def main():
update_config("configs/coco/resnet_v1_101_coco_trainval_fpn_dcn_end2end_ohem.yaml")
log_init(filename=config.TRAIN.model_prefix + "train.log")
msg = pprint.pformat(config)
logging.info(msg)
os.environ["MXNET_CUDNN_AUTOTUNE_DEFAULT"] = "0"
os.environ["MXNET_GPU_MEM_POOL_TYPE"] = "Round"
ctx = [mx.gpu(int(i)) for i in config.gpus.split(',')]
ctx = ctx * config.network.IM_PER_GPU
train_net(ctx, config.TRAIN.begin_epoch, config.TRAIN.lr, config.TRAIN.lr_step)
def main():
update_config(
"configs/voc/resnet_v1_50_voc0712_rfcn_dcn_end2end_ohem_one_gpu.yaml")
log_init(filename=config.TRAIN.model_prefix + "train.log")
msg = pprint.pformat(config)
logging.info(msg)
os.environ["MXNET_CUDNN_AUTOTUNE_DEFAULT"] = "0"
os.environ["MXNET_GPU_MEM_POOL_TYPE"] = "Round"
ctx = [mx.gpu(int(i)) for i in config.gpus.split(',')]
train_net(ctx, config.network.pretrained, config.network.pretrained_epoch,
config.TRAIN.model_prefix, config.TRAIN.begin_epoch,
config.TRAIN.end_epoch, config.TRAIN.lr, config.TRAIN.lr_step)
def main():
os.environ["MXNET_CUDNN_AUTOTUNE_DEFAULT"] = "0"
os.environ["MXNET_GPU_MEM_POOL_TYPE"] = "Round"
args = parse_args()
setattr(mobula.config, "NVCC", args.nvcc)
config = easydict.EasyDict()
config.gpus = [int(x) for x in str(args.gpus).split(',')]
config.dataset = easydict.EasyDict()
config.dataset.NUM_CLASSES = args.num_classes
config.dataset.dataset_type = args.dataset_type
config.dataset.dataset_path = args.dataset_root
config.retinanet = easydict.EasyDict()
config.retinanet.network = easydict.EasyDict()
config.retinanet.network.FPN_STRIDES = [8, 16, 32, 64, 128]
config.retinanet.network.BASE_SIZES = [(32, 32), (64, 64), (128, 128),
(256, 256), (512, 512)]
config.retinanet.network.SCALES = [2**0, 2**(1 / 2), 2**(2 / 3)]
config.retinanet.network.RATIOS = [1 / 2, 1, 2]
config.retinanet.network.bbox_norm_coef = [0.1, 0.1, 0.2, 0.2]
config.TRAIN = easydict.EasyDict()
config.TRAIN.batch_size = 2 * len(config.gpus)
config.TRAIN.lr = 0.01 * config.TRAIN.batch_size / 16
config.TRAIN.warmup_lr = config.TRAIN.lr
config.TRAIN.warmup_step = 1000
config.TRAIN.wd = 1e-4
config.TRAIN.momentum = .9
config.TRAIN.log_path = "output/{}/RetinaNet-hflip".format(
config.dataset.dataset_type, config.TRAIN.lr)
config.TRAIN.log_interval = 100
config.TRAIN.cls_focal_loss_alpha = .25
config.TRAIN.cls_focal_loss_gamma = 2
config.TRAIN.image_short_size = 600
config.TRAIN.image_max_long_size = 1333
config.TRAIN.aspect_grouping = True
# if aspect_grouping is set to False, all images will be pad to (PAD_H, PAD_W)
config.TRAIN.PAD_H = 768
config.TRAIN.PAD_W = 768
config.TRAIN.begin_epoch = 0
config.TRAIN.end_epoch = 28
config.TRAIN.lr_step = [6, 8]
config.TRAIN.FLIP = True
config.TRAIN.resume = None
config.TRAIN.trainer_resume = None
config.network = easydict.EasyDict()
config.network.FIXED_PARAMS = []
config.network.sync_bn = True
config.network.use_global_stats = False if config.network.sync_bn else True
config.network.merge_backbone_bn = False
config.val = easydict.EasyDict()
if args.demo:
config.val.params_file = args.demo_params
config.val.viz = args.viz
demo_net(config)
else:
os.makedirs(config.TRAIN.log_path, exist_ok=True)
log_init(filename=os.path.join(config.TRAIN.log_path,
"train_{}.log".format(time.time())))
msg = pprint.pformat(config)
logging.info(msg)
train_net(config)
def main():
os.environ["MXNET_CUDNN_AUTOTUNE_DEFAULT"] = "0"
load_mobula_ops()
args = parse_args()
setattr(mobula.config, "NVCC", args.nvcc)
setattr(mobula.config, "SHOW_BUILDING_COMMAND", True)
config = easydict.EasyDict()
config.gpus = [int(x) for x in str(args.gpus).split(',')]
config.dataset = easydict.EasyDict()
config.dataset.NUM_CLASSES = args.num_classes
config.dataset.dataset_type = args.dataset_type
config.dataset.dataset_path = args.dataset_root
config.dataset.max_bbox_number = 200
config.retinanet = easydict.EasyDict()
config.retinanet.network = easydict.EasyDict()
config.retinanet.network.FPN_STRIDES = [8, 16, 32, 64, 128]
config.retinanet.network.BASE_SIZES = [(32, 32), (64, 64), (128, 128),
(256, 256), (512, 512)]
config.retinanet.network.SCALES = [2**0, 2**(1 / 3), 2**(2 / 3)]
config.retinanet.network.RATIOS = [1 / 2, 1, 2]
config.retinanet.network.bbox_norm_coef = [1, 1, 1, 1]
config.TRAIN = easydict.EasyDict()
config.TRAIN.batch_size = args.im_per_gpu * len(config.gpus)
config.TRAIN.lr = 0.01 * config.TRAIN.batch_size / 16
config.TRAIN.warmup_lr = config.TRAIN.lr * 1 / 3
config.TRAIN.warmup_step = int(1000 * 16 / config.TRAIN.batch_size)
config.TRAIN.wd = 1e-4
config.TRAIN.momentum = .9
config.TRAIN.log_interval = 100
config.TRAIN.cls_focal_loss_alpha = .25
config.TRAIN.cls_focal_loss_gamma = 2
config.TRAIN.image_short_size = 800
config.TRAIN.image_max_long_size = 1333
config.TRAIN.aspect_grouping = True
config.TRAIN.negative_iou_threshold = .4
config.TRAIN.positive_iou_threshold = .5
# if aspect_grouping is set to False, all images will be pad to (PAD_H, PAD_W)
config.TRAIN.PAD_H = 768
config.TRAIN.PAD_W = 768
config.TRAIN.begin_epoch = 0
config.TRAIN.end_epoch = 28
config.TRAIN.lr_step = [5]
config.TRAIN.FLIP = True
config.TRAIN.resume = None
config.TRAIN.trainer_resume = None
config.network = easydict.EasyDict()
config.network.BACKBONE = easydict.EasyDict()
config.network.BACKBONE.name = "resnetv1b"
config.network.BACKBONE.kwargs = easydict.EasyDict()
config.network.BACKBONE.kwargs.num_layers = 50
config.network.BACKBONE.kwargs.pretrained = True
config.network.BACKBONE.kwargs.norm_kwargs = {
"num_devices": len(config.gpus)
}
config.network.BACKBONE.kwargs.norm_layer = FrozenBatchNorm2d
config.network.FIXED_PARAMS = [".*layers1.*", ".*resnetv1b_conv0.*"]
config.network.fpn_neck_feature_dim = 256
config.TRAIN.log_path = "output/{}/RetinaNet-hflip-{}-{}".format(
config.dataset.dataset_type, config.TRAIN.lr,
config.TRAIN.image_short_size, config.TRAIN.image_max_long_size)
config.val = easydict.EasyDict()
if args.demo:
config.val.params_file = args.demo_params
config.val.viz = args.viz
demo_net(config)
else:
os.makedirs(config.TRAIN.log_path, exist_ok=True)
log_init(filename=os.path.join(config.TRAIN.log_path,
"train_{}.log".format(time.time())))
msg = pprint.pformat(config)
logging.info(msg)
train_net(config)
def main():
os.environ["MXNET_CUDNN_AUTOTUNE_DEFAULT"] = "0"
# os.environ["MXNET_GPU_MEM_POOL_TYPE"] = "Round"
load_mobula_ops()
args = parse_args()
setattr(mobula.config, "NVCC", args.nvcc)
setattr(mobula.config, "SHOW_BUILDING_COMMAND", True)
config = easydict.EasyDict()
config.gpus = [int(x) for x in str(args.gpus).split(',')]
config.use_hvd = args.hvd
if config.use_hvd:
import horovod.mxnet as hvd
hvd.init()
config.gpus = [hvd.local_rank()]
config.dataset = easydict.EasyDict()
config.dataset.NUM_CLASSES = args.num_classes
config.dataset.dataset_type = args.dataset_type
config.dataset.dataset_path = args.dataset_root
config.FCOS = easydict.EasyDict()
config.FCOS.network = easydict.EasyDict()
config.FCOS.network.FPN_SCALES = [8, 16, 32, 64, 128]
config.FCOS.network.FPN_MINIMUM_DISTANCES = [0, 64, 128, 256, 512]
config.FCOS.network.FPN_MAXIMUM_DISTANCES = [64, 128, 256, 512, 4096]
config.TRAIN = easydict.EasyDict()
config.TRAIN.batch_size = args.im_per_gpu * len(config.gpus)
config.TRAIN.lr = 0.01 * config.TRAIN.batch_size / 16
config.TRAIN.warmup_lr = config.TRAIN.lr * 1 / 3
config.TRAIN.warmup_step = int(1000 * 16 / config.TRAIN.batch_size)
config.TRAIN.wd = 1e-4
config.TRAIN.momentum = .9
config.TRAIN.log_interval = 100
config.TRAIN.cls_focal_loss_alpha = .25
config.TRAIN.cls_focal_loss_gamma = 2
config.TRAIN.image_short_size = 800
config.TRAIN.image_max_long_size = 1333
config.TRAIN.aspect_grouping = True
# if aspect_grouping is set to False, all images will be pad to (PAD_H, PAD_W)
config.TRAIN.PAD_H = 768
config.TRAIN.PAD_W = 768
config.TRAIN.begin_epoch = 0
config.TRAIN.end_epoch = 28
config.TRAIN.lr_step = [5]
config.TRAIN.FLIP = True
config.TRAIN.resume = None
config.TRAIN.trainer_resume = None
config.TRAIN.USE_FP16 = False
if config.TRAIN.USE_FP16:
os.environ["MXNET_SAFE_ACCUMULATION"] = "1"
config.network = easydict.EasyDict()
config.network.BACKBONE = easydict.EasyDict()
config.network.BACKBONE.name = "resnetv1"
config.network.BACKBONE.kwargs = easydict.EasyDict()
config.network.BACKBONE.kwargs.pretrained = True
config.network.FIXED_PARAMS = [".*stage1.*", ".*resnetv10_conv0.*"]
config.network.use_global_stats = True
config.network.sync_bn = False
config.network.fpn_neck_feature_dim = 256
if config.TRAIN.USE_FP16:
assert config.network.sync_bn is False, "Sync BatchNorm is not supported by amp."
config.val = easydict.EasyDict()
if os.path.exists(args.config):
with open(args.config) as f:
config_loaded = yaml.load(f, yaml.CLoader)
def update_config(c, v):
for k in v.keys():
if isinstance(v[k], dict):
update_config(c[k], v[k])
else:
c[k] = v[k]
update_config(config, config_loaded)
else:
logging.info("Escape loading config since it does not exist.")
config.TRAIN.log_path = "output/{}/{}-{}-{}-{}/reg_weighted_by_centerness_focal_alpha_gamma_lr_{}_{}_{}".format(
"FCOS-{}-p5-{}".format(config.network.BACKBONE.name, args.extra_flag),
"fp16" if config.TRAIN.USE_FP16 else "fp32",
"sync_bn" if config.network.sync_bn else "normal_bn",
"hvd" if config.use_hvd else "", config.dataset.dataset_type,
config.TRAIN.lr, config.TRAIN.image_short_size,
config.TRAIN.image_max_long_size)
if args.demo:
config.val.params_file = args.demo_params
config.val.viz = args.viz
demo_net(config)
else:
os.makedirs(config.TRAIN.log_path, exist_ok=True)
log_init(filename=os.path.join(config.TRAIN.log_path,
"train_{}.log".format(time.time())))
msg = pprint.pformat(config)
logging.info(msg)
train_net(config)