def __init__(self, cfg, weight_dict=None, extra_features=0):
super(DepthRCNN, self).__init__()
# Weight losses
if weight_dict is None:
self.weight_dict = {
'loss_objectness': cfg.LOSS_WEIGHTS.loss_objectness,
'loss_rpn_box_reg': cfg.LOSS_WEIGHTS.loss_rpn_box_reg,
'loss_classifier': cfg.LOSS_WEIGHTS.loss_classifier,
'loss_box_reg': cfg.LOSS_WEIGHTS.loss_box_reg,
'loss_mask': cfg.LOSS_WEIGHTS.loss_mask
}
else:
self.weight_dict = weight_dict
self.image_backbone = build_backbone(cfg)
self.freeze_backbone(self.image_backbone,
cfg.MODEL.BACKBONE.FREEZE_CONV_BODY_AT)
self.roi_heads = None
if cfg.MODEL.BACKBONE.DEPTH:
self.hha_backbone = build_backbone(cfg)
self.freeze_backbone(self.hha_backbone,
cfg.MODEL.BACKBONE.FREEZE_CONV_BODY_AT)
self.rpn = build_rpn(
cfg, self.image_backbone.out_channels * 2 + extra_features)
if not cfg.MODEL.RPN_ONLY:
self.roi_heads = build_roi_heads(
cfg, self.image_backbone.out_channels * 2 + extra_features)
else:
self.hha_backbone = None
self.rpn = build_rpn(
cfg, self.image_backbone.out_channels + extra_features)
if not cfg.MODEL.RPN_ONLY:
self.roi_heads = build_roi_heads(
cfg, self.image_backbone.out_channels + extra_features)
def test_build_shufflenet_backbones(self):
''' Make sure backbones run '''
self.assertGreater(len(registry.BACKBONES), 0)
name = "ShuffleNetV2"
print('Testing {}...'.format(name))
if name in BACKBONE_CFGS:
cfg = load_config(BACKBONE_CFGS[name])
else:
# Use default config if config file is not specified
cfg = copy.deepcopy(g_cfg)
backbone = build_backbone(cfg)
# make sures the backbone has `out_channels`
self.assertIsNotNone(
getattr(backbone, 'out_channels', None),
'Need to provide out_channels for backbone {}'.format(name))
N, C_in, H, W = 2, 3, 224, 256
input = torch.rand([N, C_in, H, W], dtype=torch.float32)
out = backbone(input)
for cur_out in out:
self.assertEqual(cur_out.shape[:2],
torch.Size([N, backbone.out_channels]))
def __init__(self, exp_dict):
super().__init__()
cfg_base_path = "./models/configs/"
self.n_classes = 21
cfg_path = cfg_base_path + "e2e_mask_rcnn_R_50_FPN_1x.yaml"
self.cfg = cfg
self.cfg.merge_from_file(cfg_path)
self.cfg.MODEL.ROI_BOX_HEAD.NUM_CLASSES = self.n_classes
# ---------------
# build model
self.backbone_fpn = backbone.build_backbone(self.cfg)
self.rpn = rpn.build_rpn(self.cfg, self.backbone_fpn.out_channels)
self.roi_heads = roi_heads.build_roi_heads(
self.cfg, self.backbone_fpn.out_channels)
# ---------------
# load checkpoint
checkpoint = _load_file(self.cfg)
load_state_dict(self, checkpoint.pop("model"))
#--------
# Opt stage
self.cfg.SOLVER.BASE_LR = ((0.0025 * 8) /
(16 / float(exp_dict["batch_size"])))
optimizer = make_optimizer(self.cfg, self)
scheduler = make_lr_scheduler(self.cfg, optimizer)
self.opt = optimizer
self.scheduler = scheduler
def init_maskrcnn_benchmark(self, args):
if 'datasets' in sys.modules:
del sys.modules['datasets']
this_dir = os.path.dirname(__file__)
lib_path = os.path.join(this_dir, '../../external/maskrcnn-benchmark')
sys.path.insert(0, lib_path)
# config
from maskrcnn_benchmark.utils.env import setup_environment # noqa F401 isort:skip
from maskrcnn_benchmark.config import cfg
self.cfg = cfg
#cfg.MODEL.BACKBONE.OUT_CHANNELS = 832
#cfg.MODEL.RESNETS.RES2_OUT_CHANNELS = 208
cfg.MODEL.RPN.PRE_NMS_TOP_N_TEST = 6000
cfg.MODEL.RPN.POST_NMS_TOP_N_TEST = 1000
cfg.MODEL.ROI_HEADS.SCORE_THRESH = 0
cfg.MODEL.ROI_HEADS.DETECTIONS_PER_IMG = 0
# bind classes/methods from MaskRCNN-Benchmark
from maskrcnn_benchmark.modeling.rpn.rpn import build_rpn
from maskrcnn_benchmark.modeling.roi_heads.roi_heads import build_roi_heads
from maskrcnn_benchmark.structures.image_list import to_image_list
from maskrcnn_benchmark.structures.bounding_box import BoxList
from maskrcnn_benchmark.modeling.backbone import build_backbone
self.basenet = build_backbone(cfg)
self.rpn = build_rpn(cfg)
self.roi_heads = build_roi_heads(cfg)
self.to_image_list = to_image_list
self.BoxList = BoxList
def __init__(self, cfg):
super(RelationPredictor, self).__init__(cfg)
self.cfg = cfg
self.device = cfg.MODEL.DEVICE
self.detector_pre_calculated = self.cfg.MODEL.ROI_RELATION_HEAD.DETECTOR_PRE_CALCULATED
self.detector_force_boxes = self.cfg.MODEL.ROI_BOX_HEAD.FORCE_BOXES
self.cfg_check()
feature_dim = self.backbone.out_channels
if not self.cfg.MODEL.ROI_RELATION_HEAD.SHARE_CONV_BACKBONE:
self.rel_backbone = build_backbone(cfg)
feature_dim = self.rel_backbone.out_channels
self.force_relations = cfg.MODEL.ROI_RELATION_HEAD.FORCE_RELATIONS
if cfg.MODEL.RELATION_ON and self.cfg.MODEL.ROI_RELATION_HEAD.ALGORITHM in SCENE_MODEL_DICT:
self.relation_head = build_roi_relation_head(cfg, feature_dim)
if cfg.MODEL.ATTRIBUTE_ON:
self.attribute_head = build_roi_attribute_head(cfg, feature_dim)
# self._freeze_components(self.cfg)
for p in self.backbone.parameters():
p.requires_grad = False
for p in self.rpn.parameters():
p.requires_grad = False
for p in self.roi_heads.parameters():
p.requires_grad = False
if not self.cfg.MODEL.ROI_RELATION_HEAD.SHARE_BOX_FEATURE_EXTRACTOR:
if self.cfg.MODEL.ROI_RELATION_HEAD.SEPERATE_SO_FEATURE_EXTRACTOR:
self.subj_feature_extractor = make_roi_relation_box_feature_extractor(cfg, feature_dim)
self.obj_feature_extractor = make_roi_relation_box_feature_extractor(cfg, feature_dim)
else:
self.obj_feature_extractor = make_roi_relation_box_feature_extractor(cfg, feature_dim)
def __init__(self, cfg):
super(GeneralizedRCNN, self).__init__()
self.backbone = build_backbone(cfg)
self.rpn = build_rpn(cfg, self.backbone.out_channels)
self.roi_heads = build_roi_heads(cfg, self.backbone.out_channels)
self.feature_extractor = make_roi_box_feature_extractor(
cfg, self.backbone.out_channels)
def __init__(self, cfg):
super(PanopticModel, self).__init__()
self.backbone = build_backbone(cfg)
self.loss = SegLoss()
self.training = True
self.rpn = build_rpn(cfg, self.backbone.out_channels)
self.roi_heads = build_roi_heads(cfg, self.backbone.out_channels)
self.panoptic = PanopticRCNN(
cfg, num_classes=cfg.MODEL.PANOPTIC.NUM_CLASSES)
def optimal_bias_init(cfg, bias):
device = torch.device(cfg.MODEL.DEVICE)
arch = request_arch(cfg)
if 'retinanet' in arch:
anchor_generator = make_anchor_generator_retinanet(cfg)
fg_iou, bg_iou = cfg.MODEL.RETINANET.FG_IOU_THRESHOLD, cfg.MODEL.RETINANET.BG_IOU_THRESHOLD
num_classes = cfg.MODEL.RETINANET.NUM_CLASSES - 1
num_anchors = len(cfg.MODEL.RETINANET.ASPECT_RATIOS) \
* cfg.MODEL.RETINANET.SCALES_PER_OCTAVE
else:
assert 'faster_rcnn' in arch
anchor_generator = make_anchor_generator(cfg)
fg_iou, bg_iou = cfg.MODEL.RPN.FG_IOU_THRESHOLD, cfg.MODEL.RPN.BG_IOU_THRESHOLD
num_classes = 1
num_anchors = anchor_generator.num_anchors_per_location()[0]
prior = load_prior(cfg, arch)
if prior is not None:
nn.init.constant_(bias, -log((1 - prior) / prior))
return
data_loader = make_init_data_loader(
cfg, is_distributed=True, images_per_batch=cfg.SOLVER.IMS_PER_BATCH)
proposal_matcher = Matcher(
fg_iou,
bg_iou,
allow_low_quality_matches=True,
)
backbone = build_backbone(cfg).to(device)
num_fg, num_all = 0, 0
num_gpus = get_num_gpus()
for images, targets, _ in tqdm(data_loader):
images = images.to(device)
targets = [target.to(device) for target in targets]
h, w = images.tensors.shape[-2:]
if num_all == 0:
features = backbone(images.tensors)
n, c = features[0].shape[:2]
levels = len(features)
stride = int(h / features[0].shape[2])
features = [
torch.zeros(n,
c,
int(ceil(h / (stride * 2**i))),
int(ceil(w / (stride * 2**i))),
device=device) for i in range(levels)
]
anchors = anchor_generator(images, features)
anchors = [
cat_boxlist(anchors_per_image).to(device)
for anchors_per_image in anchors
]
for anchor, target in zip(anchors, targets):
match_quality_matrix = boxlist_iou(target, anchor)
matched_idxs = proposal_matcher(match_quality_matrix)
num_fg_per_image, num_bg_per_image = (matched_idxs >= 0).sum(), (
matched_idxs == Matcher.BELOW_LOW_THRESHOLD).sum()
num_fg += num_fg_per_image
num_all += num_fg_per_image + num_bg_per_image
fg_all_ratio = reduce_div(num_fg.float(), num_all.float()).item()
prior = fg_all_ratio / num_classes
nn.init.constant_(bias, -log((1 - prior) / prior))
if torch.cuda.current_device() == 0:
save_prior(cfg, prior, arch)
def __init__(self, cfg):
super(GeneralizedRCNN, self).__init__()
self.backbone = build_backbone(cfg)
self.rpn = build_rpn(cfg, self.backbone.out_channels)
