========\n' +
eval_result)
args.logger.info('testing cost time {:.2f}h'.format(cost_time / 3600.))
def test():
"""The function of eval."""
args = parse_args()
devid = int(os.getenv('DEVICE_ID')) if os.getenv('DEVICE_ID') else 0
context.set_context(mode=context.GRAPH_MODE,
device_target='Ascend',
save_graphs=True,
device_id=devid)
# logger
args.outputs_dir = os.path.join(
args.log_path,
datetime.datetime.now().strftime('%Y-%m-%d_time_%H_%M_%S'))
rank_id = int(
os.environ.get('RANK_ID')) if os.environ.get('RANK_ID') else 0
args.logger = get_logger(args.outputs_dir, rank_id)
context.reset_auto_parallel_context()
parallel_mode = ParallelMode.STAND_ALONE
context.set_auto_parallel_context(parallel_mode=parallel_mode,
gradients_mean=True,
device_num=1)
args.logger.info('Creating Network....')
network = SolveOutput(YOLOV3DarkNet53(is_training=False))
data_root = args.data_root
ann_file = args.annFile
args.logger.info(args.pretrained)
if os.path.isfile(args.pretrained):
param_dict = load_checkpoint(args.pretrained)
param_dict_new = {}
for key, values in param_dict.items():
if key.startswith('moments.'):
continue
elif key.startswith('yolo_network.'):
param_dict_new[key[13:]] = values
else:
param_dict_new[key] = values
load_param_into_net(network, param_dict_new)
args.logger.info('load_model {} success'.format(args.pretrained))
else:
args.logger.info('{} not exists or not a pre-trained file'.format(
args.pretrained))
assert FileNotFoundError(
'{} not exists or not a pre-trained file'.format(args.pretrained))
exit(1)
config = ConfigYOLOV3DarkNet53()
if args.testing_shape:
config.test_img_shape = conver_testing_shape(args)
ds, data_size = create_yolo_dataset(data_root,
ann_file,
is_training=False,
batch_size=1,
max_epoch=1,
device_num=1,
rank=rank_id,
shuffle=False,
config=config)
args.logger.info('testing shape : {}'.format(config.test_img_shape))
args.logger.info('totol {} images to eval'.format(data_size))
network.set_train(False)
# build attacker
attack = DeepFool(network,
num_classes=80,
model_type='detection',
reserve_ratio=0.9,
bounds=(0, 1))
input_shape = Tensor(tuple(config.test_img_shape), ms.float32)
args.logger.info('Start inference....')
batch_num = args.samples_num
adv_example = []
for i, data in enumerate(ds.create_dict_iterator(num_epochs=1)):
if i >= batch_num:
break
image = data["image"]
image_shape = data["image_shape"]
gt_boxes, gt_logits = network(image, input_shape)
gt_boxes, gt_logits = gt_boxes.asnumpy(), gt_logits.asnumpy()
gt_labels = np.argmax(gt_logits, axis=2)
adv_img = attack.generate((image.asnumpy(), image_shape.asnumpy()),
(gt_boxes, gt_labels))
adv_example.append(adv_img)
np.save('adv_example.npy', adv_example)
network = YoloWithLossCell(network)
args.logger.info('finish get network')
config.label_smooth = args.label_smooth
config.label_smooth_factor = args.label_smooth_factor
if args.training_shape:
config.multi_scale = [convert_training_shape(args.training_shape)]
if args.resize_rate:
config.resize_rate = args.resize_rate
ds, data_size = create_yolo_dataset(image_dir=args.data_root,
anno_path=args.annFile,
is_training=True,
batch_size=args.per_batch_size,
max_epoch=args.max_epoch,
device_num=args.group_size,
rank=args.rank,
config=config)
args.logger.info('Finish loading dataset')
args.steps_per_epoch = int(data_size / args.per_batch_size /
args.group_size)
if not args.ckpt_interval:
args.ckpt_interval = args.steps_per_epoch
lr = get_lr(args)
opt = Momentum(params=get_param_groups(network),
learning_rate=Tensor(lr),
def train():
"""Train function."""
args = parse_args()
devid = int(os.getenv('DEVICE_ID', '0'))
context.set_context(mode=context.GRAPH_MODE, enable_auto_mixed_precision=True,
device_target=args.device_target, save_graphs=True, device_id=devid)
if args.need_profiler:
from mindspore.profiler.profiling import Profiler
profiler = Profiler(output_path=args.outputs_dir, is_detail=True, is_show_op_path=True)
loss_meter = AverageMeter('loss')
context.reset_auto_parallel_context()
parallel_mode = ParallelMode.STAND_ALONE
degree = 1
if args.is_distributed:
parallel_mode = ParallelMode.DATA_PARALLEL
degree = get_group_size()
context.set_auto_parallel_context(parallel_mode=parallel_mode, gradients_mean=True, device_num=degree)
network = YOLOV3DarkNet53(is_training=True)
# default is kaiming-normal
default_recurisive_init(network)
load_yolov3_params(args, network)
network = YoloWithLossCell(network)
args.logger.info('finish get network')
config = ConfigYOLOV3DarkNet53()
config.label_smooth = args.label_smooth
config.label_smooth_factor = args.label_smooth_factor
if args.training_shape:
config.multi_scale = [conver_training_shape(args)]
if args.resize_rate:
config.resize_rate = args.resize_rate
ds, data_size = create_yolo_dataset(image_dir=args.data_root, anno_path=args.annFile, is_training=True,
batch_size=args.per_batch_size, max_epoch=args.max_epoch,
device_num=args.group_size, rank=args.rank, config=config)
args.logger.info('Finish loading dataset')
args.steps_per_epoch = int(data_size / args.per_batch_size / args.group_size)
if not args.ckpt_interval:
args.ckpt_interval = args.steps_per_epoch
lr = get_lr(args)
opt = Momentum(params=get_param_groups(network),
learning_rate=Tensor(lr),
momentum=args.momentum,
weight_decay=args.weight_decay,
loss_scale=args.loss_scale)
is_gpu = context.get_context("device_target") == "GPU"
if is_gpu:
loss_scale_value = 1.0
loss_scale = FixedLossScaleManager(loss_scale_value, drop_overflow_update=False)
network = amp.build_train_network(network, optimizer=opt, loss_scale_manager=loss_scale,
level="O2", keep_batchnorm_fp32=True)
keep_loss_fp32(network)
else:
network = TrainingWrapper(network, opt)
network.set_train()
if args.rank_save_ckpt_flag:
# checkpoint save
ckpt_max_num = args.max_epoch * args.steps_per_epoch // args.ckpt_interval
ckpt_config = CheckpointConfig(save_checkpoint_steps=args.ckpt_interval,
keep_checkpoint_max=ckpt_max_num)
save_ckpt_path = os.path.join(args.outputs_dir, 'ckpt_' + str(args.rank) + '/')
ckpt_cb = ModelCheckpoint(config=ckpt_config,
directory=save_ckpt_path,
prefix='{}'.format(args.rank))
cb_params = _InternalCallbackParam()
cb_params.train_network = network
cb_params.epoch_num = ckpt_max_num
cb_params.cur_epoch_num = 1
run_context = RunContext(cb_params)
ckpt_cb.begin(run_context)
old_progress = -1
t_end = time.time()
data_loader = ds.create_dict_iterator(output_numpy=True)
for i, data in enumerate(data_loader):
images = data["image"]
input_shape = images.shape[2:4]
args.logger.info('iter[{}], shape{}'.format(i, input_shape[0]))
images = Tensor.from_numpy(images)
batch_y_true_0 = Tensor.from_numpy(data['bbox1'])
batch_y_true_1 = Tensor.from_numpy(data['bbox2'])
batch_y_true_2 = Tensor.from_numpy(data['bbox3'])
batch_gt_box0 = Tensor.from_numpy(data['gt_box1'])
batch_gt_box1 = Tensor.from_numpy(data['gt_box2'])
batch_gt_box2 = Tensor.from_numpy(data['gt_box3'])
input_shape = Tensor(tuple(input_shape[::-1]), ms.float32)
loss = network(images, batch_y_true_0, batch_y_true_1, batch_y_true_2, batch_gt_box0, batch_gt_box1,
batch_gt_box2, input_shape)
loss_meter.update(loss.asnumpy())
if args.rank_save_ckpt_flag:
# ckpt progress
cb_params.cur_step_num = i + 1 # current step number
cb_params.batch_num = i + 2
ckpt_cb.step_end(run_context)
if i % args.log_interval == 0:
time_used = time.time() - t_end
epoch = int(i / args.steps_per_epoch)
fps = args.per_batch_size * (i - old_progress) * args.group_size / time_used
if args.rank == 0:
args.logger.info(
'epoch[{}], iter[{}], {}, {:.2f} imgs/sec, lr:{}'.format(epoch, i, loss_meter, fps, lr[i]))
t_end = time.time()
loss_meter.reset()
old_progress = i
if (i + 1) % args.steps_per_epoch == 0 and args.rank_save_ckpt_flag:
cb_params.cur_epoch_num += 1
if args.need_profiler:
if i == 10:
profiler.analyse()
break
args.logger.info('==========end training===============')
def train():
"""Train function."""
args = parse_args()
devid = int(os.getenv('DEVICE_ID', '0'))
context.set_context(mode=context.GRAPH_MODE,
enable_auto_mixed_precision=True,
device_target=args.device_target,
save_graphs=False,
device_id=devid)
loss_meter = AverageMeter('loss')
network = YOLOV4CspDarkNet53(is_training=True)
# default is kaiming-normal
default_recursive_init(network)
if args.pretrained_backbone:
pretrained_backbone_slice = args.pretrained_backbone.split('/')
backbone_ckpt_file = pretrained_backbone_slice[
len(pretrained_backbone_slice) - 1]
local_backbone_ckpt_path = '/cache/' + backbone_ckpt_file
# download backbone checkpoint
mox.file.copy_parallel(src_url=args.pretrained_backbone,
dst_url=local_backbone_ckpt_path)
args.pretrained_backbone = local_backbone_ckpt_path
load_yolov4_params(args, network)
network = YoloWithLossCell(network)
args.logger.info('finish get network')
config = ConfigYOLOV4CspDarkNet53()
config.label_smooth = args.label_smooth
config.label_smooth_factor = args.label_smooth_factor
if args.training_shape:
config.multi_scale = [convert_training_shape(args)]
if args.resize_rate:
config.resize_rate = args.resize_rate
# data download
local_data_path = '/cache/data'
local_ckpt_path = '/cache/ckpt_file'
print('Download data.')
mox.file.copy_parallel(src_url=args.data_url, dst_url=local_data_path)
ds, data_size = create_yolo_dataset(
image_dir=os.path.join(local_data_path, 'images'),
anno_path=os.path.join(local_data_path, 'annotation.json'),
is_training=True,
batch_size=args.per_batch_size,
max_epoch=args.max_epoch,
device_num=args.group_size,
rank=args.rank,
config=config)
args.logger.info('Finish loading dataset')
args.steps_per_epoch = int(data_size / args.per_batch_size /
args.group_size)
if not args.ckpt_interval:
args.ckpt_interval = args.steps_per_epoch * 10
lr = get_lr(args)
opt = Momentum(params=get_param_groups(network),
learning_rate=Tensor(lr),
momentum=args.momentum,
weight_decay=args.weight_decay,
loss_scale=args.loss_scale)
is_gpu = context.get_context("device_target") == "GPU"
if is_gpu:
loss_scale_value = 1.0
loss_scale = FixedLossScaleManager(loss_scale_value,
drop_overflow_update=False)
network = amp.build_train_network(network,
optimizer=opt,
loss_scale_manager=loss_scale,
level="O2",
keep_batchnorm_fp32=False)
keep_loss_fp32(network)
else:
network = TrainingWrapper(network, opt)
network.set_train()
# checkpoint save
ckpt_max_num = 10
ckpt_config = CheckpointConfig(save_checkpoint_steps=args.ckpt_interval,
keep_checkpoint_max=ckpt_max_num)
ckpt_cb = ModelCheckpoint(config=ckpt_config,
directory=local_ckpt_path,
prefix='yolov4')
cb_params = _InternalCallbackParam()
cb_params.train_network = network
cb_params.epoch_num = ckpt_max_num
cb_params.cur_epoch_num = 1
run_context = RunContext(cb_params)
ckpt_cb.begin(run_context)
old_progress = -1
t_end = time.time()
data_loader = ds.create_dict_iterator(output_numpy=True, num_epochs=1)
for i, data in enumerate(data_loader):
images = data["image"]
input_shape = images.shape[2:4]
images = Tensor.from_numpy(images)
batch_y_true_0 = Tensor.from_numpy(data['bbox1'])
batch_y_true_1 = Tensor.from_numpy(data['bbox2'])
batch_y_true_2 = Tensor.from_numpy(data['bbox3'])
batch_gt_box0 = Tensor.from_numpy(data['gt_box1'])
batch_gt_box1 = Tensor.from_numpy(data['gt_box2'])
batch_gt_box2 = Tensor.from_numpy(data['gt_box3'])
input_shape = Tensor(tuple(input_shape[::-1]), ms.float32)
loss = network(images, batch_y_true_0, batch_y_true_1, batch_y_true_2,
batch_gt_box0, batch_gt_box1, batch_gt_box2,
input_shape)
loss_meter.update(loss.asnumpy())
# ckpt progress
cb_params.cur_step_num = i + 1 # current step number
cb_params.batch_num = i + 2
ckpt_cb.step_end(run_context)
if i % args.log_interval == 0:
time_used = time.time() - t_end
epoch = int(i / args.steps_per_epoch)
fps = args.per_batch_size * (
i - old_progress) * args.group_size / time_used
if args.rank == 0:
args.logger.info(
'epoch[{}], iter[{}], {}, {:.2f} imgs/sec, lr:{}'.format(
epoch, i, loss_meter, fps, lr[i]))
t_end = time.time()
loss_meter.reset()
old_progress = i
if (i + 1) % args.steps_per_epoch == 0:
cb_params.cur_epoch_num += 1
args.logger.info('==========end training===============')
# upload checkpoint files
print('Upload checkpoint.')
mox.file.copy_parallel(src_url=local_ckpt_path, dst_url=args.train_url)
def train():
"""Train function."""
args = parse_args()
# init distributed
if args.is_distributed:
init()
args.rank = get_rank()
args.group_size = get_group_size()
# select for master rank save ckpt or all rank save, compatiable for model parallel
args.rank_save_ckpt_flag = 0
if args.is_save_on_master:
if args.rank == 0:
args.rank_save_ckpt_flag = 1
else:
args.rank_save_ckpt_flag = 1
# logger
args.outputs_dir = os.path.join(
args.ckpt_path,
datetime.datetime.now().strftime('%Y-%m-%d_time_%H_%M_%S'))
args.logger = get_logger(args.outputs_dir, args.rank)
args.logger.save_args(args)
if args.need_profiler:
from mindspore.profiler.profiling import Profiler
profiler = Profiler(output_path=args.outputs_dir,
is_detail=True,
is_show_op_path=True)
loss_meter = AverageMeter('loss')
context.reset_auto_parallel_context()
if args.is_distributed:
parallel_mode = ParallelMode.DATA_PARALLEL
degree = get_group_size()
else:
parallel_mode = ParallelMode.STAND_ALONE
degree = 1
context.set_auto_parallel_context(parallel_mode=parallel_mode,
mirror_mean=True,
device_num=degree)
network = YOLOV3DarkNet53(is_training=True)
# default is kaiming-normal
default_recurisive_init(network)
if args.resume_yolov3:
param_dict = load_checkpoint(args.resume_yolov3)
param_dict_new = {}
for key, values in param_dict.items():
args.logger.info('ckpt param name = {}'.format(key))
if key.startswith('moments.') or key.startswith('global_') or \
key.startswith('learning_rate') or key.startswith('momentum'):
continue
elif key.startswith('yolo_network.'):
key_new = key[13:]
if key_new.endswith('1.beta'):
key_new = key_new.replace('1.beta', 'batchnorm.beta')
if key_new.endswith('1.gamma'):
key_new = key_new.replace('1.gamma', 'batchnorm.gamma')
if key_new.endswith('1.moving_mean'):
key_new = key_new.replace('1.moving_mean',
'batchnorm.moving_mean')
if key_new.endswith('1.moving_variance'):
key_new = key_new.replace('1.moving_variance',
'batchnorm.moving_variance')
if key_new.endswith('.weight'):
if key_new.endswith('0.weight'):
key_new = key_new.replace('0.weight', 'conv.weight')
else:
key_new = key_new.replace('.weight', '.conv.weight')
if key_new.endswith('.bias'):
key_new = key_new.replace('.bias', '.conv.bias')
param_dict_new[key_new] = values
args.logger.info('in resume {}'.format(key_new))
else:
param_dict_new[key] = values
args.logger.info('in resume {}'.format(key))
args.logger.info('resume finished')
for _, param in network.parameters_and_names():
args.logger.info('network param name = {}'.format(param.name))
if param.name not in param_dict_new:
args.logger.info('not match param name = {}'.format(
param.name))
load_param_into_net(network, param_dict_new)
args.logger.info('load_model {} success'.format(args.resume_yolov3))
config = ConfigYOLOV3DarkNet53()
# convert fusion network to quantization aware network
if config.quantization_aware:
network = quant.convert_quant_network(network,
bn_fold=True,
per_channel=[True, False],
symmetric=[True, False])
network = YoloWithLossCell(network)
args.logger.info('finish get network')
config.label_smooth = args.label_smooth
config.label_smooth_factor = args.label_smooth_factor
if args.training_shape:
config.multi_scale = [conver_training_shape(args)]
if args.resize_rate:
config.resize_rate = args.resize_rate
ds, data_size = create_yolo_dataset(image_dir=args.data_root,
anno_path=args.annFile,
is_training=True,
batch_size=args.per_batch_size,
max_epoch=args.max_epoch,
device_num=args.group_size,
rank=args.rank,
config=config)
args.logger.info('Finish loading dataset')
args.steps_per_epoch = int(data_size / args.per_batch_size /
args.group_size)
if not args.ckpt_interval:
args.ckpt_interval = args.steps_per_epoch
# lr scheduler
if args.lr_scheduler == 'exponential':
lr = warmup_step_lr(
args.lr,
args.lr_epochs,
args.steps_per_epoch,
args.warmup_epochs,
args.max_epoch,
gamma=args.lr_gamma,
)
elif args.lr_scheduler == 'cosine_annealing':
lr = warmup_cosine_annealing_lr(args.lr, args.steps_per_epoch,
args.warmup_epochs, args.max_epoch,
args.T_max, args.eta_min)
elif args.lr_scheduler == 'cosine_annealing_V2':
lr = warmup_cosine_annealing_lr_V2(args.lr, args.steps_per_epoch,
args.warmup_epochs, args.max_epoch,
args.T_max, args.eta_min)
elif args.lr_scheduler == 'cosine_annealing_sample':
lr = warmup_cosine_annealing_lr_sample(args.lr, args.steps_per_epoch,
args.warmup_epochs,
args.max_epoch, args.T_max,
args.eta_min)
else:
raise NotImplementedError(args.lr_scheduler)
opt = Momentum(params=get_param_groups(network),
learning_rate=Tensor(lr),
momentum=args.momentum,
weight_decay=args.weight_decay,
loss_scale=args.loss_scale)
network = TrainingWrapper(network, opt)
network.set_train()
if args.rank_save_ckpt_flag:
# checkpoint save
ckpt_max_num = args.max_epoch * args.steps_per_epoch // args.ckpt_interval
ckpt_config = CheckpointConfig(
save_checkpoint_steps=args.ckpt_interval,
keep_checkpoint_max=ckpt_max_num)
ckpt_cb = ModelCheckpoint(config=ckpt_config,
directory=args.outputs_dir,
prefix='{}'.format(args.rank))
cb_params = _InternalCallbackParam()
cb_params.train_network = network
cb_params.epoch_num = ckpt_max_num
cb_params.cur_epoch_num = 1
run_context = RunContext(cb_params)
ckpt_cb.begin(run_context)
old_progress = -1
t_end = time.time()
data_loader = ds.create_dict_iterator()
shape_record = ShapeRecord()
for i, data in enumerate(data_loader):
images = data["image"]
input_shape = images.shape[2:4]
args.logger.info('iter[{}], shape{}'.format(i, input_shape[0]))
shape_record.set(input_shape)
images = Tensor(images)
annos = data["annotation"]
if args.group_size == 1:
batch_y_true_0, batch_y_true_1, batch_y_true_2, batch_gt_box0, batch_gt_box1, batch_gt_box2 = \
batch_preprocess_true_box(annos, config, input_shape)
else:
batch_y_true_0, batch_y_true_1, batch_y_true_2, batch_gt_box0, batch_gt_box1, batch_gt_box2 = \
batch_preprocess_true_box_single(annos, config, input_shape)
batch_y_true_0 = Tensor(batch_y_true_0)
batch_y_true_1 = Tensor(batch_y_true_1)
batch_y_true_2 = Tensor(batch_y_true_2)
batch_gt_box0 = Tensor(batch_gt_box0)
batch_gt_box1 = Tensor(batch_gt_box1)
batch_gt_box2 = Tensor(batch_gt_box2)
input_shape = Tensor(tuple(input_shape[::-1]), ms.float32)
loss = network(images, batch_y_true_0, batch_y_true_1, batch_y_true_2,
batch_gt_box0, batch_gt_box1, batch_gt_box2,
input_shape)
loss_meter.update(loss.asnumpy())
if args.rank_save_ckpt_flag:
# ckpt progress
cb_params.cur_step_num = i + 1 # current step number
cb_params.batch_num = i + 2
ckpt_cb.step_end(run_context)
if i % args.log_interval == 0:
time_used = time.time() - t_end
epoch = int(i / args.steps_per_epoch)
fps = args.per_batch_size * (
i - old_progress) * args.group_size / time_used
if args.rank == 0:
args.logger.info(
'epoch[{}], iter[{}], {}, {:.2f} imgs/sec, lr:{}'.format(
epoch, i, loss_meter, fps, lr[i]))
t_end = time.time()
loss_meter.reset()
old_progress = i
if (i + 1) % args.steps_per_epoch == 0 and args.rank_save_ckpt_flag:
cb_params.cur_epoch_num += 1
if args.need_profiler:
if i == 10:
profiler.analyse()
break
args.logger.info('==========end training===============')
def test_yolov3_darknet53():
devid = int(os.getenv('DEVICE_ID')) if os.getenv('DEVICE_ID') else 0
context.set_context(mode=context.GRAPH_MODE,
enable_auto_mixed_precision=True,
device_target="Ascend",
device_id=devid)
rank = 0
device_num = 1
lr_init = 0.001
epoch_size = 3
batch_size = 32
loss_scale = 1024
mindrecord_dir = DATA_DIR
# It will generate mindrecord file in args_opt.mindrecord_dir,
# and the file name is yolo.mindrecord0, 1, ... file_num.
if not os.path.isdir(mindrecord_dir):
raise KeyError("mindrecord path is not exist.")
data_root = os.path.join(mindrecord_dir, 'train2014')
annFile = os.path.join(mindrecord_dir,
'annotations/instances_train2014.json')
# print("yolov3 mindrecord is ", mindrecord_file)
if not os.path.exists(annFile):
print("instances_train2014 file is not exist.")
assert False
loss_meter = AverageMeter('loss')
context.reset_auto_parallel_context()
parallel_mode = ParallelMode.STAND_ALONE
context.set_auto_parallel_context(parallel_mode=parallel_mode,
gradients_mean=True,
device_num=1)
network = YOLOV3DarkNet53(is_training=True)
# default is kaiming-normal
default_recurisive_init(network)
network = YoloWithLossCell(network)
print('finish get network')
config = ConfigYOLOV3DarkNet53()
label_smooth = 0
label_smooth_factor = 0.1
config.label_smooth = label_smooth
config.label_smooth_factor = label_smooth_factor
# When create MindDataset, using the fitst mindrecord file, such as yolo.mindrecord0.
print("Create dataset begin!")
training_shape = [int(416), int(416)]
config.multi_scale = [training_shape]
num_samples = 256
ds, data_size = create_yolo_dataset(image_dir=data_root,
anno_path=annFile,
is_training=True,
batch_size=batch_size,
max_epoch=epoch_size,
device_num=device_num,
rank=rank,
config=config,
num_samples=num_samples)
print("Create dataset done!")
per_batch_size = batch_size
group_size = 1
print("data_size:", data_size)
steps_per_epoch = int(data_size / per_batch_size / group_size)
print("steps_per_epoch:", steps_per_epoch)
warmup_epochs = 0.
max_epoch = epoch_size
T_max = 1
eta_min = 0
lr = warmup_cosine_annealing_lr(lr_init, steps_per_epoch, warmup_epochs,
max_epoch, T_max, eta_min)
opt = Momentum(params=get_param_groups(network),
learning_rate=Tensor(lr),
momentum=0.9,
weight_decay=0.0005,
loss_scale=loss_scale)
network = TrainingWrapper(network, opt)
network.set_train()
old_progress = -1
t_end = time.time()
data_loader = ds.create_dict_iterator(output_numpy=True)
train_starttime = time.time()
time_used_per_epoch = 0
print("time:", time.time())
for i, data in enumerate(data_loader):
images = data["image"]
input_shape = images.shape[2:4]
print('iter[{}], shape{}'.format(i, input_shape[0]))
images = Tensor.from_numpy(images)
batch_y_true_0 = Tensor.from_numpy(data['bbox1'])
batch_y_true_1 = Tensor.from_numpy(data['bbox2'])
batch_y_true_2 = Tensor.from_numpy(data['bbox3'])
batch_gt_box0 = Tensor.from_numpy(data['gt_box1'])
batch_gt_box1 = Tensor.from_numpy(data['gt_box2'])
batch_gt_box2 = Tensor.from_numpy(data['gt_box3'])
input_shape = Tensor(tuple(input_shape[::-1]), ms.float32)
loss = network(images, batch_y_true_0, batch_y_true_1, batch_y_true_2,
batch_gt_box0, batch_gt_box1, batch_gt_box2,
input_shape)
loss_meter.update(loss.asnumpy())
if (i + 1) % steps_per_epoch == 0:
time_used = time.time() - t_end
epoch = int(i / steps_per_epoch)
fps = per_batch_size * (i - old_progress) * group_size / time_used
if rank == 0:
print(
'epoch[{}], iter[{}], {}, {:.2f} imgs/sec, lr:{}, time_used:{}'
.format(epoch, i, loss_meter, fps, lr[i], time_used))
t_end = time.time()
loss_meter.reset()
old_progress = i
time_used_per_epoch = time_used
train_endtime = time.time() - train_starttime
print('train_time_used:{}'.format(train_endtime))
expect_loss_value = 3210.0
loss_value = re.findall(r"\d+\.?\d*", str(loss_meter))
print('loss_value:{}'.format(loss_value[0]))
assert float(loss_value[0]) < expect_loss_value
export_time_used = 20.0
print('time_used_per_epoch:{}'.format(time_used_per_epoch))
assert time_used_per_epoch < export_time_used
print('==========test case passed===========')
def test():
"""The function of eval."""
start_time = time.time()
args = parse_args()
devid = int(os.getenv('DEVICE_ID')) if os.getenv('DEVICE_ID') else 0
context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target,
save_graphs=False, device_id=devid)
# logger
args.outputs_dir = os.path.join(args.log_path,
datetime.datetime.now().strftime('%Y-%m-%d_time_%H_%M_%S'))
rank_id = int(os.environ.get('RANK_ID')) if os.environ.get('RANK_ID') else 0
args.logger = get_logger(args.outputs_dir, rank_id)
args.logger.info('Creating Network....')
network = YOLOV4CspDarkNet53(is_training=False)
ckpt_file_slice = args.checkpoint_path.split('/')
ckpt_file = ckpt_file_slice[len(ckpt_file_slice)-1]
local_ckpt_path = '/cache/'+ckpt_file
# download checkpoint
mox.file.copy_parallel(src_url=args.checkpoint_path, dst_url=local_ckpt_path)
args.logger.info(local_ckpt_path)
if os.path.isfile(local_ckpt_path):
param_dict = load_checkpoint(local_ckpt_path)
param_dict_new = {}
for key, values in param_dict.items():
if key.startswith('moments.'):
continue
elif key.startswith('yolo_network.'):
param_dict_new[key[13:]] = values
else:
param_dict_new[key] = values
load_param_into_net(network, param_dict_new)
args.logger.info('load_model {} success'.format(local_ckpt_path))
else:
args.logger.info('{} not exists or not a pre-trained file'.format(local_ckpt_path))
assert FileNotFoundError('{} not exists or not a pre-trained file'.format(local_ckpt_path))
exit(1)
config = ConfigYOLOV4CspDarkNet53()
if args.testing_shape:
config.test_img_shape = convert_testing_shape(args)
local_data_path = '/cache/data'
# data download
print('Download data.')
mox.file.copy_parallel(src_url=args.data_url, dst_url=local_data_path)
args.data_root = os.path.join(local_data_path, 'images')
args.ann_file = os.path.join(local_data_path, 'eval_annotation.json')
ds, data_size = create_yolo_dataset(args.data_root, args.ann_file,
is_training=False, batch_size=args.per_batch_size,
max_epoch=1, device_num=1, rank=rank_id, shuffle=False,
config=config)
args.logger.info('testing shape: {}'.format(config.test_img_shape))
args.logger.info('total {} images to eval'.format(data_size))
network.set_train(False)
# init detection engine
detection = DetectionEngine(args)
input_shape = Tensor(tuple(config.test_img_shape), ms.float32)
args.logger.info('Start inference....')
for i, data in enumerate(ds.create_dict_iterator(num_epochs=1)):
image = data["image"]
image_shape = data["image_shape"]
image_id = data["img_id"]
prediction = network(image, input_shape)
output_big, output_me, output_small = prediction
output_big = output_big.asnumpy()
output_me = output_me.asnumpy()
output_small = output_small.asnumpy()
image_id = image_id.asnumpy()
image_shape = image_shape.asnumpy()
detection.detect([output_small, output_me, output_big], args.per_batch_size,
image_shape, image_id, config)
if i % 1000 == 0:
args.logger.info('Processing... {:.2f}% '.format(i * args.per_batch_size / data_size * 100))
args.logger.info('Calculating mAP...')
detection.do_nms_for_results()
result_file_path = detection.write_result()
args.logger.info('result file path: {}'.format(result_file_path))
eval_result = detection.get_eval_result()
cost_time = time.time() - start_time
args.logger.info('\n=============coco eval result=========\n' + eval_result)
args.logger.info('testing cost time {:.2f}h'.format(cost_time / 3600.))
def train():
"""Train function."""
args = parse_args()
args.logger.save_args(args)
if args.need_profiler:
from mindspore.profiler.profiling import Profiler
profiler = Profiler(output_path=args.outputs_dir,
is_detail=True,
is_show_op_path=True)
loss_meter = AverageMeter('loss')
context.reset_auto_parallel_context()
parallel_mode = ParallelMode.STAND_ALONE
degree = 1
if args.is_distributed:
parallel_mode = ParallelMode.DATA_PARALLEL
degree = get_group_size()
context.set_auto_parallel_context(parallel_mode=parallel_mode,
gradients_mean=True,
device_num=degree)
network = YOLOV3DarkNet53(is_training=True)
# default is kaiming-normal
default_recurisive_init(network)
load_yolov3_quant_params(args, network)
config = ConfigYOLOV3DarkNet53()
# convert fusion network to quantization aware network
if config.quantization_aware:
network = quant.convert_quant_network(network,
bn_fold=True,
per_channel=[True, False],
symmetric=[True, False])
network = YoloWithLossCell(network)
args.logger.info('finish get network')
config.label_smooth = args.label_smooth
config.label_smooth_factor = args.label_smooth_factor
if args.training_shape:
config.multi_scale = [conver_training_shape(args)]
if args.resize_rate:
config.resize_rate = args.resize_rate
ds, data_size = create_yolo_dataset(image_dir=args.data_root,
anno_path=args.annFile,
is_training=True,
batch_size=args.per_batch_size,
max_epoch=args.max_epoch,
device_num=args.group_size,
rank=args.rank,
config=config)
args.logger.info('Finish loading dataset')
args.steps_per_epoch = int(data_size / args.per_batch_size /
args.group_size)
if not args.ckpt_interval:
args.ckpt_interval = args.steps_per_epoch
lr = get_lr(args)
opt = Momentum(params=get_param_groups(network),
learning_rate=Tensor(lr),
momentum=args.momentum,
weight_decay=args.weight_decay,
loss_scale=args.loss_scale)
network = TrainingWrapper(network, opt)
network.set_train()
if args.rank_save_ckpt_flag:
# checkpoint save
ckpt_max_num = args.max_epoch * args.steps_per_epoch // args.ckpt_interval
ckpt_config = CheckpointConfig(
save_checkpoint_steps=args.ckpt_interval,
keep_checkpoint_max=ckpt_max_num)
save_ckpt_path = os.path.join(args.outputs_dir,
'ckpt_' + str(args.rank) + '/')
ckpt_cb = ModelCheckpoint(config=ckpt_config,
directory=save_ckpt_path,
prefix='{}'.format(args.rank))
cb_params = _InternalCallbackParam()
cb_params.train_network = network
cb_params.epoch_num = ckpt_max_num
cb_params.cur_epoch_num = 1
run_context = RunContext(cb_params)
ckpt_cb.begin(run_context)
old_progress = -1
t_end = time.time()
data_loader = ds.create_dict_iterator(output_numpy=True, num_epochs=1)
shape_record = ShapeRecord()
for i, data in enumerate(data_loader):
images = data["image"]
input_shape = images.shape[2:4]
args.logger.info('iter[{}], shape{}'.format(i, input_shape[0]))
shape_record.set(input_shape)
images = Tensor.from_numpy(images)
annos = data["annotation"]
if args.group_size == 1:
batch_y_true_0, batch_y_true_1, batch_y_true_2, batch_gt_box0, batch_gt_box1, batch_gt_box2 = \
batch_preprocess_true_box(annos, config, input_shape)
else:
batch_y_true_0, batch_y_true_1, batch_y_true_2, batch_gt_box0, batch_gt_box1, batch_gt_box2 = \
batch_preprocess_true_box_single(annos, config, input_shape)
batch_y_true_0 = Tensor.from_numpy(batch_y_true_0)
batch_y_true_1 = Tensor.from_numpy(batch_y_true_1)
batch_y_true_2 = Tensor.from_numpy(batch_y_true_2)
batch_gt_box0 = Tensor.from_numpy(batch_gt_box0)
batch_gt_box1 = Tensor.from_numpy(batch_gt_box1)
batch_gt_box2 = Tensor.from_numpy(batch_gt_box2)
input_shape = Tensor(tuple(input_shape[::-1]), ms.float32)
loss = network(images, batch_y_true_0, batch_y_true_1, batch_y_true_2,
batch_gt_box0, batch_gt_box1, batch_gt_box2,
input_shape)
loss_meter.update(loss.asnumpy())
if args.rank_save_ckpt_flag:
# ckpt progress
cb_params.cur_step_num = i + 1 # current step number
cb_params.batch_num = i + 2
ckpt_cb.step_end(run_context)
if i % args.log_interval == 0:
time_used = time.time() - t_end
epoch = int(i / args.steps_per_epoch)
fps = args.per_batch_size * (
i - old_progress) * args.group_size / time_used
if args.rank == 0:
args.logger.info(
'epoch[{}], iter[{}], {}, {:.2f} imgs/sec, lr:{}'.format(
epoch, i, loss_meter, fps, lr[i]))
t_end = time.time()
loss_meter.reset()
old_progress = i
if (i + 1) % args.steps_per_epoch == 0 and args.rank_save_ckpt_flag:
cb_params.cur_epoch_num += 1
if args.need_profiler:
if i == 10:
profiler.analyse()
break
args.logger.info('==========end training===============')
def train():
"""Train function."""
args = parse_args()
devid = int(os.getenv('DEVICE_ID')) if os.getenv('DEVICE_ID') else 0
context.set_context(mode=context.GRAPH_MODE, enable_auto_mixed_precision=True,
device_target=args.device_target, save_graphs=True, device_id=devid)
# init distributed
if args.is_distributed:
if args.device_target == "Ascend":
init()
else:
init("nccl")
args.rank = get_rank()
args.group_size = get_group_size()
# select for master rank save ckpt or all rank save, compatiable for model parallel
args.rank_save_ckpt_flag = 0
if args.is_save_on_master:
if args.rank == 0:
args.rank_save_ckpt_flag = 1
else:
args.rank_save_ckpt_flag = 1
# logger
args.outputs_dir = os.path.join(args.ckpt_path,
datetime.datetime.now().strftime('%Y-%m-%d_time_%H_%M_%S'))
args.logger = get_logger(args.outputs_dir, args.rank)
args.logger.save_args(args)
if args.need_profiler:
from mindspore.profiler.profiling import Profiler
profiler = Profiler(output_path=args.outputs_dir, is_detail=True, is_show_op_path=True)
loss_meter = AverageMeter('loss')
context.reset_auto_parallel_context()
if args.is_distributed:
parallel_mode = ParallelMode.DATA_PARALLEL
degree = get_group_size()
else:
parallel_mode = ParallelMode.STAND_ALONE
degree = 1
context.set_auto_parallel_context(parallel_mode=parallel_mode, mirror_mean=True, device_num=degree)
network = YOLOV3DarkNet53(is_training=True)
# default is kaiming-normal
default_recurisive_init(network)
if args.pretrained_backbone:
network = load_backbone(network, args.pretrained_backbone, args)
args.logger.info('load pre-trained backbone {} into network'.format(args.pretrained_backbone))
else:
args.logger.info('Not load pre-trained backbone, please be careful')
if args.resume_yolov3:
param_dict = load_checkpoint(args.resume_yolov3)
param_dict_new = {}
for key, values in param_dict.items():
if key.startswith('moments.'):
continue
elif key.startswith('yolo_network.'):
param_dict_new[key[13:]] = values
args.logger.info('in resume {}'.format(key))
else:
param_dict_new[key] = values
args.logger.info('in resume {}'.format(key))
args.logger.info('resume finished')
load_param_into_net(network, param_dict_new)
args.logger.info('load_model {} success'.format(args.resume_yolov3))
network = YoloWithLossCell(network)
args.logger.info('finish get network')
config = ConfigYOLOV3DarkNet53()
config.label_smooth = args.label_smooth
config.label_smooth_factor = args.label_smooth_factor
if args.training_shape:
config.multi_scale = [conver_training_shape(args)]
if args.resize_rate:
config.resize_rate = args.resize_rate
ds, data_size = create_yolo_dataset(image_dir=args.data_root, anno_path=args.annFile, is_training=True,
batch_size=args.per_batch_size, max_epoch=args.max_epoch,
device_num=args.group_size, rank=args.rank, config=config)
args.logger.info('Finish loading dataset')
args.steps_per_epoch = int(data_size / args.per_batch_size / args.group_size)
if not args.ckpt_interval:
args.ckpt_interval = args.steps_per_epoch
# lr scheduler
if args.lr_scheduler == 'exponential':
lr = warmup_step_lr(args.lr,
args.lr_epochs,
args.steps_per_epoch,
args.warmup_epochs,
args.max_epoch,
gamma=args.lr_gamma,
)
elif args.lr_scheduler == 'cosine_annealing':
lr = warmup_cosine_annealing_lr(args.lr,
args.steps_per_epoch,
args.warmup_epochs,
args.max_epoch,
args.T_max,
args.eta_min)
elif args.lr_scheduler == 'cosine_annealing_V2':
lr = warmup_cosine_annealing_lr_V2(args.lr,
args.steps_per_epoch,
args.warmup_epochs,
args.max_epoch,
args.T_max,
args.eta_min)
elif args.lr_scheduler == 'cosine_annealing_sample':
lr = warmup_cosine_annealing_lr_sample(args.lr,
args.steps_per_epoch,
args.warmup_epochs,
args.max_epoch,
args.T_max,
args.eta_min)
else:
raise NotImplementedError(args.lr_scheduler)
opt = Momentum(params=get_param_groups(network),
learning_rate=Tensor(lr),
momentum=args.momentum,
weight_decay=args.weight_decay,
loss_scale=args.loss_scale)
enable_amp = False
is_gpu = context.get_context("device_target") == "GPU"
if is_gpu:
enable_amp = True
if enable_amp:
loss_scale_value = 1.0
loss_scale = FixedLossScaleManager(loss_scale_value, drop_overflow_update=False)
network = amp.build_train_network(network, optimizer=opt, loss_scale_manager=loss_scale,
level="O2", keep_batchnorm_fp32=True)
keep_loss_fp32(network)
else:
network = TrainingWrapper(network, opt)
network.set_train()
if args.rank_save_ckpt_flag:
# checkpoint save
ckpt_max_num = args.max_epoch * args.steps_per_epoch // args.ckpt_interval
ckpt_config = CheckpointConfig(save_checkpoint_steps=args.ckpt_interval,
keep_checkpoint_max=ckpt_max_num)
ckpt_cb = ModelCheckpoint(config=ckpt_config,
directory=args.outputs_dir,
prefix='{}'.format(args.rank))
cb_params = _InternalCallbackParam()
cb_params.train_network = network
cb_params.epoch_num = ckpt_max_num
cb_params.cur_epoch_num = 1
run_context = RunContext(cb_params)
ckpt_cb.begin(run_context)
old_progress = -1
t_end = time.time()
data_loader = ds.create_dict_iterator()
for i, data in enumerate(data_loader):
images = data["image"]
input_shape = images.shape[2:4]
args.logger.info('iter[{}], shape{}'.format(i, input_shape[0]))
images = Tensor(images)
batch_y_true_0 = Tensor(data['bbox1'])
batch_y_true_1 = Tensor(data['bbox2'])
batch_y_true_2 = Tensor(data['bbox3'])
batch_gt_box0 = Tensor(data['gt_box1'])
batch_gt_box1 = Tensor(data['gt_box2'])
batch_gt_box2 = Tensor(data['gt_box3'])
input_shape = Tensor(tuple(input_shape[::-1]), ms.float32)
loss = network(images, batch_y_true_0, batch_y_true_1, batch_y_true_2, batch_gt_box0, batch_gt_box1,
batch_gt_box2, input_shape)
loss_meter.update(loss.asnumpy())
if args.rank_save_ckpt_flag:
# ckpt progress
cb_params.cur_step_num = i + 1 # current step number
cb_params.batch_num = i + 2
ckpt_cb.step_end(run_context)
if i % args.log_interval == 0:
time_used = time.time() - t_end
epoch = int(i / args.steps_per_epoch)
fps = args.per_batch_size * (i - old_progress) * args.group_size / time_used
if args.rank == 0:
args.logger.info(
'epoch[{}], iter[{}], {}, {:.2f} imgs/sec, lr:{}'.format(epoch, i, loss_meter, fps, lr[i]))
t_end = time.time()
loss_meter.reset()
old_progress = i
if (i + 1) % args.steps_per_epoch == 0 and args.rank_save_ckpt_flag:
cb_params.cur_epoch_num += 1
if args.need_profiler:
if i == 10:
profiler.analyse()
break
args.logger.info('==========end training===============')
load_param_into_net(network, param_dict_new)
args.logger.info('load_model {} success'.format(args.pretrained))
else:
args.logger.info('{} not exists or not a pre-trained file'.format(args.pretrained))
assert FileNotFoundError('{} not exists or not a pre-trained file'.format(args.pretrained))
exit(1)
data_root = args.data_root
ann_file = args.ann_file
config = ConfigYOLOV4CspDarkNet53()
if args.testing_shape:
config.test_img_shape = convert_testing_shape(args.testing_shape)
ds, data_size = create_yolo_dataset(data_root, ann_file, is_training=False, batch_size=args.per_batch_size,
max_epoch=1, device_num=1, rank=rank_id, shuffle=False,
config=config)
args.logger.info('testing shape : {}'.format(config.test_img_shape))
args.logger.info('totol {} images to eval'.format(data_size))
network.set_train(False)
# init detection engine
detection = DetectionEngine(args)
input_shape = Tensor(tuple(config.test_img_shape), ms.float32)
args.logger.info('Start inference....')
for index, data in enumerate(ds.create_dict_iterator(num_epochs=1)):
image = data["image"]
def train():
"""Train function."""
args = parse_args()
# logger
args.outputs_dir = os.path.join(
args.ckpt_path,
datetime.datetime.now().strftime('%Y-%m-%d_time_%H_%M_%S'))
args.logger = get_logger(args.outputs_dir, args.rank)
args.logger.save_args(args)
loss_meter = AverageMeter('loss')
network = YOLOV3DarkNet53(is_training=True)
# default is kaiming-normal
default_recursive_init(network)
pretrained_backbone_slice = args.pretrained_backbone.split('/')
backbone_ckpt_file = pretrained_backbone_slice[
len(pretrained_backbone_slice) - 1]
local_backbone_ckpt_path = '/cache/' + backbone_ckpt_file
# download backbone checkpoint
mox.file.copy_parallel(src_url=args.pretrained_backbone,
dst_url=local_backbone_ckpt_path)
if args.pretrained_backbone:
network = load_backbone(network, local_backbone_ckpt_path, args)
args.logger.info('load pre-trained backbone {} into network'.format(
args.pretrained_backbone))
else:
args.logger.info('Not load pre-trained backbone, please be careful')
if args.resume_yolov3:
param_dict = load_checkpoint(args.resume_yolov3)
param_dict_new = {}
for key, values in param_dict.items():
if key.startswith('moments.'):
continue
elif key.startswith('yolo_network.'):
param_dict_new[key[13:]] = values
args.logger.info('in resume {}'.format(key))
else:
param_dict_new[key] = values
args.logger.info('in resume {}'.format(key))
args.logger.info('resume finished')
load_param_into_net(network, param_dict_new)
args.logger.info('load_model {} success'.format(args.resume_yolov3))
network = YoloWithLossCell(network)
args.logger.info('finish get network')
config = ConfigYOLOV3DarkNet53()
config.label_smooth = args.label_smooth
config.label_smooth_factor = args.label_smooth_factor
if args.training_shape:
config.multi_scale = [convert_training_shape(args)]
if args.resize_rate:
config.resize_rate = args.resize_rate
# data download
local_data_path = '/cache/data'
local_ckpt_path = '/cache/ckpt_file'
print('Download data.')
mox.file.copy_parallel(src_url=args.data_url, dst_url=local_data_path)
ds, data_size = create_yolo_dataset(
image_dir=os.path.join(local_data_path, 'images'),
anno_path=os.path.join(local_data_path, 'annotation.json'),
is_training=True,
batch_size=args.per_batch_size,
max_epoch=args.epoch_size,
device_num=args.group_size,
rank=args.rank,
config=config)
args.logger.info('Finish loading dataset')
args.steps_per_epoch = int(data_size / args.per_batch_size /
args.group_size)
if not args.ckpt_interval:
args.ckpt_interval = args.steps_per_epoch * 10
# lr scheduler
if args.lr_scheduler == 'exponential':
lr = warmup_step_lr(
args.lr,
args.lr_epochs,
args.steps_per_epoch,
args.warmup_epochs,
args.epoch_size,
gamma=args.lr_gamma,
)
elif args.lr_scheduler == 'cosine_annealing':
lr = warmup_cosine_annealing_lr(args.lr, args.steps_per_epoch,
args.warmup_epochs, args.max_epoch,
args.T_max, args.eta_min)
elif args.lr_scheduler == 'cosine_annealing_V2':
lr = warmup_cosine_annealing_lr_V2(args.lr, args.steps_per_epoch,
args.warmup_epochs, args.max_epoch,
args.T_max, args.eta_min)
elif args.lr_scheduler == 'cosine_annealing_sample':
lr = warmup_cosine_annealing_lr_sample(args.lr, args.steps_per_epoch,
args.warmup_epochs,
args.max_epoch, args.T_max,
args.eta_min)
else:
raise NotImplementedError(args.lr_scheduler)
opt = Momentum(params=get_param_groups(network),
learning_rate=Tensor(lr),
momentum=args.momentum,
weight_decay=args.weight_decay,
loss_scale=args.loss_scale)
network = TrainingWrapper(network, opt)
network.set_train()
# checkpoint save
ckpt_max_num = 10
ckpt_config = CheckpointConfig(save_checkpoint_steps=args.ckpt_interval,
keep_checkpoint_max=ckpt_max_num)
ckpt_cb = ModelCheckpoint(config=ckpt_config,
directory=local_ckpt_path,
prefix='yolov3')
cb_params = _InternalCallbackParam()
cb_params.train_network = network
cb_params.epoch_num = ckpt_max_num
cb_params.cur_epoch_num = 1
run_context = RunContext(cb_params)
ckpt_cb.begin(run_context)
old_progress = -1
t_end = time.time()
data_loader = ds.create_dict_iterator()
shape_record = ShapeRecord()
for i, data in enumerate(data_loader):
images = data["image"]
input_shape = images.shape[2:4]
shape_record.set(input_shape)
images = Tensor(images)
annos = data["annotation"]
if args.group_size == 1:
batch_y_true_0, batch_y_true_1, batch_y_true_2, batch_gt_box0, batch_gt_box1, batch_gt_box2 = \
batch_preprocess_true_box(annos, config, input_shape)
else:
batch_y_true_0, batch_y_true_1, batch_y_true_2, batch_gt_box0, batch_gt_box1, batch_gt_box2 = \
batch_preprocess_true_box_single(annos, config, input_shape)
batch_y_true_0 = Tensor(batch_y_true_0)
batch_y_true_1 = Tensor(batch_y_true_1)
batch_y_true_2 = Tensor(batch_y_true_2)
batch_gt_box0 = Tensor(batch_gt_box0)
batch_gt_box1 = Tensor(batch_gt_box1)
batch_gt_box2 = Tensor(batch_gt_box2)
input_shape = Tensor(tuple(input_shape[::-1]), ms.float32)
loss = network(images, batch_y_true_0, batch_y_true_1, batch_y_true_2,
batch_gt_box0, batch_gt_box1, batch_gt_box2,
input_shape)
loss_meter.update(loss.asnumpy())
# ckpt progress
cb_params.cur_step_num = i + 1 # current step number
cb_params.batch_num = i + 2
ckpt_cb.step_end(run_context)
if i % args.log_interval == 0:
time_used = time.time() - t_end
epoch = int(i / args.steps_per_epoch)
fps = args.per_batch_size * (
i - old_progress) * args.group_size / time_used
if args.rank == 0:
args.logger.info(
'epoch[{}], iter[{}], {}, {:.2f} imgs/sec, lr:{}'.format(
epoch, i, loss_meter, fps, lr[i]))
t_end = time.time()
loss_meter.reset()
old_progress = i
if (i + 1) % args.steps_per_epoch == 0:
cb_params.cur_epoch_num += 1
args.logger.info('==========end training===============')
# upload checkpoint files
print('Upload checkpoint.')
mox.file.copy_parallel(src_url=local_ckpt_path, dst_url=args.train_url)
