def init_modules(self):
self.feature_extractor = feature_extractors.build(
self.feature_extractor_config)
self.rpn_model = detectors.build(self.rpn_config)
if self.pooling_mode == 'align':
self.rcnn_pooling = ROIAlign(
(self.pooling_size, self.pooling_size), 1.0 / 16.0, 2)
# note that roi extractor is shared but heads not
# self.rcnn_cls_preds = nn.ModuleList(
# [nn.Linear(2048, self.n_classes) for _ in range(self.num_stages)])
in_channels = 2048
# construct many branches for each attr of instance
branches = {}
for attr in self.instance_info:
col = self.instance_info[attr]
branches[attr] = nn.ModuleList([nn.Linear(in_channels, self.n_classes) for _ in range(self.num_stages)])
self.branches = nn.ModuleDict(branches)
# if self.class_agnostic:
# rcnn_bbox_pred = nn.Linear(in_channels, 4)
# else:
# rcnn_bbox_pred = nn.Linear(in_channels, 4 * self.n_classes)
# self.rcnn_bbox_preds = nn.ModuleList(
# [rcnn_bbox_pred for _ in range(self.num_stages)])
# loss module
# if self.use_focal_loss:
# self.rcnn_cls_loss = FocalLoss(self.n_classes)
# else:
self.rcnn_cls_loss = nn.CrossEntropyLoss(reduce=False)
self.rcnn_bbox_loss = nn.modules.SmoothL1Loss(reduce=False)
def init_modules(self):
self.feature_extractor = feature_extractors.build(
self.feature_extractor_config)
self.rpn_model = detectors.build(self.rpn_config)
if self.pooling_mode == 'align':
self.rcnn_pooling = ROIAlign(
(self.pooling_size, self.pooling_size), 1.0 / 16.0, 2)
# note that roi extractor is shared but heads not
self.rcnn_cls_preds = nn.ModuleList(
[nn.Linear(2048, self.n_classes) for _ in range(self.num_stages)])
in_channels = 2048
if self.class_agnostic:
rcnn_bbox_pred = nn.Linear(in_channels, 4)
else:
rcnn_bbox_pred = nn.Linear(in_channels, 4 * self.n_classes)
self.rcnn_bbox_preds = nn.ModuleList(
[rcnn_bbox_pred for _ in range(self.num_stages)])
self.rcnn_orient_preds = nn.ModuleList(
[nn.Linear(2048, 4) for _ in range(self.num_stages)])
self.rcnn_dim_preds = nn.ModuleList(
[nn.Linear(2048, 3) for _ in range(self.num_stages)])
# loss module
# if self.use_focal_loss:
# self.rcnn_cls_loss = FocalLoss(self.n_classes)
# else:
self.rcnn_cls_loss = nn.CrossEntropyLoss(reduction='none')
self.rcnn_bbox_loss = nn.modules.SmoothL1Loss(reduction='none')
self.rcnn_orient_loss = OrientationLoss()
def init_modules(self):
self.feature_extractor = feature_extractors.build(
self.feature_extractor_config)
self.rpn_model = detectors.build(self.rpn_config)
in_channels = 1024
self.rcnn_cls_preds = nn.ModuleList([
nn.Linear(in_channels, self.n_classes)
for _ in range(self.num_stages)
])
if self.class_agnostic:
rcnn_bbox_pred = nn.Linear(in_channels, 4)
else:
rcnn_bbox_pred = nn.Linear(in_channels, 4 * self.n_classes)
self.rcnn_bbox_preds = nn.ModuleList(
[rcnn_bbox_pred for _ in range(self.num_stages)])
# loss module
if self.use_focal_loss:
self.rcnn_cls_loss = FocalLoss(self.n_classes, gamma=2, alpha=0.75)
else:
self.rcnn_cls_loss = nn.CrossEntropyLoss(reduction='none')
self.rcnn_bbox_loss = nn.modules.SmoothL1Loss(reduction='none')
# TODO add feat scale adaptive roi pooling
self.rcnn_pooling = AdaptiveROIAlign(
(self.pooling_size, self.pooling_size), 2)
def init_modules(self):
self.feature_extractor = feature_extractors.build(
self.feature_extractor_config)
self.rpn_model = detectors.build(self.rpn_config)
# self.rcnn_pooling = ROIAlign(
# (self.pooling_size, self.pooling_size), 1.0 / 16.0, 2)
self.rcnn_pooling = AdaptiveROIAlign(
(self.pooling_size, self.pooling_size), 2)
# construct many branches for each attr of instance
branches = {}
for attr in self.instance_info:
num_channels = self.instance_info[attr].num_channels
branches[attr] = nn.ModuleList([
nn.Linear(self.in_channels, num_channels)
for _ in range(self.num_stages)
])
self.branches = nn.ModuleDict(branches)
self.rcnn_cls_loss = nn.CrossEntropyLoss(reduce=False)
self.rcnn_bbox_loss = nn.modules.SmoothL1Loss(reduce=False)
def init_modules(self):
# self.feature_extractor = ResNetFeatureExtractor(
# self.feature_extractor_config)
self.feature_extractor = feature_extractors.build(
self.feature_extractor_config)
self.rpn_model = RPNModel(self.rpn_config)
if self.pooling_mode == 'align':
self.rcnn_pooling = ROIAlign(
(self.pooling_size, self.pooling_size), 1.0 / 16.0, 2)
elif self.pooling_mode == 'ps':
self.rcnn_pooling = PSRoIPool(7, 7, 1.0 / 16, 7, self.n_classes)
elif self.pooling_mode == 'psalign':
raise NotImplementedError('have not implemented yet!')
elif self.pooling_mode == 'deformable_psalign':
raise NotImplementedError('have not implemented yet!')
# self.rcnn_cls_pred = nn.Conv2d(2048, self.n_classes, 3, 1, 1)
self.rcnn_cls_preds = nn.Linear(self.in_channels, self.n_classes)
if self.class_agnostic:
self.rcnn_bbox_preds = nn.Linear(self.in_channels, 4)
else:
self.rcnn_bbox_preds = nn.Linear(self.in_channels,
4 * self.n_classes)
# loss module
if self.use_focal_loss:
self.rcnn_cls_loss = FocalLoss(self.n_classes)
else:
self.rcnn_cls_loss = functools.partial(F.cross_entropy,
reduce=False)
self.rcnn_bbox_loss = nn.modules.SmoothL1Loss(reduce=False)
# self.rcnn_3d_pred = nn.Linear(c, 3 + 4 + 11 + 2 + 1)
if self.class_agnostic_3d:
self.rcnn_3d_pred = nn.Linear(self.in_channels, 3 + 5)
else:
self.rcnn_3d_pred = nn.Linear(self.in_channels,
3 * self.n_classes + 5)
self.rcnn_3d_loss = OrientationLoss(split_loss=True)
def init_modules(self):
super().init_modules()
self.feature_extractor = feature_extractors.build(
self.feature_extractor_config)