def __init__(
self,
in_channels=512,
mid_channels=512,
ratios=[0.5, 1, 2],
anchor_scales=[8, 16, 32],
feat_stride=16,
mode="training",
):
super(RegionProposalNetwork, self).__init__()
self.anchor_base = generate_anchor_base(anchor_scales=anchor_scales,
ratios=ratios)
# 步长，压缩的倍数
self.feat_stride = feat_stride
self.proposal_layer = ProposalCreator(mode)
# 每一个网格上默认先验框的数量
n_anchor = self.anchor_base.shape[0]
# 先进行一个3x3的卷积
self.conv1 = nn.Conv2d(in_channels, mid_channels, 3, 1, 1)
# 分类预测先验框内部是否包含物体
self.score = nn.Conv2d(mid_channels, n_anchor * 2, 1, 1, 0)
# 回归预测对先验框进行调整
self.loc = nn.Conv2d(mid_channels, n_anchor * 4, 1, 1, 0)
normal_init(self.conv1, 0, 0.01)
normal_init(self.score, 0, 0.01)
normal_init(self.loc, 0, 0.01)
def __init__(
self, in_channels=512, mid_channels=512, ratios=[0.5, 1, 2],
anchor_scales=[8, 16, 32], feat_stride=16,
mode = "training",
):
super(RegionProposalNetwork, self).__init__()
self.feat_stride = feat_stride
self.proposal_layer = ProposalCreator(mode)
#-----------------------------------------#
# 生成基础先验框，shape为[9, 4]
#-----------------------------------------#
self.anchor_base = generate_anchor_base(anchor_scales=anchor_scales, ratios=ratios)
n_anchor = self.anchor_base.shape[0]
#-----------------------------------------#
# 先进行一个3x3的卷积，可理解为特征整合
#-----------------------------------------#
self.conv1 = nn.Conv2d(in_channels, mid_channels, 3, 1, 1)
#-----------------------------------------#
# 分类预测先验框内部是否包含物体
#-----------------------------------------#
self.score = nn.Conv2d(mid_channels, n_anchor * 2, 1, 1, 0)
#-----------------------------------------#
# 回归预测对先验框进行调整
#-----------------------------------------#
self.loc = nn.Conv2d(mid_channels, n_anchor * 4, 1, 1, 0)
#--------------------------------------#
# 对FPN的网络部分进行权值初始化
#--------------------------------------#
normal_init(self.conv1, 0, 0.01)
normal_init(self.score, 0, 0.01)
normal_init(self.loc, 0, 0.01)
def __init__(self,
in_channels=512,
mid_channels=512,
ratios=[0.5, 1, 2],
scales=[0.5, 1, 2],
feat_stride=16):
super(RegionProposalNetwork, self).__init__()
# prepare anchor base
self.anchor_base = generate_anchor_base(side_length=16,
ratios=ratios,
scales=scales,
strides=feat_stride)
self.feat_stride = feat_stride
# network params
n_anchor = self.anchor_base.shape[0]
self.conv1 = nn.Conv2d(in_channels, mid_channels, 3, 1, 1)
self.score = nn.Conv2d(mid_channels, n_anchor * 2, 1, 1, 0)
self.loc = nn.Conv2d(mid_channels, n_anchor * 4, 1, 1, 0)
normal_init(self.conv1, 0, 0.01)
normal_init(self.score, 0, 0.01)
normal_init(self.loc, 0, 0.01)
def __init__(self,
in_channels=256,
mid_channels=256,
ratios=[0.5, 1., 2.],
anchor_scales=[8, 16, 32],
feat_stride=16,
mode="training"):
super(RPN, self).__init__()
self.base_anchor = generate_anchor_base(ratios=ratios,
anchor_scales=anchor_scales)
self.K = self.base_anchor.shape[
0] # number of anchors in a anchor base
self.feat_stride = feat_stride
self.proposal_layer = region_proposal(mode)
# layers after conv5
self.conv1 = nn.Conv2d(in_channels=in_channels,
out_channels=mid_channels,
kernel_size=3,
stride=1,
padding=1)
self.cls = nn.Conv2d(in_channels=mid_channels,
out_channels=self.K * 2,
kernel_size=1,
stride=1,
padding=0)
self.reg = nn.Conv2d(in_channels=mid_channels,
out_channels=self.K * 4,
kernel_size=1,
stride=1,
padding=0)
# channels inistalization
normal_init(self.conv1, 0, 0.01)
normal_init(self.cls, 0, 0.01)
normal_init(self.reg, 0, 0.01)
batch_size = 1
start_epoch, max_epoch = 0, 30
data_root = cfg_net['data_path']
train_file = os.path.join(cfg_net['train_data_path'], "VOC2007_train.txt")
train_transform = Compose([ToTensor(), RandomHorizontalFlip(0.5)])
# step 1: data
train_set = datasets.VOC2007Dataset(data_root, vocfile=train_file, transforms=train_transform)
train_loader = DataLoader(train_set, batch_size=batch_size, collate_fn=collate_fn)
base_size = 16
ratios = [0.5, 1, 2]
anchor_scales = [8, 16, 32]
anchors_gen = Anchors.AnchorTargetCreator()
anchor_base = Anchors.generate_anchor_base(base_size=base_size, anchor_scales=anchor_scales, ratios=ratios)
net = mobilenet_v2(num_classes=2, width_mult=0.35, inverted_residual_setting=None, round_nearest=8).to(device)
rpn_net = RPN(1280, 512, 9).to(device)
# inputs = torch.zeros((2, 3, 128, 128))
# cls = net(inputs)
# feature_map = net.features(inputs)
# rpn_class, rpn_prob, rpn_bbox = rpn_net(feature_map)
classify_loss = rpn_loss.ClassifyLoss()
for data in train_loader:
images, targets = data
images_size = list(images.shape[2:])