def __init__(self, faster_rcnn):
super(FasterRCNNTrainer, self).__init__()
self.faster_rcnn = faster_rcnn
self.rpn_sigma = opt.rpn_sigma
self.roi_sigma = opt.roi_sigma
self.rpn_pen = opt.rpn_pen
self.roi_pen = opt.roi_pen
# target creator create gt_bbox gt_label etc as training targets.
# FLAG: add params
# Initail best: pos 0.2, neg 0.1
self.anchor_target_creator = AnchorTargetCreator(pos_ratio=0.5,
pos_iou_thresh=0.7,
neg_iou_thresh=0.3)
# Initial best: pos 0.2, neg 0.2
self.proposal_target_creator = ProposalTargetCreator(pos_ratio=0.5,
pos_iou_thresh=0.5,
neg_iou_thresh_hi=0.5)
self.loc_normalize_mean = faster_rcnn.loc_normalize_mean
self.loc_normalize_std = faster_rcnn.loc_normalize_std
self.optimizer = self.faster_rcnn.get_optimizer()
# visdom wrapper
self.vis = Visualizer(env=opt.env)
# indicators for training status
self.rpn_cm = ConfusionMeter(2)
self.roi_cm = ConfusionMeter(4)
self.meters = {k: AverageValueMeter() for k in LossTuple._fields} # average loss
def __init__(self, faster_rcnn):
super(FasterRCNNTrainer, self).__init__()
self.faster_rcnn = faster_rcnn
self.rpn_sigma = opt.rpn_sigma #是在_faster_rcnn_loc_loss调用用来计算位置损失函数用到的超参数，
self.roi_sigma = opt.roi_sigma
self.anchor_target_creator = AnchorTargetCreator(
) #从上万个anchor中挑选256个来训练rpn，其中正样本不超过128
self.proposal_target_creator = ProposalTargetCreator(
) #从rpn给的2000个框中挑出128个来训练roihead，其中正样本不超过32个
self.loc_normalize_mean = faster_rcnn.loc_normalize_mean
self.loc_normalize_std = faster_rcnn.loc_normalize_std
self.optimizer = self.faster_rcnn.get_optimizer()
#可视化
self.vis = Visualizer(env=opt.env)
#验证预测值和真实值的精度
self.rpn_cm = ConfusionMeter(
2) #混淆矩阵，就是验证预测值与真实值精确度的矩阵ConfusionMeter(2)括号里的参数指的是类别数
self.roi_cm = ConfusionMeter(opt.class_num + 1)
self.meters = {k: AverageValueMeter()
for k in LossTuple._fields} #验证平均loss
def __init__(self, faster_rcnn):
# 继承父模块的初始化
super(FasterRCNNTrainer, self).__init__()
self.faster_rcnn = faster_rcnn
# 下面2个参数是在_faster_rcnn_loc_loss调用用来计算位置损失函数用到的超参数
self.rpn_sigma = opt.rpn_sigma
self.roi_sigma = opt.roi_sigma
# target creator create gt_bbox gt_label etc as training targets.
# 用于从20000个候选anchor中产生256个anchor进行二分类和位置回归，也就是
# 为rpn网络产生的预测位置和预测类别提供真正的ground_truth标准
self.anchor_target_creator = AnchorTargetCreator()
# AnchorTargetCreator和ProposalTargetCreator是为了生成训练的目标
# （或称ground truth），只在训练阶段用到，ProposalCreator是RPN为Fast
# R-CNN生成RoIs，在训练和测试阶段都会用到。所以测试阶段直接输进来300
# 个RoIs，而训练阶段会有AnchorTargetCreator的再次干预
self.proposal_target_creator = ProposalTargetCreator()
# (0., 0., 0., 0.)
self.loc_normalize_mean = faster_rcnn.loc_normalize_mean
# (0.1, 0.1, 0.2, 0.2)
self.loc_normalize_std = faster_rcnn.loc_normalize_std
# SGD
self.optimizer = self.faster_rcnn.get_optimizer()
# 可视化，vis_tool.py
self.vis = Visualizer(env=opt.env)
# 混淆矩阵，就是验证预测值与真实值精确度的矩阵ConfusionMeter
# (2)括号里的参数指的是类别数
self.rpn_cm = ConfusionMeter(2)
# roi的类别有21种（20个object类+1个background）
self.roi_cm = ConfusionMeter(21)
# 平均损失
self.meters = {k: AverageValueMeter()
for k in LossTuple._fields} # average loss
def __init__(self, fasterrcnn):
super(Trainer, self).__init__()
self.fasterrcnn = fasterrcnn
self.rpn_sigma = 3.
self.roi_sigma = 1.
self.anchor_target_creator = AnchorTargetCreator()
self.proposal_target_creator = ProposalTargetCreator()
self.loc_normalize_mean = fasterrcnn.loc_normalize_mean
self.loc_normalize_std = fasterrcnn.loc_normalize_std
lr = 1e-3
params = []
for key, value in dict(fasterrcnn.named_parameters()).items():
if value.requires_grad:
if 'bias' in key:
params += [{
'params': [value],
'lr': lr * 2,
'weight_decay': 0
}]
else:
params += [{
'params': [value],
'lr': lr,
'weight_decay': 0.0005
}]
self.optimizer = torch.optim.SGD(params, momentum=0.9)
def __init__(self, faster_rcnn):
# 继承父模块的初始化
super(FasterRCNNTrainer, self).__init__()
self.faster_rcnn = faster_rcnn
self.rpn_sigma = opt.rpn_sigma
self.roi_sigma = opt.roi_sigma # 超参：在_faster_rcnn_loc_loss调用用来计算位置损失函数
# target creator create gt_bbox gt_label etc as training targets.
# 用于从20000个候选anchor中产生256个anchor进行二分类和位置回归，也就是为rpn网络产生的预测位置和预测类别提供真正的ground_truth标准
self.anchor_target_creator = AnchorTargetCreator()
# AnchorTargetCreator和ProposalTargetCreator是为了生成训练的目标（或称ground truth），只在训练阶段用到，ProposalCreator是RPN为Fast R-CNN生成RoIs，在训练和测试阶段都会用到。所以测试阶段直接输进来300个RoIs，而训练阶段会有AnchorTargetCreator的再次干预。
self.proposal_target_creator = ProposalTargetCreator()
# (0., 0., 0., 0.)
self.loc_normalize_mean = faster_rcnn.loc_normalize_mean
# (0.1, 0.1, 0.2, 0.2)
self.loc_normalize_std = faster_rcnn.loc_normalize_std
self.optimizer = self.faster_rcnn.get_optimizer() # SGD
# visdom wrapper
self.vis = Visualizer(env=opt.env) # 可视化工具
# indicators for training status
# 混淆矩阵，验证预测值和真实值精确度，2为类别数
self.rpn_cm = ConfusionMeter(2)
#
self.roi_cm = ConfusionMeter(21)
self.meters = {k: AverageValueMeter()
for k in LossTuple._fields} # average loss
def __init__(self,
faster_rcnn,
attacker=None,
layer_idx=None,
attack_mode=False):
super(BRFasterRcnnTrainer, self).__init__()
self.faster_rcnn = faster_rcnn
self.attacker = attacker
self.layer_idx = layer_idx
self.rpn_sigma = opt.rpn_sigma
self.roi_sigma = opt.roi_sigma
self.attack_mode = attack_mode
self.anchor_target_creator = AnchorTargetCreator()
self.proposal_target_creator = ProposalTargetCreator()
self.loc_normalize_mean = faster_rcnn.loc_normalize_mean
self.loc_normalize_std = faster_rcnn.loc_normalize_std
self.optimizer = self.faster_rcnn.get_optimizer()
self.vis = Visualizer(env=opt.env)
self.rpn_cm = ConfusionMeter(2)
self.roi_cm = ConfusionMeter(21)
self.meters = {k: AverageValueMeter() for k in LossTuple._fields}
self.BR_meters = {k: AverageValueMeter() for k in LossTupleBR._fields}
def __init__(self, faster_rcnn):
super(FasterRCNNTrainer, self).__init__()
#传入的是FasterRCNNVGG16模型，继承了FasterRCNN模型，而参数根据说明 是FasterRCNN模型
#即初始化的是FasterRCNN模型
#FasterRCNN模型是父类 FasterRCNNVGG16模型是子类
self.faster_rcnn = faster_rcnn
#sigma for l1_smooth_loss
self.rpn_sigma = opt.rpn_sigma
self.roi_sigma = opt.roi_sigma
# target creator create gt_bbox gt_label etc as training targets.
#目标框creator 目标是产生 真实的bbox 类别标签等
#将真实的bbox分配给锚点
self.anchor_target_creator = AnchorTargetCreator()
self.proposal_target_creator = ProposalTargetCreator()
#得到faster网络权重，均值 和方差
self.loc_normalize_mean = faster_rcnn.loc_normalize_mean
self.loc_normalize_std = faster_rcnn.loc_normalize_std
#得到faster网络的优化器
self.optimizer = self.faster_rcnn.get_optimizer()
# visdom wrapper
self.vis = Visualizer(env=opt.env)
# indicators for training status
#训练状态指标 两个混淆矩阵 2×2（前景后景） 21×21（20类+背景）
self.rpn_cm = ConfusionMeter(2)
self.roi_cm = ConfusionMeter(21)
self.meters = {k: AverageValueMeter()
for k in LossTuple._fields} # average loss 平均损失
def __init__(self, faster_rcnn, logger):
super(FasterRCNNTrainer, self).__init__()
self.logger = logger
self.faster_rcnn = faster_rcnn
self.rpn_sigma = opt.rpn_sigma
self.roi_sigma = opt.roi_sigma
self.losses = AverageVal()
# target creator create gt_bbox gt_label etc as training targets.
self.anchor_target_creator = AnchorTargetCreator()
self.proposal_target_creator = ProposalTargetCreator()
self.loc_normalize_mean = faster_rcnn.loc_normalize_mean
self.loc_normalize_std = faster_rcnn.loc_normalize_std
self.optimizer = self.faster_rcnn.get_optimizer()
# visdom wrapper
self.vis = Visualizer(env=opt.env)
# indicators for training status
self.rpn_cm = ConfusionMeter(2)
self.roi_cm = ConfusionMeter(21)
self.meters = {k: AverageValueMeter()
for k in LossTuple._fields} # average loss
def __init__(self, faster_rcnn):
super(FasterRCNNTrainer, self).__init__()
self.faster_rcnn = faster_rcnn
#在faster_rcnn_loc_losss中调用，用来计算位置损失函数时用到的超参
self.rpn_sigma = opt.rpn_sigma
self.roi_sigma = opt.roi_sigma
# target creator create gt_bbox gt_label etc as training targets.
#用于从20000个候选anchor中产生256个anchor进行二分类和位置回归,用于rpn的训练
self.anchor_target_creator = AnchorTargetCreator()
#从2000个筛选出的ROIS中再次选出128个ROIs用于ROIhead训练
self.proposal_target_creator = ProposalTargetCreator()
#定义位置信息的均值方差。因为送入网络训练的位置信息需全部归一化处理
self.loc_normalize_mean = faster_rcnn.loc_normalize_mean
self.loc_normalize_std = faster_rcnn.loc_normalize_std
self.optimizer = self.faster_rcnn.get_optimizer()
# visdom wrapper
self.vis = Visualizer(env=opt.env)
# indicators for training status
self.rpn_cm = ConfusionMeter(2)
self.roi_cm = ConfusionMeter(21)
self.meters = {k: AverageValueMeter()
for k in LossTuple._fields} # average loss
def __init__(self, faster_rcnn):
super(FasterRCNNTrainer, self).__init__()
self.faster_rcnn = faster_rcnn
self.rpn_sigma = opt.rpn_sigma
self.roi_sigma = opt.roi_sigma
# target creator create gt_bbox gt_label etc as training targets.
self.anchor_target_creator = AnchorTargetCreator()
self.proposal_target_creator = ProposalTargetCreator()
self.loc_normalize_mean = faster_rcnn.loc_normalize_mean
self.loc_normalize_std = faster_rcnn.loc_normalize_std
self.optimizer = self.faster_rcnn.get_optimizer()
# visdom wrapper
self.vis = Visualizer(env=opt.env)
def __init__(self, faster_rcnn):
super(FasterRCNNTrainer, self).__init__()
self.faster_rcnn = faster_rcnn
self.rpn_sigma = opt.rpn_sigma
self.roi_sigma = opt.roi_sigma
self.anchor_target_creator = AnchorTargetCreator()
# 训练RPN的时候，从20000个anchor中选择256个进行训练，
# 以使得正负样本比例大概是1:1. 同时给出训练的 位置参数目标。
self.proposal_target_creator = ProposalTargetCreator()
# 在训练RoIHead/Fast R-CNN的时候，从 2000 个 rois 中选择 128 个用以训练。
# 同时给定训练目标, 返回（sample_RoI, gt_RoI_loc, gt_RoI_label）
self.loc_normalize_mean = faster_rcnn.loc_normalize_mean
self.loc_normalize_std = faster_rcnn.loc_normalize_std
self.optimizer = self.faster_rcnn.get_optimizer()
def __init__(self, siam_reid):
super(Trainer, self).__init__()
self.siam_reid = siam_reid
self.rpn_sigma = opt.rpn_sigma
self.roi_sigma = opt.roi_sigma
self.anchor_target_creator = AnchorTargetCreator()
self.proposal_filter = ProposalFilter()
self.loc_normalize_mean = (0., 0., 0., 0.)
self.loc_normalize_std = (0.1, 0.1, 0.2, 0.2)
self.optimizer = self.siam_reid.get_optimizer()
self.vis = Visualizer(env=opt.env)
# self.rpn_cm = ConfusionMeter(2)
# self.roi_cm = ConfusionMeter(2)
self.meters = {k: AverageValueMeter() for k in LossTuple._fields}
def __init__(self, faster_rcnn):
super(FasterRCNNTrainer, self).__init__()
self.faster_rcnn = faster_rcnn
self.rpn_sigma = opt.rpn_sigma
self.roi_sigma = opt.roi_sigma
# target creator create gt_bbox gt_label etc as training targets.
self.anchor_target_creator = AnchorTargetCreator()
self.proposal_target_creator = ProposalTargetCreator()
self.loc_normalize_mean = faster_rcnn.loc_normalize_mean
self.loc_normalize_std = faster_rcnn.loc_normalize_std
self.optimizer = self.faster_rcnn.get_optimizer()
self.rpn_cm = ConfusionMeter(2)
self.roi_cm = ConfusionMeter(21)
self.meters = {k: AverageValueMeter() for k in LossTuple._fields} # average loss
def __init__(self, faster_rcnn):
super(FasterRCNNTrainer, self).__init__()
self.faster_rcnn = faster_rcnn # faster-rcnn的对象
self.rpn_sigma = opt.rpn_sigma # RPN的超参数 3.
self.roi_sigma = opt.roi_sigma # roi的超参数 1.
# target creator create gt_bbox gt_label etc as training targets.
# generate anchor 生成20000多个框 然后 RPN 训练阶段 会 挑 256个 anchor
# 用于和 bbox 一起计算 loc label部分会赋值 -1 是忽略 0是负样本 128个 1是正样本 128个
self.anchor_target_creator = AnchorTargetCreator()
# 从 训练 20000-->12000-->2000 个ROIS 挑选 128个 正负比例 1:3
# 测试 12000-->6000>--》300
self.proposal_target_creator = ProposalTargetCreator()
self.loc_normalize_mean = faster_rcnn.loc_normalize_mean
self.loc_normalize_std = faster_rcnn.loc_normalize_std
self.optimizer = self.faster_rcnn.get_optimizer()
def __init__(self, faster_rcnn, lstm, final_predictor):
super(FasterRCNN_LSTM, self).__init__()
self.faster_rcnn = faster_rcnn
self.lstm = lstm
self.n_class = faster_rcnn.n_class
self.final_predictor = final_predictor
self.bce = nn.BCELoss()
self.rpn_sigma = opt.rpn_sigma
self.roi_sigma = opt.roi_sigma
# target creator create gt_bbox gt_label etc as training targets.
self.anchor_target_creator = AnchorTargetCreator()
self.proposal_target_creator = ProposalTargetCreator()
self.loc_normalize_mean = faster_rcnn.loc_normalize_mean
self.loc_normalize_std = faster_rcnn.loc_normalize_std
self.optimizer = self.faster_rcnn.get_optimizer()
def __init__(self, faster_rcnn):
super(FasterRCNNTrainer, self).__init__()
self.faster_rcnn = faster_rcnn #训练的网络
self.rpn_sigma = opt.rpn_sigma #计算rpn_loc_loss的超参数
self.roi_sigma = opt.roi_sigma #计算roi_loc_loss的超参数
# target creator create gt_bbox gt_label etc as training targets.
self.anchor_target_creator = AnchorTargetCreator() #rpn处使用的，实例化anchor和gt匹配函数
self.proposal_target_creator = ProposalTargetCreator() #roi处使用的，实例化提取roi正负样本函数
self.loc_normalize_mean = faster_rcnn.loc_normalize_mean #用于坐标归一化的mean
self.loc_normalize_std = faster_rcnn.loc_normalize_std #用于坐标归一化的std
self.optimizer = self.faster_rcnn.get_optimizer() #实例化得到optimizer函数
# visdom wrapper
self.vis = Visualizer(env=opt.env) #可视化内容，（跳过）
# indicators for training status
self.rpn_cm = ConfusionMeter(2) #可视化内容，（跳过）
self.roi_cm = ConfusionMeter(21) #可视化内容，（跳过）
self.meters = {k: AverageValueMeter() for k in LossTuple._fields} # average loss，可视化内容，（跳过）
def __init__(self, faster_rcnn):
super(FasterRCNNTrainer, self).__init__()
self.faster_rcnn = faster_rcnn
self.rpn_sigma = opt.rpn_sigma
self.roi_sigma = opt.roi_sigma
# target creator create gt_bbox gt_label etc as training targets.
self.anchor_target_creator = AnchorTargetCreator()#用于从20000个候选anchor中产生256个anchor进行二分类和位置回归，也就是为rpn网络产生的预测位置和预测类别提供真正的ground_truth标准
self.proposal_target_creator = ProposalTargetCreator()#AnchorTargetCreator和ProposalTargetCreator是为了生成训练的目标
self.loc_normalize_mean = faster_rcnn.loc_normalize_mean#
self.loc_normalize_std = faster_rcnn.loc_normalize_std
self.optimizer = self.faster_rcnn.get_optimizer()#SGD
# visdom wrapper
self.vis = Visualizer(env=opt.env)
# indicators for training status
self.rpn_cm = ConfusionMeter(2)#构造一个用于多类分类问题的混淆矩阵
self.roi_cm = ConfusionMeter(21)
self.meters = {k: AverageValueMeter() for k in LossTuple._fields} # average loss
def __init__(self, faster_rcnn):
super(FasterRCNNTrainer, self).__init__()
self.faster_rcnn = faster_rcnn
self.rpn_sigma = configurations.rpn_sigma
self.roi_sigma = configurations.roi_sigma
# target creator create gt_bbox gt_label etc as training targets.
self.anchor_target_creator = AnchorTargetCreator()
self.proposal_target_creator = ProposalTargetCreator()
self.loc_normalize_mean = faster_rcnn.loc_normalize_mean
self.loc_normalize_std = faster_rcnn.loc_normalize_std
lr = configurations.lr
params = []
for key, value in dict(faster_rcnn.named_parameters()).items():
if value.requires_grad:
if 'bias' in key:
params += [{
'params': [value],
'lr': lr * 2,
'weight_decay': 0
}]
else:
params += [{
'params': [value],
'lr': lr,
'weight_decay': configurations.weight_decay
}]
self.optimizer = t.optim.SGD(params, momentum=0.9)
#self.optimizer = self.faster_rcnn.get_optimizer()
# indicators for training status
self.rpn_cm = ConfusionMeter(2)
self.roi_cm = ConfusionMeter(21)
self.meters = {k: AverageValueMeter()
for k in LossTuple._fields} # average loss