ctdet2onnx.md

1. Replace THC-based DCNv2 with ATen-based DCNv2. If it is not replaced, you will get (TypeError: int() not supported on cdata 'struct THLongTensor *') when converting onnx, and I have no idea to solve it. So I use DCNv2 from mmdetection.

   • copy the dcn to lib/models/netowrks

     cp -r dcn lib/models/netowrks

   • upgrade pytorch to 1.0+
   • complie Deform Conv

     cd lib/models/netowrks/dcn
     python setup.py build_ext --inplace

2. Add symbolic to DeformConvFunction.

   class ModulatedDeformConvFunction(Function):
   
       @staticmethod
       def symbolic(g, input, offset, mask, weight, bias,stride,padding,dilation,groups,deformable_groups):
           return g.op("DCNv2", input, offset, mask, weight, bias,
                       stride_i = stride,padding_i = padding,dilation_i = dilation,
                       groups_i = groups,deformable_group_i = deformable_groups)
       @staticmethod
       def forward(ctx,
                   input,
                   offset,
                   mask,
                   weight,
                   bias=None,
                   stride=1,
                   padding=0,
                   dilation=1,
                   groups=1,
                   deformable_groups=1):
                   pass#.......

3. Change import

   • change (from .DCNv2.dcn_v2 import DCN) to (from .dcn.modules.deform_conv import ModulatedDeformConvPack as DCN) in pose_dla_dcn.py and resnet_dcn.py.
   • Now you can convert the model using Deform Conv to onnx.

4. For dla34.

   • convert ctdet_coco_dla_2x.pth to ctdet_coco_dla_2x.onnx

     from lib.opts import opts
     from lib.models.model import create_model, load_model
     from types import MethodType
     import torch.onnx as onnx
     import torch
     from torch.onnx import OperatorExportTypes
     from collections import OrderedDict
     ## onnx is not support dict return value
     ## for dla34
     def pose_dla_forward(self, x):
         x = self.base(x)
         x = self.dla_up(x)
         y = []
         for i in range(self.last_level - self.first_level):
             y.append(x[i].clone())
         self.ida_up(y, 0, len(y))
         ret = []  ## change dict to list
         for head in self.heads:
             ret.append(self.__getattr__(head)(y[-1]))
         return ret
     ## for dla34v0
     def dlav0_forward(self, x):
         x = self.base(x)
         x = self.dla_up(x[self.first_level:])
         # x = self.fc(x)
         # y = self.softmax(self.up(x))
         ret = []  ## change dict to list
         for head in self.heads:
             ret.append(self.__getattr__(head)(x))
         return ret
     ## for resdcn
     def resnet_dcn_forward(self, x):
         x = self.conv1(x)
         x = self.bn1(x)
         x = self.relu(x)
         x = self.maxpool(x)
     
         x = self.layer1(x)
         x = self.layer2(x)
         x = self.layer3(x)
         x = self.layer4(x)
         x = self.deconv_layers(x)
         ret = []  ## change dict to list
         for head in self.heads:
             ret.append(self.__getattr__(head)(x))
         return ret
     
     forward = {'dla':pose_dla_forward,'dlav0':dlav0_forward,'resdcn':resnet_dcn_forward}
     
     opt = opts().init()  ## change lib/opts.py add_argument('task', default='ctdet'....) to add_argument('--task', default='ctdet'....)
     opt.arch = 'dla_34'
     opt.heads = OrderedDict([('hm', 80), ('reg', 2), ('wh', 2)])
     opt.head_conv = 256 if 'dla' in opt.arch else opt.head_conv=64
     print(opt)
     model = create_model(opt.arch, opt.heads, opt.head_conv)
     model.forward = MethodType(forward[opt.arch.split('_')[0]], model)
     load_model(model, 'ctdet_coco_dla_2x.pth')
     model.eval()
     model.cuda()
     input = torch.zeros([1, 3, 512, 512]).cuda()
     onnx.export(model, input, "ctdet_coco_dla_2x.onnx", verbose=True,
                 operator_export_type=OperatorExportTypes.ONNX)

   • If you get (ValueError: Auto nesting doesn't know how to process an input object of type int. Accepted types: Tensors, or lists/tuples of them) You need to change (def _iter_filter) in torch.autograd.function.

        def _iter_filter(....):
            ....
            if condition(obj):
                 yield obj
            elif isinstance(obj,int):  ## int to tensor
                 yield torch.tensor(obj)
            ....

5. onnx-tensorrt DCNv2 plugin

   • Related code
     • onnx-tensorrt/builtin_op_importers.cpp
     • onnx-tensorrt/builtin_plugins.cpp
     • onnx-tensorrt/DCNv2.hpp
     • onnx-tensorrt/DCNv2.cpp
     • onnx-tensorrt/dcn_v2_im2col_cuda.cu
     • onnx-tensorrt/dcn_v2_im2col_cuda.h
   • Not only support centernet. If you want to convert other model to tensorrt engine, please refer to src/ctdetNet.cpp or contact me.