def __init__(self, phase, conf):
super(RPN, self).__init__()
train = phase.lower() == 'train'
self.base = resnet.ResNetDilate(num_layer=50)
self.depthnet = resnet.ResNetDilate(num_layer=50)
# settings
self.phase = phase
self.num_classes = len(conf['lbls']) + 1
self.num_anchors = conf['anchors'].shape[0]
self.prev_num = 1 if conf.prev_num is None else conf.prev_num
self.dropout_rate = 0.2 if conf.dropout_rate is None else conf.dropout_rate
self.prop_feats = nn.Sequential(
nn.Conv2d(2048, 512, 3, padding=1),
nn.ReLU(inplace=True),
)
self.channel_attention_1 = ResidualGroup(RCAB,
n_resblocks=12,
n_feat=512,
kernel_size=3,
reduction=16,
act=nn.LeakyReLU(0.2, True),
norm=False)
self.channel_attention_2 = ResidualGroup(RCAB,
n_resblocks=12,
n_feat=512,
kernel_size=3,
reduction=16,
act=nn.LeakyReLU(0.2, True),
norm=False)
self.channel_attention_3 = ResidualGroup(RCAB,
n_resblocks=12,
n_feat=512,
kernel_size=3,
reduction=16,
act=nn.LeakyReLU(0.2, True),
norm=False)
self.prop_feats_multi = nn.Sequential(
nn.Conv2d(2048 * 3, 512, 3, padding=1),
nn.ReLU(inplace=True),
)
self.dropout = nn.Dropout(p=self.dropout_rate)
self.dropout_channel = nn.Dropout2d(p=0.3)
# outputs
self.cls = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_classes * self.num_anchors, 1)
# motion
self.motion_x = nn.Conv2d(self.prop_feats_multi[0].out_channels,
self.num_anchors, 1)
self.motion_y = nn.Conv2d(self.prop_feats_multi[0].out_channels,
self.num_anchors, 1)
self.motion_z = nn.Conv2d(self.prop_feats_multi[0].out_channels,
self.num_anchors, 1)
# bbox 2d
self.bbox_x = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_y = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_w = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_h = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
# bbox 3d
self.bbox_x3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_y3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_z3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_w3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_h3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_l3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_rY3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.softmax = nn.Softmax(dim=1)
self.feat_stride = conf.feat_stride
self.feat_size = calc_output_size(np.array(conf.crop_size),
self.feat_stride)
self.rois = locate_anchors(conf.anchors,
self.feat_size,
conf.feat_stride,
convert_tensor=True)
self.rois = self.rois.type(torch.cuda.FloatTensor)
self.anchors = conf.anchors
motion_scale = np.load(
'./lib/anchor_motion_scale_multi_objects_split1.npy')
self.motion_scale = torch.from_numpy(motion_scale).unsqueeze(
0).unsqueeze(2).unsqueeze(3)
def __init__(self, phase, conf):
super(RPN, self).__init__()
self.base = resnet.ResNetDilate(conf.base_model)
self.adaptive_diated = conf.adaptive_diated
self.dropout_position = conf.dropout_position
self.use_dropout = conf.use_dropout
self.drop_channel = conf.drop_channel
self.use_corner = conf.use_corner
self.corner_in_3d = conf.corner_in_3d
self.deformable = conf.deformable
if conf.use_rcnn_pretrain:
# print(self.base.state_dict().keys())
if conf.base_model == 101:
pretrained_model = torch.load(
'faster_rcnn_1_10_14657.pth')['model']
rename_dict = {
'RCNN_top.0': 'layer4',
'RCNN_base.0': 'conv1',
'RCNN_base.1': 'bn1',
'RCNN_base.2': 'relu',
'RCNN_base.3': 'maxpool',
'RCNN_base.4': 'layer1',
'RCNN_base.5': 'layer2',
'RCNN_base.6': 'layer3'
}
change_dict = {}
for item in pretrained_model.keys():
for rcnn_name in rename_dict.keys():
if rcnn_name in item:
change_dict[item] = item.replace(
rcnn_name, rename_dict[rcnn_name])
break
pretrained_model = {
change_dict[k]: v
for k, v in pretrained_model.items() if k in change_dict
}
self.base.load_state_dict(pretrained_model)
elif conf.base_model == 50:
pretrained_model = torch.load(
'res50_faster_rcnn_iter_1190000.pth',
map_location=lambda storage, loc: storage)
pretrained_model = {
k.replace('resnet.', ''): v
for k, v in pretrained_model.items() if 'resnet' in k
}
# print(pretrained_model.keys())
self.base.load_state_dict(pretrained_model)
self.depthnet = resnet.ResNetDilate(50)
if self.adaptive_diated:
self.adaptive_layers = nn.Sequential(
nn.AdaptiveMaxPool2d(3),
nn.Conv2d(512, 512 * 3, 3, padding=0),
)
self.adaptive_softmax = nn.Softmax(dim=3)
if self.deformable:
self.deform_layer = DeformConv2d(512,
512,
3,
padding=1,
bias=False,
modulation=True)
# settings
self.phase = phase
self.num_classes = len(conf['lbls']) + 1
self.num_anchors = conf['anchors'].shape[0]
self.prop_feats = nn.Sequential(
nn.Conv2d(2048, 512, 3, padding=1),
nn.ReLU(inplace=True),
)
if self.use_dropout:
self.dropout = nn.Dropout(p=conf.dropout_rate)
if self.drop_channel:
self.dropout_channel = nn.Dropout2d(p=0.3)
# outputs
self.cls = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_classes * self.num_anchors, 1)
# bbox 2d
self.bbox_x = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_y = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_w = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_h = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
# bbox 3d
self.bbox_x3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_y3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_z3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_w3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_h3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_l3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_rY3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
if self.corner_in_3d:
self.bbox_3d_corners = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors * 18,
1) # 2 * 8 + 2
self.bbox_vertices = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors * 24, 1) # 3 * 8
elif self.use_corner:
self.bbox_vertices = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors * 24, 1)
self.softmax = nn.Softmax(dim=1)
self.feat_stride = conf.feat_stride
self.feat_size = calc_output_size(np.array(conf.crop_size),
self.feat_stride)
self.rois = locate_anchors(conf.anchors,
self.feat_size,
conf.feat_stride,
convert_tensor=True)
self.rois = self.rois.type(torch.cuda.FloatTensor)
self.anchors = conf.anchors
def __init__(self, phase, conf):
super(RPN, self).__init__()
self.base = resnet.ResNetDilate(
conf.base_model
) # 去掉最后全连接层和池化层并修改layer4中部分网络的预训练好参数的ResNet-50,作为特征提取网络的主干
self.adaptive_diated = conf.adaptive_diated # True
self.dropout_position = conf.dropout_position # 'early' # 'early' 'late' 'adaptive'
self.use_dropout = conf.use_dropout # True
self.drop_channel = conf.drop_channel # True
self.use_corner = conf.use_corner # False
self.corner_in_3d = conf.corner_in_3d # False
self.deformable = conf.deformable # False
# 遮挡模块:
self.occlusion = conf.occlusion if "occlusion" in conf else False
self.threshold = conf.threshold if "threshold" in conf else 1
if conf.use_rcnn_pretrain: # False
# print(self.base.state_dict().keys())
if conf.base_model == 101:
pretrained_model = torch.load(
'faster_rcnn_1_10_14657.pth')['model']
rename_dict = {
'RCNN_top.0': 'layer4',
'RCNN_base.0': 'conv1',
'RCNN_base.1': 'bn1',
'RCNN_base.2': 'relu',
'RCNN_base.3': 'maxpool',
'RCNN_base.4': 'layer1',
'RCNN_base.5': 'layer2',
'RCNN_base.6': 'layer3'
}
change_dict = {}
for item in pretrained_model.keys():
for rcnn_name in rename_dict.keys():
if rcnn_name in item:
change_dict[item] = item.replace(
rcnn_name, rename_dict[rcnn_name])
break
pretrained_model = {
change_dict[k]: v
for k, v in pretrained_model.items() if k in change_dict
}
self.base.load_state_dict(pretrained_model)
elif conf.base_model == 50:
pretrained_model = torch.load(
'res50_faster_rcnn_iter_1190000.pth',
map_location=lambda storage, loc: storage)
pretrained_model = {
k.replace('resnet.', ''): v
for k, v in pretrained_model.items() if 'resnet' in k
}
# print(pretrained_model.keys())
self.base.load_state_dict(pretrained_model)
self.depthnet = resnet.ResNetDilate(50)
if self.adaptive_diated: # True
self.adaptive_softmax = nn.Softmax(dim=3)
self.adaptive_layers = nn.Sequential(
nn.AdaptiveMaxPool2d(
3
), # 常见的池化参数为kernel_size,‘’图像‘’的输出尺寸另算（可变）；AdaptiveMaxPool2d()参数为output_size，对于任何输入大小，其‘’图像‘’输出尺寸就是output_siz；
nn.Conv2d(512, 512 * 3, 3, padding=0),
) # 用在layer2与layer3之间; # size: ->(512,512*3,1,1)
self.adaptive_bn = nn.BatchNorm2d(512)
self.adaptive_relu = nn.ReLU(inplace=True)
self.adaptive_layers1 = nn.Sequential(
nn.AdaptiveMaxPool2d(3),
nn.Conv2d(1024, 1024 * 3, 3, padding=0),
) # # 用在layer3与layer4之间; # size: ->(1024,1024*3,1,1)
self.adaptive_bn1 = nn.BatchNorm2d(1024)
self.adaptive_relu1 = nn.ReLU(inplace=True)
if self.deformable: # False
self.deform_layer = DeformConv2d(512,
512,
3,
padding=1,
bias=False,
modulation=True)
# settings
self.phase = phase
self.num_classes = len(conf['lbls']) + 1 # +1的理解：不属于3个中的任何一类？
self.num_anchors = conf['anchors'].shape[0]
self.prop_feats = nn.Sequential(
nn.Conv2d(2048, 512, 3, padding=1),
nn.ReLU(inplace=True),
) # [N,2018,H,W]->[N,512,H,W]
if self.use_dropout: # True
self.dropout = nn.Dropout(p=conf.dropout_rate)
if self.drop_channel: # True
self.dropout_channel = nn.Dropout2d(
p=0.3) # 通常输入为 nn.Conv2d modules.
# outputs
self.cls = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_classes * self.num_anchors,
1) # [N,512,H,W]->[N,36*4,H,W]
# bbox 2d
self.bbox_x = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors,
1) # # [N,512,H,W]->[N,36,H,W]
self.bbox_y = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_w = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_h = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
# bbox 3d
self.bbox_x3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_y3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_z3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_w3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_h3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_l3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.bbox_rY3d = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
# 遮挡模块:
if self.occlusion:
self.occ_correct = nn.Conv2d(self.prop_feats[0].out_channels, 7, 1)
self.occ_correct = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors, 1)
self.avg_pool = nn.AdaptiveMaxPool2d((1, 1))
self.fc1 = nn.Linear(self.occ_correct.out_channels, 133)
self.fc2 = nn.Linear(133, 7)
if self.corner_in_3d: # False
self.bbox_3d_corners = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors * 18,
1) # 2 * 8 + 2
self.bbox_vertices = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors * 24, 1) # 3 * 8
elif self.use_corner: # False
self.bbox_vertices = nn.Conv2d(self.prop_feats[0].out_channels,
self.num_anchors * 24, 1)
self.softmax = nn.Softmax(dim=1)
self.feat_stride = conf.feat_stride # 16
self.feat_size = calc_output_size(np.array(
conf.crop_size), self.feat_stride) # feat_size=[32,106]；
self.rois = locate_anchors(conf.anchors,
self.feat_size,
conf.feat_stride,
convert_tensor=True)
self.rois = self.rois.type(torch.cuda.FloatTensor)
self.anchors = conf.anchors