训练
yishunzhijian opened this issue · 17 comments
训练自己的数据集出现这个问题，请问怎么修改程序？
Traceback (most recent call last):
File "train.py", line 25, in
import eval as eval_script
File "/home/song/CIoU-master/eval.py", line 882
net.detect.use_cluster_nms = args.cluster_nms
^
TabError: inconsistent use of tabs and spaces in indentation
这个是对应位置的缩进不对齐所致，我上传的文件在我本地上都是对齐的，可能有些地方上传后不对齐了，你需要调整一下，再运行eval.py
这个是对应位置的缩进不对齐所致，我上传的文件在我本地上都是对齐的，可能有些地方上传后不对齐了，你需要调整一下，再运行eval.py
这修改好之后出现新问题，我觉得下面eval.py程序需要改一下：
if args.cross_class_nms = True:
nms='cross class'
else:
nms='not use cross class'
if args.fast_nms = True:
num_count = num_count + 1
if args.cluster_nms = True:
num_count = num_count + 1
if args.cluster_diounms = True:
num_count = num_count + 1
if args.spm = True:
num_count = num_count + 1
if args.spm_dist = True:
num_count = num_count + 1
if args.spm_dist_weighted = True:
改成
if args.cross_class_nms == True:
nms='cross class'
else:
nms='not use cross class'
if args.fast_nms == True:
num_count = num_count + 1
if args.cluster_nms == True:
num_count = num_count + 1
if args.cluster_diounms == True:
num_count = num_count + 1
if args.spm == True:
num_count = num_count + 1
if args.spm_dist == True:
num_count = num_count + 1
if args.spm_dist_weighted == True:
num_count = num_count + 1
训练出现了新的问题：
Traceback (most recent call last):
File "train.py", line 504, in
train()
File "train.py", line 371, in train
compute_validation_map(epoch, iteration, yolact_net, val_dataset, log if args.log else None)
File "train.py", line 492, in compute_validation_map
val_info = eval_script.evaluate(yolact_net, dataset, train_mode=True)
File "/home/song/CIoU-master/eval.py", line 964, in evaluate
preds = net(batch)
File "/home/song/.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 541, in call
result = self.forward(*input, **kwargs)
File "/home/song/CIoU-master/yolact.py", line 676, in forward
return self.detect(pred_outs, self)
File "/home/song/CIoU-master/layers/functions/detection.py", line 76, in call
result = self.detect(batch_idx, conf_preds, decoded_boxes, mask_data, inst_data)
File "/home/song/CIoU-master/layers/functions/detection.py", line 122, in detect
boxes, masks, classes, scores = self.cluster_diounms(boxes, masks, scores, self.nms_thresh, self.top_k)
File "/home/song/CIoU-master/layers/functions/detection.py", line 347, in cluster_diounms
scores, idx = scores.sort(1, descending=True)
IndexError: Dimension out of range (expected to be in range of [-1, 0], but got 1)
请问这个怎么改？
eval.py中，换成其他的nms看看是否会报错，比如fast_nms
eval.py中，换成其他的nms看看是否会报错，比如fast_nms
fast_nms好使，可以训练，其他的有相同的问题
这就很奇怪了，因为我看你报错是cluster_diounms的第一步scores.sort(1, descending=True)，而这一步对fast_nms也是一样的第一步。详情可查看layers/functions/detection.py
这就很奇怪了，因为我看你报错是cluster_diounms的第一步scores.sort(1, descending=True)，而这一步对fast_nms也是一样的第一步。详情可查看layers/functions/detection.py
其他的nms都可以训练，就这个还是有问题 spm_dist_weighted
Traceback (most recent call last):
File "train.py", line 504, in
train()
File "train.py", line 371, in train
compute_validation_map(epoch, iteration, yolact_net, val_dataset, log if args.log else None)
File "train.py", line 492, in compute_validation_map
val_info = eval_script.evaluate(yolact_net, dataset, train_mode=True)
File "/home/song/CIoU-master/eval.py", line 964, in evaluate
preds = net(batch)
File "/home/song/.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 541, in call
result = self.forward(*input, *kwargs)
File "/home/song/CIoU-master/yolact.py", line 676, in forward
return self.detect(pred_outs, self)
File "/home/song/CIoU-master/layers/functions/detection.py", line 76, in call
result = self.detect(batch_idx, conf_preds, decoded_boxes, mask_data, inst_data)
File "/home/song/CIoU-master/layers/functions/detection.py", line 128, in detect
boxes, masks, classes, scores = self.cluster_SPM_dist_weighted_nms(boxes, masks, scores, self.nms_thresh, self.top_k)
File "/home/song/CIoU-master/layers/functions/detection.py", line 489, in cluster_SPM_dist_weighted_nms
weights = (B(B>0.8).float() + torch.eye(n).cuda().expand(80,n,n)) * (scores.unsqueeze(2).expand(80,n,n))
RuntimeError: The size of tensor a (4) must match the size of tensor b (80) at non-singleton dimension 0
detection.py的489行改成 weights = (B*(B>0.8).float() + torch.eye(n).cuda().expand(4,n,n)) * (scores.unsqueeze(2).expand(4,n,n))(scores.unsqueeze(2).expand(4,n,n))就可以训练了
请问加入ciou损失和Cluster-NMS后，为什么ap mask的值会提升？我的理解是ap box会提升，因为是对边界框的优化。yolact检测和分割是分开进行的，如何解释ap mask会提升呢？谢谢回复。
YOLACT是基于检测的实例分割算法，因此box会影响裁剪出的mask的效果
YOLACT是基于检测的实例分割算法，因此box会影响裁剪出的mask的效果
可以讲的具体一点嘛，因为分割分支直接产生mask，似乎和box没什么关系
你看NMS的代码，mask是根据box的结果来裁剪的。box越正确，mask的结果也就越正确
你看NMS的代码，mask是根据box的结果来裁剪的。box越正确，mask的结果也就越正确
谢谢，那ciou对ap mask的提升怎么理解呢
CIoU用于bbox regression，可以提升box的质量
CIoU用于bbox regression，可以提升box的质量
非常感谢
还有一个小问题，看了一下代码，还是不太清楚ap mask是如何计算的，是计算每个像素吗
我看到你论文中Balancing precision and recall by different values of β in Cluster-NMS S+D，并且让β等于0.6，但是代码中并没有β的设置
@yishunzhijian 看 layers/box_utils.py distance函数
cluster_nms 出现同样问题，请问您解决了吗