【编程BUG】Pytorch中的+=操作

用于行为识别的TRN代码中有一段代码

for scaleID in range(1, len(self.scales)):
# iterate over the scales
	idx_relations_randomsample = np.random.choice(len(self.relations_scales[scaleID]), self.subsample_scales[scaleID], replace=False)
	for idx in idx_relations_randomsample:
		act_relation= input[:, self.relations_scales[scaleID][idx], :]
		act_relation = act_relation.view(act_relation.size(0), self.scales[scaleID] * self.img_feature_dim)
		act_relation = self.fc_fusion_scales[scaleID](act_relation)
		act_all +=act_relation
return act_all

运行会报错：RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.HalfTensor [64, 256]], which is output 0 of ReluBackward0, is at version 18; expected version 0 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).

解决方案为将倒数第二行改为:

act_all =act_all+act_relation

所以要慎用+=