Toybrick

标题: pytorch ConvTranspose2d出错.版本1.4.0 [打印本页]

作者: 一袋米要扛几楼 时间: 2020-9-14 19:09
标题: pytorch ConvTranspose2d出错.版本1.4.0
class Darknet(nn.Module):
"""YOLOv3 object detection model"""

def __init__(self, config_path, img_size=448):
      super(Darknet, self).__init__()
      self.img_size = img_size
      self.seen = 0
      self.blocks = []
      self.yolo_layers = []

      innitial_channel = 1
      self.block2 = creat_block_maxpool(3, innitial_channel * 1, conv_kernel=(3, 3), conv_stride=(1, 1), conv_padding=(1, 1), pool_kernel=(2, 2), pool_stride=(2, 2))
      self.block4 = creat_block_maxpool(innitial_channel * 1, innitial_channel * 2, conv_kernel=(3, 3), conv_stride=(1, 1), conv_padding=(1, 1), pool_kernel=(2, 2), pool_stride=(2, 2))
      self.block6 = creat_block_maxpool(innitial_channel * 2, innitial_channel * 4, conv_kernel=(3, 3), conv_stride=(1, 1), conv_padding=(1, 1), pool_kernel=(2, 2), pool_stride=(2, 2))

      self.block8 = creat_block(innitial_channel * 4, innitial_channel * 8, conv_kernel=(3, 3), conv_stride=(1, 1), conv_padding=(1, 1))
      self.block9 = nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2), padding=0)

      self.block10 = creat_block(innitial_channel * 8, innitial_channel * 16, conv_kernel=(3, 3), conv_stride=(1, 1), conv_padding=(1, 1))
      self.block11 = nn.MaxPool2d(kernel_size=(2, 2), stride=(1, 1), padding=0)

      self.block12 = creat_block(innitial_channel * 16, innitial_channel * 32, conv_kernel=(3, 3), conv_stride=(1, 1), conv_padding=(1, 1))
      self.block13 = creat_block(innitial_channel * 32, innitial_channel * 8, conv_kernel=(3, 3), conv_stride=(1, 1), conv_padding=(1, 1))
      self.block14 = creat_block(innitial_channel * 8, innitial_channel * 16, conv_kernel=(3, 3), conv_stride=(1, 1), conv_padding=(1, 1))
      self.block15_yolo1 = creat_block(innitial_channel * 16, 3*(5+1), conv_kernel=(3, 3), conv_stride=(1, 1), conv_padding=(1, 1))

      # self.block17 = 13
      self.block18 = creat_block(innitial_channel * 8, innitial_channel * 4, conv_kernel=(3, 3), conv_stride=(1, 1), conv_padding=(1, 1))
      self.block19 = torch.nn.ConvTranspose2d(innitial_channel * 4, innitial_channel * 8, kernel_size=3, stride=2, padding=0, output_padding=1, bias=True)
      # self.block20_rote = 19+8
      self.block21 = creat_block(innitial_channel * 8, innitial_channel * 16, conv_kernel=(3, 3), conv_stride=(1, 1), conv_padding=(1, 1))
      self.block22 = creat_block(innitial_channel * 16, innitial_channel * 16, conv_kernel=(3, 3), conv_stride=(1, 1), conv_padding=(1, 1))
      self.block23_yolo2 = nn.Conv2d(innitial_channel * 16, 3*(5+1), kernel_size=3, stride=1, padding=1)

def forward(self, x):

      x2 = self.block2(x)
      x4 = self.block4(x2)
      x6 = self.block6(x4)
      x8 = self.block8(x6)
      x9 = self.block9(x8)
      x10 = self.block10(x9)
      x11 = self.block11(x10)
      x12 = self.block12(x11)
      x13 = self.block13(x12)

      x14 = self.block14(x13)
      x15 = self.block15_yolo1(x14)

      x18 = self.block18(x13)# return x19 之前的结果运行通过，return x19之后的结果报错。也就是torch.nn.ConvTranspose2d
      x19 = self.block19(x18)
      # # print(x_8.shape, x_19.shape)
      x20 = x8 + x19

      x21 = self.block21(x20)
      x22 = self.block22(x21)
      x23 = self.block23_yolo2(x22)
      # return x23 #报错
      return x15, x18 #通过rk版本1.4.0--> config model
done
--> Loading model
WARNING: The TensorFlow contrib module will not be included in TensorFlow 2.0.
For more information, please see:
  * https://github.com/tensorflow/co ... 7-contrib-sunset.md
  * https://github.com/tensorflow/addons
If you depend on functionality not listed there, please file an issue.
/usr/local/lib/python3.5/dist-packages/onnx_tf/common/__init__.py:87: UserWarning: FrontendHandler.get_outputs_names is deprecated. It will be removed in future release.. Use node.outputs instead.
  warnings.warn(message)
./rknn_models_test2/yoloV3.pt ********************
WARNING: Token 'COMMENT' defined, but not used
WARNING: There is 1 unused token
E Try match aten::_convolution_at_x19.1

ut0 failed, catch exception!
E Catch exception when loading pytorch model: ./rknn_models_test2/yoloV3.pt!
E Traceback (most recent call last):
E File "rknn/base/RKNNlib/converter/convert_pytorch.py", line 1567, in rknn.base.RKNNlib.converter.convert_pytorch.convert_pytorch.match_paragraph_and_param
E File "rknn/base/RKNNlib/converter/convert_pytorch.py", line 1444, in rknn.base.RKNNlib.converter.convert_pytorch.convert_pytorch._torch_try_match_ruler
E ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()
E During handling of the above exception, another exception occurred:
E Traceback (most recent call last):
E File "rknn/api/rknn_base.py", line 657, in rknn.api.rknn_base.RKNNBase.load_pytorch
E File "rknn/base/RKNNlib/app/importer/import_pytorch.py", line 104, in rknn.base.RKNNlib.app.importer.import_pytorch.ImportPytorch.run
E File "rknn/base/RKNNlib/converter/convert_pytorch.py", line 1590, in rknn.base.RKNNlib.converter.convert_pytorch.convert_pytorch.match_paragraph_and_param
E File "rknn/api/rknn_log.py", line 312, in rknn.api.rknn_log.RKNNLog.e
E ValueError: Try match aten::_convolution_at_x19.1

ut0 failed, catch exception!
Load pytorch model failed!

作者: jefferyzhang 时间: 2020-9-14 22:26
置顶帖的qq群里问问

作者: 一袋米要扛几楼 时间: 2020-9-15 08:38

jefferyzhang 发表于 2020-9-14 22:26
置顶帖的qq群里问问

好的谢谢

作者: 一袋米要扛几楼 时间: 2020-9-15 08:43

jefferyzhang 发表于 2020-9-14 22:26
置顶帖的qq群里问问

555555，原谅我没找到QQ群

，能给我帖一个吗？

作者: jefferyzhang 时间: 2020-9-15 09:08

一袋米要扛几楼发表于 2020-9-15 08:43
555555，原谅我没找到QQ群，能给我帖一个吗？

置顶贴就有：http://t.rock-chips.com/forum.ph ... &extra=page%3D1

作者: 一袋米要扛几楼 时间: 2020-9-15 09:30

jefferyzhang 发表于 2020-9-15 09:08
置顶贴就有：http://t.rock-chips.com/forum.php?mod=viewthread&tid=1967&extra=page%3D1

get，

作者: xiaqing10 时间: 2021-1-12 14:08
nn.ConvTranspose2d 这个rknn 还不支持，用upsample来代替。

作者: changchenliang 时间: 2021-4-13 15:37

xiaqing10 发表于 2021-1-12 14:08
nn.ConvTranspose2d 这个rknn 还不支持，用upsample来代替。

请问目前最新的rknn1.6.0支持nn.ConvTranspose2d了吗？
我的网络里有nn.ConvTranspose2d运算，在之前的rknn1.4.0+torch1.2.0版本下竟然能正确转化和推理，但是放到目前的rknn1.6.0+torch1.6.0下就出错了，这好奇怪啊？？这是为何。。

作者: bobby_jiang 时间: 2021-4-16 17:54
最新的rknn1.6.0支持nn.ConvTranspose2d。
你要检查，你的计算棒固件是否升级到最新版本了？请按照wiki在线升级到最新的1.6.0版本。

欢迎光临 Toybrick (https://t.rock-chips.com/)