Toybrick

标题: onnx 2 rknn 的问题 [打印本页]

作者: nopattern    时间: 2019-7-10 14:40
标题: onnx 2 rknn 的问题
本帖最后由 nopattern 于 2019-7-10 14:46 编辑

    转换的时候出现了以下问题:

  1. I Match r_relu [['Relu_195']] [['Relu']] to [['relu']]
  2. I Try match Relu_202:out0
  3. I Match r_relu [['Relu_202']] [['Relu']] to [['relu']]
  4. I Try match ConvTranspose_54:out0
  5. W Not match tensor ConvTranspose_54:out0
  6. E Catch exception when loading onnx model: ssd_mbv2512.onnx!
  7. T Traceback (most recent call last):
  8. T   File "rknn/api/rknn_base.py", line 344, in rknn.api.rknn_base.RKNNBase.load_onnx
  9. T   File "rknn/base/RKNNlib/converter/convert_onnx.py", line 1037, in rknn.base.RKNNlib.converter.convert_onnx.convert_onnx.match_paragraph_and_param
  10. T   File "rknn/base/RKNNlib/converter/convert_onnx.py", line 950, in rknn.base.RKNNlib.converter.convert_onnx.convert_onnx._onnx_push_ready_tensor
  11. T TypeError: 'NoneType' object is not iterable
  12. Load mobilenet_v2 failed! Ret = -1
复制代码



   不支持 op ConvTranspose ? ConvTranspose是不是 对应 Deconvolution?



作者: nopattern    时间: 2019-7-11 09:41
是不是还不支持 Deconvolution 还是其他问题引起的?
作者: nopattern    时间: 2019-7-12 16:45
去掉了 Deconvolution , 发现了以下错误:
  1. D Process Conv_30_93 ...
  2. D RKNN output shape(convolution): (1 128 128 6)
  3. D Process Transpose_31_75 ...
  4. E Catch exception when loading onnx model: ssd_mbv2_512.onnx!
  5. T Traceback (most recent call last):
  6. T   File "rknn/api/rknn_base.py", line 350, in rknn.api.rknn_base.RKNNBase.load_onnx
  7. T   File "rknn/base/RKNNlib/RKNNnetbuilder.py", line 281, in rknn.base.RKNNlib.RKNNnetbuilder.RKNNNetBuilder.build
  8. T   File "rknn/base/RKNNlib/RKNNnetbuilder.py", line 300, in rknn.base.RKNNlib.RKNNnetbuilder.RKNNNetBuilder.build_layer
  9. T   File "rknn/base/RKNNlib/RKNNnetbuilder.py", line 300, in rknn.base.RKNNlib.RKNNnetbuilder.RKNNNetBuilder.build_layer
  10. T   File "rknn/base/RKNNlib/RKNNnetbuilder.py", line 300, in rknn.base.RKNNlib.RKNNnetbuilder.RKNNNetBuilder.build_layer
  11. T   [Previous line repeated 3 more times]
  12. T   File "rknn/base/RKNNlib/RKNNnetbuilder.py", line 306, in rknn.base.RKNNlib.RKNNnetbuilder.RKNNNetBuilder.build_layer
  13. T   File "rknn/base/RKNNlib/layer/RKNNlayer.py", line 262, in rknn.base.RKNNlib.layer.RKNNlayer.RKNNLayer.compute_shape
  14. T   File "rknn/base/RKNNlib/layer/permute.py", line 21, in rknn.base.RKNNlib.layer.permute.Permute.compute_out_shape
  15. T ValueError: invalid literal for int() with base 10: '[0,'
  16. Load mobilenet_v2 failed! Ret = -1
复制代码


看上去是transpose 的参数解析 axes  除了问题?
作者: nopattern    时间: 2019-7-12 21:51
本帖最后由 nopattern 于 2019-7-13 18:59 编辑

论坛上有人吗?能不能看一下,是不是rknn 有bug 还是哪里有问题?
作者: zhangyu    时间: 2019-8-2 15:20
bisa of convtranspose must be True, if bias=false, the problem occurs
作者: zxs    时间: 2019-10-21 15:42
反卷积的问题解决了吗,我也是发现反卷积出现了问题
作者: jefferyzhang    时间: 2019-10-22 08:39
W Not match tensor ConvTranspose_54ut0
这个OP不支持。
从rknn1.1起就可以自定义OP了,可以自己写OP来支持
作者: jefferyzhang    时间: 2019-11-5 10:39
hello,你们能否提供下这个模型源文件和转换脚本给我们,我们debug需要。




欢迎光临 Toybrick (https://t.rock-chips.com/) Powered by Discuz! X3.3