AttributeError: ‘list‘ object has no attribute ‘detach‘ | torchsta | stat(model, (3, 224, 224)) | 模型

场景：

用torchstat分析模型结构，当model中包含Concat时，调用stat(model, (3, 224, 224))函数，会报错

---

问题描述

完整报错内容

Traceback (most recent call last):
  File "/home/yangzhanshan/disk/temp/my-Pruning/models/mobilenet_v2.py", line 1008, in <module>
    stat(model_yaml, (3, 224, 224))
  File "/home/yangzhanshan/disk/anaconda3/envs/yolov5/lib/python3.8/site-packages/torchstat/statistics.py", line 71, in stat
    ms.show_report()
  File "/home/yangzhanshan/disk/anaconda3/envs/yolov5/lib/python3.8/site-packages/torchstat/statistics.py", line 64, in show_report
    collected_nodes = self._analyze_model()
  File "/home/yangzhanshan/disk/anaconda3/envs/yolov5/lib/python3.8/site-packages/torchstat/statistics.py", line 57, in _analyze_model
    model_hook = ModelHook(self._model, self._input_size)
  File "/home/yangzhanshan/disk/anaconda3/envs/yolov5/lib/python3.8/site-packages/torchstat/model_hook.py", line 24, in __init__
    self._model(x)
  File "/home/yangzhanshan/disk/anaconda3/envs/yolov5/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/yangzhanshan/disk/temp/my-Pruning/models/mobilenet_v2.py", line 958, in forward
    return self.forward_once(x, profile)
  File "/home/yangzhanshan/disk/temp/my-Pruning/models/mobilenet_v2.py", line 970, in forward_once
    x = m(x)  # run
  File "/home/yangzhanshan/disk/anaconda3/envs/yolov5/lib/python3.8/site-packages/torchstat/model_hook.py", line 47, in wrap_call
    itemsize = input[0].detach().numpy().itemsize
AttributeError: 'list' object has no attribute 'detach'

Process finished with exit code 1

---

原因分析：

当模型运行到Concat层时，该层的输入为list，当input为list时，运行input[0].detach().numpy().itemsize会报错

---

解决方案：

进入报错的文件，入下图中的标红部分

更改该文件的47-50行
原来为

            itemsize = input[0].detach().numpy().itemsize

            start = time.time()
            output = self._origin_call[module.__class__](module, *input, **kwargs)

改为

            start = time.time()
            output = self._origin_call[module.__class__](module, *input, **kwargs)

            if type(input[0]) is not list:
                itemsize = input[0].detach().numpy().itemsize
            else:
                input = output
                itemsize = input[0].detach().numpy().itemsize

更改完成后，stat函数即可顺利运行

引流

from torchstat import stat
stat(model_yaml, (3, 224, 224))

深度学习 模型 访存量
深度学习 模型 参数量
深度学习 模型 空间复杂度