tensorboy / centerpose Goto Github PK

i use coco2017 and your pretrained model dla34.pth to train again,the loss is still growing from 3.1 to round 17.why the loss does keep unchanged or change smaller,

I found the code just for single test, but how to batch test?

Hi, I'm still a little confused about func gaussian_radius, which defined like this:

def gaussian_radius(det_size, min_overlap=0.7):
  height, width = det_size

  a1  = 1
  b1  = (height + width)
  c1  = width * height * (1 - min_overlap) / (1 + min_overlap)
  sq1 = np.sqrt(b1 ** 2 - 4 * a1 * c1)
  r1  = (b1 + sq1) / 2

  a2  = 4
  b2  = 2 * (height + width)
  c2  = (1 - min_overlap) * width * height
  sq2 = np.sqrt(b2 ** 2 - 4 * a2 * c2)
  r2  = (b2 + sq2) / 2

  a3  = 4 * min_overlap
  b3  = -2 * min_overlap * (height + width)
  c3  = (min_overlap - 1) * width * height
  sq3 = np.sqrt(b3 ** 2 - 4 * a3 * c3)
  r3  = (b3 + sq3) / 2
  return min(r1, r2, r3)

I have test some values like this:

Dose this a little bigger for drawing gaussian distribution? Beacause in keypoints detection job, sigma always be 2.5 or 3.0. What do you think of this?

Hi, I am confused about this. The output feature of HRNet-W32 has 32 channels but the targe maps has more than 51 channel (17 keypoint hmps + 17*2 offset x, y maps).
Usually, the target feature maps have more than 4~5 times the channel number of the output feature maps of the backbone network.

@tensorboy 你好，看你的表格里面DLA-34在coco val上的精度是62.7%，我这边用官方的model测出来是59.2%，请问你做了什么别的优化措施嘛

How do I train the model in a custom dataset?

Hi, I'm still a little confused about func gaussian_radius, which defined like this:

def gaussian_radius(det_size, min_overlap=0.7):
  height, width = det_size

  a1  = 1
  b1  = (height + width)
  c1  = width * height * (1 - min_overlap) / (1 + min_overlap)
  sq1 = np.sqrt(b1 ** 2 - 4 * a1 * c1)
  r1  = (b1 + sq1) / 2

  a2  = 4
  b2  = 2 * (height + width)
  c2  = (1 - min_overlap) * width * height
  sq2 = np.sqrt(b2 ** 2 - 4 * a2 * c2)
  r2  = (b2 + sq2) / 2

  a3  = 4 * min_overlap
  b3  = -2 * min_overlap * (height + width)
  c3  = (min_overlap - 1) * width * height
  sq3 = np.sqrt(b3 ** 2 - 4 * a3 * c3)
  r3  = (b3 + sq3) / 2
  return min(r1, r2, r3)

I have test some values like this:

Dose this a little bigger for drawing gaussian distribution? Beacause in keypoints detection job, sigma always be 2.5 or 3.0. What do you think of this?

At 'results, meta = results[0]' ,got this error

你好，模型网盘链接已失效，可以重新上传下吗，万分感激！

@tensorboy hi, i have finish the train with your code ,but when i test the model i get RuntimeError：storage has wrong size:expected 0 got 589824, do u know the reason

Can you send the model in Baidu pan?please
thank you, good people.

Thank the author! The new version of centernet has been released（centernetv2）. Do you have any idea of a V2 based pose estimation version? If not, can you provide some ideas for improvement?

Drop parameter dla_up.ida_2.up_1.weight.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.node_1.actf.0.weight.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.node_1.actf.0.bias.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.node_1.actf.0.running_mean.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.node_1.actf.0.running_var.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.node_1.actf.0.num_batches_tracked.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.node_1.conv.weight.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.node_1.conv.bias.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.node_1.conv.conv_offset_mask.weight.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.node_1.conv.conv_offset_mask.bias.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.proj_2.actf.0.weight.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.proj_2.actf.0.bias.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.proj_2.actf.0.running_mean.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.proj_2.actf.0.running_var.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.proj_2.actf.0.num_batches_tracked.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.proj_2.conv.weight.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.proj_2.conv.bias.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.proj_2.conv.conv_offset_mask.weight.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.
Drop parameter dla_up.ida_2.proj_2.conv.conv_offset_mask.bias.If you see this, your model does not fully load the pre-trained weight. Please make sure you have correctly specified --arch xxx or set the correct --num_classes for your own dataset.

Traceback (most recent call last):
File "/home/centerpose/tools/train.py", line 151, in
main(cfg, local_rank)
File "/home/centerpose/tools/train.py", line 120, in main
mAP = val_dataset.run_eval(preds, cfg.OUTPUT_DIR)
File "/home/centerpose/tools/../lib/datasets/coco_hp.py", line 100, in run_eval
coco_dets = self.coco.loadRes(self.convert_eval_format(results))
File "/home/centerpose/tools/../lib/datasets/coco_hp.py", line 64, in convert_eval_format
bbox = dets[:4]
IndexError: invalid index to scalar variable.

The model zoo links are broken now, can you fix it?

Hi, what are the settings corresponding to the results in the table in READE:

1. batch test or single input one by one?
2. flip augmentation is used when you report the FPS?
3. what kind of GPU is used?

Thank you for your help in advance.

Hi, thanks for your work. Is there some error in ReadMe about wrong figure for showing mAP vs Run time? Based your tabel, it may be mAP vs FPS not mAP vs Run time.

请教：如何更优的去设置mobilenetv2的参数呢or策略？看到您yml里面的参数，觉得很多不甚合理

when i load efficientdet model,it loaded a darnet model

I have test the train process using first 100 samples of MSCOCO. The cmd of training is

 python3 train.py --cfg ../experiments/mobilenetv3_512x512.yaml

and when run mAP calculate in line 121 of train.py, I got an error like this:

but when I run

python3 evaluate.py --cfg ../experiments/mobilenetv3_512x512.yaml --NMS false --TESTMODEL ../models/model_zoo/mobilenetV3_1x.pth

It's ok and can output result.

Update:
Maybe change code like this will be work in here

dets_out = np.concatenate(
    [detection[1] for detection in dets_out], axis=0).astype(np.float32)           
if self.cfg.TEST.NMS or len(self.cfg.TEST.TEST_SCALES) > 1:
    soft_nms_39(dets_out, Nt=0.5, method=2)
dets_out = dets_out.tolist()
# results[batch['meta']['img_id'].cpu().numpy()[0]] = dets_out[0]
results[batch['meta']['img_id'].cpu().numpy()[0]] = dets_out

我使用了预训练的High Resolution，得到的ap值如下：

Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets= 20 ] = 0.401
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets= 20 ] = 0.617
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets= 20 ] = 0.429
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets= 20 ] = 0.481
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets= 20 ] = 0.386
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 20 ] = 0.632
 Average Recall     (AR) @[ IoU=0.50      | area=   all | maxDets= 20 ] = 0.876
 Average Recall     (AR) @[ IoU=0.75      | area=   all | maxDets= 20 ] = 0.680
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets= 20 ] = 0.561
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets= 20 ] = 0.733

但README中给出的ap值是0.495，远高于测试的结果。请问预训练模型是不是传错了？

Hello, when I was testing your model, I found that the "TEST_SCALES" in "hrnet_ *.yaml" was set to [1, 2] and [1] in other YAML files, which had a great impact on the speed of HRNet model. Is there any special purpose for you to set it like this?

I tried to export themodel to onnx, but the main issue is the DCN layers are not supported in onnx as far as I know.