torch模型转换rknn模型加载中出错 toolkit: 1.7.3

[复制链接] · 发表于 2023-7-13 18:02:39

本帖最后由 toytoyTAC 于 2023-7-13 18:04 编辑

1. 首先出现 E NotImplementedError: The following operators are not implemented: ['aten::reciprocal']
2. 因为模型中有用到GeM，把p设成了1/3，pow(self.p)，之后出现load pytorch报错E ValueError: need more than 1 value to unpack
...
...
D Process clip_at_1417_176_176 ...
D RKNN output shape(clipbyvalue): (0 10 10 512)
D Tensor @clip_at_1417_176_176

ut0 type: float32
E Catch exception when loading pytorch model: /home/kerry/test/project0710/models/torch_models/743newtest.pt!
E Traceback (most recent call last):
E File "rknn/base/RKNNlib/RKNNnetbuilder.py", line 437, in rknn.base.RKNNlib.RKNNnetbuilder.RKNNNetBuilder.build_layer
E File "rknn/base/RKNNlib/RKNNnetbuilder.py", line 437, in rknn.base.RKNNlib.RKNNnetbuilder.RKNNNetBuilder.build_layer
E File "rknn/base/RKNNlib/RKNNnetbuilder.py", line 437, in rknn.base.RKNNlib.RKNNnetbuilder.RKNNNetBuilder.build_layer
E [Previous line repeated 1 more time]
E File "rknn/base/RKNNlib/RKNNnetbuilder.py", line 435, in rknn.base.RKNNlib.RKNNnetbuilder.RKNNNetBuilder.build_layer

E ValueError: need more than 1 value to unpack
...
...
3. 尝试把GeM中的pool_p4层拿掉后可以成功load，但是转rknn后精度下降比较明显。

只看该作者 · 发表于 2023-7-14 09:16:49

'aten::reciprocal' 这个算子不支持，你可以替换或者把输出节点放在这个操作之前，后处理时候再去实现他

只看该作者 · 发表于 2023-7-14 14:34:43

模型的backbone是resnet50，后面接的GeM

这是一开始的GeM：
def gem(x, p=3, eps=1e-6):
return F.avg_pool2d(x.clamp(min=eps).pow(p), (x.size(-2), x.size(-1))).pow(1.0 / p)

class GeM(nn.Module):
def __init__(self, p=3, eps=1e-6):
      super(GeM, self).__init__()
      self.p = nn.Parameter(torch.ones(1) * p)
      self.eps = eps
def forward(self, x):
      return gem(x, p=self.p, eps=self.eps)[:,:,0,0]
def __repr__(self):
      return (
         self.__class__.__name__
         + f"(p={self.p.data.tolist()[0]:.4f}, eps={str(self.eps)})"
      )

为了避免reciprocal这个算子改成了下面这样：
def gem(x, p=1/3, eps=1e-6):
return F.avg_pool2d(x.clamp(min=eps).pow(p), (x.size(-2), x.size(-1))).pow(p)

class GeM(nn.Module):
def __init__(self, p=1/3, eps=1e-6):
      super(GeM, self).__init__()
      self.p = nn.Parameter(torch.ones(1) * p)
      self.eps = eps
def forward(self, x):
      return gem(x, p=self.p, eps=self.eps)[:,:,0,0]
def __repr__(self):
      return (
         self.__class__.__name__
         + f"(p={self.p.data.tolist()[0]:.4f}, eps={str(self.eps)})"
      )

大佬帮忙看看是不是修改GeM出现了问题？