search

MIV-XJTU / ARTrack

PyTorch implementation of paper "ARTrack" and "ARTrackV2"
https://artrackv2.github.io/
Apache License 2.0
241 stars 34 forks source link

No coco #90

Open sunqimin opened 2 weeks ago

sunqimin commented 2 weeks ago

你好！请问如果不使用coco数据集的话是不是只需要训练一次就可以了，那这里的config是怎么设置呢？

AlexDotHam commented 2 weeks ago

分两个阶段主要是加速训练，一阶段的训练比较快容易收敛，二阶段的速度慢，我没有尝试过只使用二阶段是否能收敛，我认为应该还是比较困难的。

sunqimin commented 2 weeks ago

可以分享一下两个阶段训练的loss吗，我感觉我这里好像是有问题的

AlexDotHam commented 2 weeks ago

artrackv2-artrackv2_256_got.log artrackv2_seq-artrackv2_seq_256_got.log That will be my pleasure

sunqimin commented 3 days ago

Traceback (most recent call last): File "D:\SQM\ARTrack-main\lib\train../..\lib\train\trainers\base_trainer.py", line 85, in train self.train_epoch() File "D:\SQM\ARTrack-main\lib\train../..\lib\train\trainers\ltr_seq_trainer.py", line 227, in train_epoch self.cycle_dataset(loader) File "D:\SQM\ARTrack-main\lib\train../..\lib\train\trainers\ltr_seq_trainer.py", line 83, in cycle_dataset explore_result = self.actor.explore(data) File "D:\SQM\ARTrack-main\lib\train../..\lib\train\actors\artrack_seq.py", line 472, in explore outputs_template = self.batch_init(template, template_bbox, here_gt_bbox) File "D:\SQM\ARTrack-main\lib\train../..\lib\train\actors\artrack_seq.py", line 325, in batch_init self.channel_average.append(np.mean(img, axis=(0, 1))) File "<__array_function__ internals>", line 180, in mean File "C:\Users\Mylog06.conda\envs\artrack\lib\site-packages\numpy\core\fromnumeric.py", line 3432, in mean return _methods._mean(a, axis=axis, dtype=dtype, File "C:\Users\Mylog06.conda\envs\artrack\lib\site-packages\numpy\core_methods.py", line 168, in _mean rcount = _count_reduce_items(arr, axis, keepdims=keepdims, where=where) File "C:\Users\Mylog06.conda\envs\artrack\lib\site-packages\numpy\core_methods.py", line 76, in _count_reduce_items items *= arr.shape[mu.normalize_axis_index(ax, arr.ndim)] numpy.AxisError: axis 0 is out of bounds for array of dimension 0 你好！请问在第二阶段第一个epoch训练结束之后报错该怎么解决呢

AlexDotHam commented 3 days ago

Emm，我这个代码我从来没在Windows上跑过，这个问题我也不太了解，不过看起来可能是Numpy版本的问题，建议你尝试一下1.23或者更低的版本，您可以在之后告知我结果。 另外，如果您是在特定的某个iteration才会报错的话，我猜测可能是数据集的原因，TrackingNet存在一些非法数据，这些数据的读取可能是错误的原因，您可以在”"D:\SQM\ARTrack-main\lib\train../..\lib\train\actors\artrack_seq.py", line 325, in batch_init“ 这一行附近增加一个try，如果读取到潜在的非法数据直接return，形如该文件中491行的对应代码

AlexDotHam commented 3 days ago

您还可以参考： https://github.com/MIV-XJTU/ARTrack/issues/36 他和您有相似的问题。