Baseline method,map problem

[byfate]

final: Map 77.8%

final loss. about range in [2.1,2.3] (run several times)

batch_size 12;two 2080 gpu other parameter as main.py

• change base.py 改动如何生成标签 def generate_ground_truth_base

def generate_ground_truth_base
cen_x, cen_y, bbox_w, bbox_h, theta = rect  # data_transform后四个点的最小外接矩形的 x_center y_center , w ,h ,theta 角度
# print(theta)
radius = gaussian_radius((math.ceil(bbox_h), math.ceil(bbox_w)))  # 计算高斯核半径
radius = max(0, int(radius))  # 保证>=0
ct = np.asarray([cen_x, cen_y], dtype=np.float32)  # ct float 中心点
ct_int = ct.astype(np.int32)  # ct int 中心点
# generate gt heatmap
draw_umich_gaussian(hm[annotation['cat'][k]], ct_int, radius)  # 所传参数为H*W, (center_x,center_y),radius
ind[k] = ct_int[1] * image_w + ct_int[0]                      
# generate wh gt
wh[k, 0:2] = 1. * bbox_w, 1. * bbox_h
# generate reg gt offset_x&y
reg[k] = ct - ct_int  # 记录回归的center offset_x offset_y
reg_mask[k] = 1  # 类别mask 置为1
assert 0 >= theta >= -90, "error angle"
# generate angle
ang[k] = theta

# so in getitem 
# ---- 省略
image = self.load_image(index)
        # opencv格式 numpy
image_h, image_w, c = image.shape
elif
    ----
elif self.phase == 'train':
    img_id = self.img_ids[index]
    annotation = self.load_annotation(index)
    data_dict = self.generate_ground_truth_base(image, annotation)
return data_dict

• nochange ctrbox_net.py change nothing, didn't clamp the angle value range [-9-,0]

  just use the same code as the BBAvector method

• change decoder.py

  ctdet_decode changeto ctdet_decode_base

  didnot change much,just add angle heatmap

heat = pr_decs['hm']                            # N*C*H*W # center*class*h*w
wh = pr_decs['wh']                              # N*2*H*W  w h
reg = pr_decs['reg']                            # N*2*H*W (offset x ,y)
cls_theta = pr_decs['ang']                      # N*1*H*W  表示是否是水平或者旋转框
batch, c, height, width = heat.size()
heat = self._nms(heat)                           # 经过NMS找出极大值点，没有最大分类的点变为0 # N*C*H*W
# print("heat",heat.shape)
scores, inds, clses, ys, xs = self._topk(heat)  # batch*K
reg = self._tranpose_and_gather_feat(reg, inds)
# print("reg_shape", reg.shape) # N*k*2
reg = reg.view(batch, self.K, 2)  # list 进行形状的reshape
xs = xs.view(batch, self.K, 1) + reg[:, :, 0:1]
ys = ys.view(batch, self.K, 1) + reg[:, :, 1:2]
clses = clses.view(batch, self.K, 1).float()
scores = scores.view(batch, self.K, 1)
wh = self._tranpose_and_gather_feat(wh, inds)
wh = wh.view(batch, self.K, 2)  # list 进行形状的reshape
# add
ang = self._tranpose_and_gather_feat(cls_theta, inds)
ang = ang.view(batch, self.K, 1)
detections = torch.cat([xs,  # cen_x
                        ys,  # cen_y
                        wh,
                        ang,
                        scores,
                        clses],
                        dim=2)
index = (scores > self.conf_thresh).squeeze(0).squeeze(1)  # 此处的conference决定检测出来的框保留的置信度
detections = detections[:, index, :]  # 返回的为1*M*12 ，此处不支持多张图同时进行
# print("decode base shape", detections.shape )
return detections.data.cpu().numpy()

• loss.py

  I follow you reply,and the weigh loss is used.

  and I clamp the angle heatmap to [-90,0]

class LossBase(torch.nn.Module):
    def __init__(self):
        super(LossBase, self).__init__()
        self.L_hm = FocalLoss()          # 类别
        self.L_wh = OffSmoothL1Loss()    # box parameter
        self.L_off = OffSmoothL1Loss()   # center offset
        self.L_ang = OffSmoothL1Loss()     # theta

    def forward(self, pr_decs, gt_batch):
        hm_loss = self.L_hm(pr_decs['hm'], gt_batch['hm'])
        wh_loss = self.L_wh(pr_decs['wh'], gt_batch['reg_mask'], gt_batch['ind'], gt_batch['wh'])

        off_loss = self.L_off(pr_decs['reg'], gt_batch['reg_mask'], gt_batch['ind'], gt_batch['reg'])
        # change
        pr_decs['ang'] = torch.clamp(pr_decs['ang'], min=-90., max=0)

        ang_loss = self.L_ang(pr_decs['ang'], gt_batch['reg_mask'], gt_batch['ind'], gt_batch['ang'])
        if isnan(hm_loss) or isnan(wh_loss) or isnan(off_loss):
            print('hm loss is {}'.format(hm_loss))
            print('wh loss is {}'.format(wh_loss))
            print('off loss is {}'.format(off_loss))

        loss = hm_loss + wh_loss + off_loss + 0.1*ang_loss
        return loss

• func_utils.py

  change function decode_prediction to function decode_prediction_new_base

  1 I follow your advice using cv2.points instead of using my own affine.

  and I clamp the cordinates to [0,input_w/input_h]

  def decode_prediction_new_base(predictions, dsets, args, img_id, down_ratio):
    predictions = predictions[0, :, :]
    ori_image = dsets.load_image(dsets.img_ids.index(img_id))
    h, w, c = ori_image.shape
  
    pts0 = {cat: [] for cat in dsets.category}
    scores0 = {cat: [] for cat in dsets.category}
    for pred in predictions:
        cen_x = int(pred[0])    # center_y
        cen_y = int(pred[1])    # center_y
  
        bbox_w = int(pred[2])   # w
        bbox_h = int(pred[3])   # h
        theta = pred[4]         # 角度
  
        pts = cv2.boxPoints(((cen_x, cen_y), (bbox_w, bbox_h), theta))
        # 限制坐标
        pts = np.asarray(np.clip(pts, a_min=0, a_max=args.input_w), np.float32)
  
        pts[:, 0] = pts[:, 0] * down_ratio / args.input_w * w
        pts[:, 1] = pts[:, 1] * down_ratio / args.input_h * h
  
        # dsets.category 如['ship'] clse为整数
        score = pred[5]  # 分数
        clse = pred[6]  # 类别
        pts0[dsets.category[int(clse)]].append(pts)
  
        scores0[dsets.category[int(clse)]].append(score)             # 每个class 每个类别
    return pts0, scores0

  So in the function write_results, I use the modified decode_prediction_new_base

[byfate]

I check my code several times,and follow your adivce(ge, weight loss cv2.points),would your please help me when you are convenient?Hope you can point out my problem if you are convenient,appreciate

[yijingru]

Wait, you get 77.8 on DOTA's testing set? Input is 0.5 and 1 scale with 600x600 crop?

[byfate]

Wait, you get 77.8 on DOTA's testing set? Input is 0.5 and 1 scale with 600x600 crop?

Sorry,its HRSC dataset,I didn't mention it .

[yijingru]

Oh, I see. The batch size would matter, I used about 100 epochs for HRSC2016. The codes look good to me. I think the current mAP means your code is good too.

[byfate]

OK, looks like that I haveto rent some gpu to see if the map can be higher.when the result come out, I will tell you .thanks for your replying.

[yijingru]

No problem.

[byfate]

No problem.

when I change the baseline method's learning rate as *1.0e-4，the batch size is 12 ,using two 2080ti. the highest map** is 83.6,but it is not stable during 100 epoch,and I observe the confidece of the box, lots of box confidence are not high.

what is really strange is when I run the baseline method using 1 Tesla V100 ,learning rate *1.25e-4,batch-size 20*the hightest map is about 77.7%(only in one card which has 32gm video card),and when I change the batch-size to 12 ,learning rate 1.0e-4，its highest map is 83.7%,also the traing process is not stable,it's hard to get the highest map depending on batch-size and learning rate.

It seems that I need to adjust different learning rate and batch-size during traing. In all,thanks for you kind replying.appreciated.

[yijingru]

Thank you for your report. I'm thinking maybe the Cuda or library matters, the newest GPU would not be greatly supported. Just guess... But I agree learning rate does matter when the batch size is small in the backpropagation.

[byfate]

Thank you for your report. I'm thinking maybe the Cuda or library matters, the newest GPU would not be greatly supported. Just guess... But I agree learning rate does matter when the batch size is small in the backpropagation.

OK,thank you.