try to add a python layer deformable_conv2d implement

Author: vera121

python/deform_conv2d.py

@@ -0,0 +1,76 @@
+import caffe
+import numpy as np
+import torch
+import torchvision.ops
+
+class Deform_Conv2D(caffe.Layer):
+ """
+ Implemention of pytorch deform_conv2d method.
+ Refer to https://pytorch.org/vision/main/generated/torchvision.ops.deform_conv2d.html
+
+ Used for DBNet deploy https://github.com/MhLiao/DB/blob/master/assets/ops/dcn/functions/deform_conv.py#L111
+ Input:
+ input,
+ offset,
+ mask,
+ weight,
+ # bias=None, #
+ Param:
+ stride=1,
+ padding=1, #
+ dilation=1,
+ groups=1, #
+ deformable_groups=1 #
+ """
+
+ def setup(self, bottom, top):
+ # check number of inputs and outputs
+ if len(bottom) != 4:
+ raise Exception("Only supporting input 4 Tensors now!")
+ if len(top) != 1:
+ raise Exception("Only output one Tensor at a time!")
+ 
+ d = eval(self.param_str)
+ if d["stride"] != None:
+ self.stride = d["stride"]
+ else:
+ self.stride = 1
+ if d["padding"] != None:
+ self.padding = d["padding"]
+ else:
+ self.padding = 1 
+ if d["dilation"] != None:
+ self.dilation = d["dilation"]
+ else:
+ self.dilation = 1
+
+
+ def reshape(self, bottom, top):
+ # check input dimensions
+ #if bottom[0].count == 0:
+ # raise Exception("Input must not be empty!")
+ top[0].reshape(*bottom[0].data.shape)
+
+ def forward(self, bottom, top):
+ input = bottom[0].data
+ offset = bottom[1].data
+ mask = bottom[2].data
+ weight = bottom[3].data
+ # bias #
+ x = torchvision.ops.deform_conv2d(
+ input=torch.from_numpy(input),
+ weight=torch.from_numpy(weight),
+ # bias=torch.from_numpy(bias),
+ offset=torch.from_numpy(offset),
+ mask=torch.from_numpy(mask),
+ stride=int(self.stride),
+ padding=int(self.padding),
+ dilation=int(self.dilation),
+ )
+ top[0].data[...] = x.detach().cpu().numpy()
+
+ def backward(self, top, propagate_down, bottom):
+ for i in range(len(propagate_down)):
+ if not propagate_down[i]:
+ continue
+ bottom[i].diff[...] = top[i].diff[:]