num_expected_pos = int(self.num * self.pos_fraction)
pos_inds = self._sample_pos(assign_result, num_expected_pos, bboxes=bboxes, **kwargs)
# We found that sampled indices have duplicated items occasionally.
# (may be a bug of PyTorch)
pos_inds = pos_inds.unique() # 这里会报错
num_sampled_pos = pos_inds.numel()
num_expected_neg = self.num - num_sampled_pos
观察发现由于pos_inds调用的 _sample_pos() 函数有问题
def _sample_pos(self, assign_result, num_expected, **kwargs):
"""Randomly sample some positive samples."""
pos_inds = jt.nonzero(assign_result.gt_inds > 0) # 此函数偶尔返回错误inds
print(pos_inds.size())
if pos_inds.numel() != 0:
pos_inds = pos_inds.squeeze(1)
if pos_inds.numel() <= num_expected:
return pos_inds
else:
return self.random_choice(pos_inds, num_expected)
下面是返回的pos_inds的size().
···
[16,1,]
[20,1,]
[45,1,]
[6,1,]
[11,1,]
[15,1,]
[15,1,]
[4,1,]
[11,1,]
[27,1,]
[290,1,]
[8,1,]
[65,1,]
[18,1,]
[5,1,]
[23,1,]
[2,1,]
[5,1,]
[50,1,]
[2,1,]
[31,1,]
[1,1,]
[0,1,] # 由于这个返回的结果有问题
Traceback (most recent call last):
File "tools/run_net.py", line 56, in <module>
main()
File "tools/run_net.py", line 47, in main
runner.run()
File "/home/yuxk/.conda/envs/pytorch/lib/python3.8/site-packages/jdet-0.2.0.0-py3.8.egg/jdet/runner/runner.py", line 84, in run
self.train()
File "/home/yuxk/.conda/envs/pytorch/lib/python3.8/site-packages/jdet-0.2.0.0-py3.8.egg/jdet/runner/runner.py", line 126, in train
losses = self.model(images,targets)
File "/home/yuxk/.conda/envs/pytorch/lib/python3.8/site-packages/jittor/__init__.py", line 951, in __call__
return self.execute(*args, **kw)
File "/home/yuxk/.conda/envs/pytorch/lib/python3.8/site-packages/jdet-0.2.0.0-py3.8.egg/jdet/models/networks/rcnn.py", line 39, in execute
proposals_list, rpn_losses = self.rpn(features,targets)
File "/home/yuxk/.conda/envs/pytorch/lib/python3.8/site-packages/jittor/__init__.py", line 951, in __call__
return self.execute(*args, **kw)
File "/home/yuxk/.conda/envs/pytorch/lib/python3.8/site-packages/jdet-0.2.0.0-py3.8.egg/jdet/models/roi_heads/oriented_rpn_head.py", line 487, in execute
losses = self.loss(*outs,targets)
File "/home/yuxk/.conda/envs/pytorch/lib/python3.8/site-packages/jdet-0.2.0.0-py3.8.egg/jdet/models/roi_heads/oriented_rpn_head.py", line 457, in loss
labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,num_total_pos, num_total_neg = self.get_targets(anchor_list, valid_flag_list, targets)
File "/home/yuxk/.conda/envs/pytorch/lib/python3.8/site-packages/jdet-0.2.0.0-py3.8.egg/jdet/models/roi_heads/oriented_rpn_head.py", line 376, in get_targets
all_bbox_weights, pos_inds_list, neg_inds_list, sampling_results_list) = multi_apply(self._get_targets_single, anchor_list, valid_flag_list, targets)
File "/home/yuxk/.conda/envs/pytorch/lib/python3.8/site-packages/jdet-0.2.0.0-py3.8.egg/jdet/utils/general.py", line 53, in multi_apply
return tuple(map(list, zip(*map_results)))
File "/home/yuxk/.conda/envs/pytorch/lib/python3.8/site-packages/jdet-0.2.0.0-py3.8.egg/jdet/models/roi_heads/oriented_rpn_head.py", line 308, in _get_targets_single
sampling_result = self.sampler.sample(assign_result, anchors, target_bboxes)
File "/home/yuxk/.conda/envs/pytorch/lib/python3.8/site-packages/jdet-0.2.0.0-py3.8.egg/jdet/models/boxes/sampler.py", line 96, in sample
pos_inds = pos_inds.unique()
File "/home/yuxk/.conda/envs/pytorch/lib/python3.8/site-packages/jittor/misc.py", line 573, in unique
input_flatten = input_flatten.view(orig_shape[0], -1)
File "/home/yuxk/.conda/envs/pytorch/lib/python3.8/site-packages/jittor/__init__.py", line 553, in reshape
return origin_reshape(x, shape)
RuntimeError: Wrong inputs arguments, Please refer to examples(help(jt.ops.reshape)).
Types of your inputs are:
self = module,
args = (Var, tuple, ),
The function declarations are:
VarHolder* reshape(VarHolder* x, NanoVector shape)
Failed reason:[f 0919 21:23:05.270914 56 reshape_op.cc:47] Check failed: y_items != 0 && x_items % y_items == 0 reshape shape is invalid for input of size 0
可以看出由于 pos_inds 的返回出问题，导致后续的问题，至于为啥会出来一个size是 （0，1） 的，我搞不懂。为啥第一维会是0.