出现 RuntimeError: adaptive_avg_pool2d_backward_cuda does not have a deterministic implementation的解决方法

目录

• 1. 问题所示
• 2. 原理分析
• 3. 解决方法

1. 问题所示

在目标检测的模型中加入CA注意力机制或者CBAM注意力机制的时候出现如下问题：

   Epoch    GPU_mem   box_loss   obj_loss   cls_loss  Instances       Size
  0%|          | 0/143 [00:02<?, ?it/s]
Traceback (most recent call last):
  File "train.py", line 760, in <module>
    main(opt)
  File "train.py", line 643, in main
    train(opt.hyp, opt, device, callbacks)
  File "train.py", line 402, in train
    scaler.scale(loss).backward()
  File "/home/kkjz/anaconda3/envs/py17/lib/python3.7/site-packages/torch/_tensor.py", line 488, in backward
    self, gradient, retain_graph, create_graph, inputs=inputs
  File "/home/kkjz/anaconda3/envs/py17/lib/python3.7/site-packages/torch/autograd/__init__.py", line 199, in backward
    allow_unreachabl