Ncnn: keras转onnx后再用onnx2ncnn转换，出现flatten not supported axis=0错误

Created on 14 Jan 2019 · 5Comments · Source: Tencent/ncnn

原keras模型有个flatten层：输入：[2，2，20]，格式：channels last。输出：80
转换onnx成功，但在onnx2ncnn转换时，报错：
unsupported flatten axis = 0
identity not supported yet
请教，目前ncnn支持keras转换过来的这种格式的flatten吗？如果不支持，怎么修改？
我目前打算修改ncnn读入flatten层部分的代码，不知能否实现。

Source

xuqiong1989

Most helpful comment

遇到同样问题的小伙伴，可以查看我的总结：
https://blog.csdn.net/weixin_41770169/article/details/86488493
https://blog.csdn.net/weixin_41770169/article/details/86488825

xuqiong1989 on 15 Jan 2019

👍2

All 5 comments

axis=0 警告暂时无视
identity 可以用 onnx optimizer 提前去掉

nihui on 15 Jan 2019

❤1

axis=0 警告暂时无视
identity 可以用 onnx optimizer 提前去掉

非常感谢，已经转换成功！

xuqiong1989 on 15 Jan 2019

遇到同样问题的小伙伴，可以查看我的总结：
https://blog.csdn.net/weixin_41770169/article/details/86488493
https://blog.csdn.net/weixin_41770169/article/details/86488825

xuqiong1989 on 15 Jan 2019

👍2

你好，我也在尝试keras2onnx2ncnn，我也是keras的flatten层出现问题，如下所述：

1、在keras2onnx之后，发现flatten层被分割成很多层，如下
%TFNodes_pool2_MaxPool_0 = Identity(%pool2_MaxPool_0)
%TFNodes_flatten_stack_Unsqueeze__12_0 = Unsqueezeaxes = [0]
%TFNodes_flatten_Shape_0 = Shape(%TFNodes_pool2_MaxPool_0)
%TFNodes_flatten_Shape__7_0 = Castto = 1
%TFNodes_flatten_strided_slice_0 = Slice(%TFNodes_flatten_Shape__7_0, %const_slice__8, const_slice__9, %const_slice__10)
%TFNodes_flatten_strided_slice__11_0 = Castto = 6
%TFNodes_flatten_Prod_0 = ReduceProd[axes = [0], keepdims = 0]
(%TFNodes_flatten_strided_slice__11_0)
%TFNodes_flatten_stack_Unsqueeze__13_0 = Unsqueezeaxes = [0]
%TFNodes_flatten_stack_Concat__14_0 = Concataxis = 0
%TFNodes_flatten_Reshape__15_0 = Castto = 7
%TFNodes_flatten_Reshape_0 = Reshape(%TFNodes_pool2_MaxPool_0,
%TFNodes_flatten_Reshape__15_0)
感觉不是很正常，因为其他卷积层池化层等转换后还是一层，所以想请问flatten层转换后被拆分是正常的吗？
2、在使用onnx2ncnn.exe时，出现如下报错：
Identity not supported yet!
Unsqueeze not supported yet!
# axes 7
Shape not supported yet!
Cast not supported yet!
# to=1
Cast not supported yet!
# to=6
ReduceProd not supported yet!
# axes 7
# keepdims=0
Unsqueeze not supported yet!
# axes 7
Cast not supported yet!
# to=7
Identity not supported yet!
可以发现除了Identity层外，其他不支持的层都是flaten的拆分层，向请问这个问题怎么解决。
3、我参考了你的csdn帖子去除identity层，报错如下：
Traceback (most recent call last):
File "G:/keras/FE_LeNet/keras_to_onnx.py", line 26, in
optimized_model = optimizer.optimize(onnx_model, passes)
File "F:\Anaconda3\lib\site-packages\onnx\optimizer.py", line 55, in optimize
optimized_model_str = C.optimize(model_str, passes)
IndexError: invalid unordered_map key
程序基本跟官方的一样，想请问这个报错如何处理。