Pytorch——报错解决：RuntimeError: Output 0 of SelectBackward is a view and is being modified inplace.

我在做Semi-Supervised过程中，需要分别计算labeled和unlabeled data的loss，但是在多卡分布式过程中，不允许出现同一个model前向两次的情况，会报错（RuntimeError: Expected to mark a variable ready only once. This error is caused by one），因此我们需要将labeled和unlabeled data合并在一起前向一次model，然后再把结果拆分出来即可，实现思路如下：

合并labeled和unlabeled data；
输入模型前向传播得到预测结果；
再将预测结果分为labeled和unlabeled data两部分；
分别计算loss即可。

    def foward_student_train(self, sup_data, unsup_data):'''forward student'''# 合并输入datastudent_data, sup_data_length = \self.combine_student_data(sup_data, unsup_data)# 前向传播img_feats, _ = self.student.extract_feat(points=None, img=student_data['img_inputs'], img_metas=student_data['img_metas'])    # bev_feats, Nonestudent_info = self.student.pts_bbox_head(img_feats)# 分开预测结果datatsup_info, unsup_info = self.split_student_data(student_info, sup_data_length)# 计算有监督部分的lossloss_inputs = [sup_data['gt_bboxes_3d'], sup_data['gt_labels_3d'], sup_info]sup_loss = self.student.pts_bbox_head.loss(*loss_inputs)return sup_loss, unsup_info# 合并输入datadef combine_student_data(self, sup_data, unsup_data):'''combine sup and unsup data for student model'''assert isinstance(sup_data, dict) and \isinstance(unsup_data, dict)new_student_data = deepcopy(sup_data)keys = sup_data.keys()for key in keys:if key == 'img_inputs':new_student_data[key] = self.combine_imgs(sup_data[key], unsup_data[key])else:new_student_data[key] = sup_data[key] + \unsup_data[key]return new_student_data, (len(sup_data['img_metas']), len(unsup_data['img_metas']))# 分开预测结果datadef split_student_data(self, student_data, student_data_length):sup_data, unsup_data = [], []for idx, data in enumerate(student_data):   # len(list) = 6sup_data.append([{}])unsup_data.append([{}])# 默认len(list) = 1for key in data[0].keys():sup_data[idx][0][key] = data[0][key][:student_data_length[0], ...].clone()# detach upservised data gradientunsup_data[idx][0][key] = data[0][key][student_data_length[0]:, ...].clone()return tuple(sup_data), tuple(unsup_data)

Pytorch——报错解决：RuntimeError: Output 0 of SelectBackward is a view and is being modified inplace.相关推荐

RuntimeError: Output 0 of SelectBackward is a view and is being modified inplace.
1. 问题背景今天在浏览一些代码的时候,总是出现了以下的错误描述 RuntimeError: Output 0 of SelectBackward is a view and is being mo ...
RuntimeError: Output 0 of UnbindBackward is a view and is being modified inplace.
1. 问题描述: 在使用utils中的make_grid函数时 utils.make_grid(fmap, normalize=True, scale_each=True, nrow=nrow) 如果 ...
使用tensorboard遇到:Output 0 of UnbindBackward is a view and is being modified inplace.....
使用tensorboad时遇到问题: RuntimeError: Output 0 of UnbindBackward is a view and is being modified inplace. ...
Output 0 of UnbindBackward is a view and is being modified inplace
出错原因,参考:https://blog.csdn.net/pearl8899/article/details/112185772. 对于torch1.7,问题发生在utils.py,make_gri ...
PyTorch报错：RuntimeError: CUDA error: device-side assert triggered at /pytorch/aten/src/THC/generic
训练模型报错 RuntimeError: cuda runtime error (710) : device-side assert triggered at /pytorch/aten/src/TH ...
Redis报错解决：RDB: 0 MB of memory used by copy-on-write
我在使用Redis的时候报错如下: 2225:M 06 May 12:01:45.066 * 10 changes in 300 seconds. Saving... 32225:M 06 May 1 ...
Pytorch——报错解决：TypeError: default_collate: batch must contain tensors, numpy arrays, numbers, found
报错信息 TypeError: default_collate: batch must contain tensors, numpy arrays, numbers, dicts or lists; ...
pytorch报错：RuntimeError: mat1 and mat2 shapes cannot be multiplied (64x2500 and 3020x1600)
相信同学们在刚接触深度学习的时候老会遇到类似的问题: 这个其实就是卷积层到全连接层之间的参数数量不对应的问题卷积层的的输出结果是64*2500的矩阵,那么全连接层的第一个第一层的就应该是:2500* ...
PyTorch报错“/.../Loss.cu: ... [59,0,0] Assertion input_val ＞= zero input_val ＜= one failed.”
1 问题描述今天在调试代码的时候,出现这样一个错误: PyTorch Assertion: /opt/conda/conda-bld/pytorch_1640811803361/work/aten/ ...

Pytorch——报错解决：RuntimeError: Output 0 of SelectBackward is a view and is being modified inplace.

Pytorch——报错解决：RuntimeError: Output 0 of SelectBackward is a view and is being modified inplace.相关推荐

最新文章

热门文章