CPU 1060
CUDA 10.0
PADLE 2.0
ubutun 16.04
docker 2.0rc linux 10.0
code:

!/usr/bin/env python3

from paddle.fluid.initializer import Constant
from paddle.fluid.param_attr import ParamAttr
import paddle
import paddle.fluid as fluid
import cv2
import numpy as np
from time import time
import paddle.nn.functional as F

class DcnModel(paddle.nn.Layer):
definit(self, name_scope, in_feature_number, out_feature_number, kernel_size):
super(DcnModel, self).init()
self.feature_number = in_feature_number
self.kernel_size = kernel_size
self.dcn_offset = paddle.nn.Conv2D(weight_attr=name_scope + '.offset',
bias_attr=name_scope + '.offset.bias',
in_channels=in_feature_number,
out_channels=kernel_size * kernel_size * 2,
kernel_size=kernel_size,
stride=1,
padding=kernel_size//2
)
self.dcn_amplitude = paddle.nn.Conv2D(weight_attr=name_scope + '.amplitude',
bias_attr=name_scope + '.amplitude.bias',
in_channels=in_feature_number,
out_channels=kernel_size*kernel_size,
kernel_size=kernel_size,
stride=1,
padding=kernel_size//2)
self.name_scope = name_scope

self.dcn_kernel = self.create_parameter(shape=[128, 128, 3, 3], attr=name_scope + '.kernel', dtype='float32')

self.dcn_name = name_scope + '.kernel'

self.dcn_conv = paddle.stastic.nn.deform_conv2d (x, offset, mask, num_filters, filter_size, stride=1, padding=0, dilation=1, groups=1, deformable_groups=1, im2col_step=1, weight_attr=None, bias_attr=None, name=None)

def forward(self, inputs):
B, C, H, W = inputs.shape
K = self.kernel_size

get offset [B, K * K * 2, H, W] via self.conv_offset

offset = self.dcn_offset(inputs)

reshape offset to [B, H, W, K * K, 2]

offset = offset.transpose((0, 2, 3, 1)).reshape((B, H, W, K * K, 2))

get amplitute [B, K*K, H, W]

amplitute = self.dcn_amplitude(inputs)
amplitute = F.sigmoid(amplitute)

dcn_kernel = np.array(self.dcn_kernel)

print(self.name_scope + '.kernel')

out_feature = paddle.fluid.layers.deformable_conv(
input= inputs,
offset=offset,
mask=amplitute,
num_filters=C,
filter_size=K,
stride=1,
padding=(K - 1) // 2,
groups=1,
deformable_groups=1,
im2col_step=1,
param_attr=ParamAttr(name=self.name_scope + '.kernel'),

param_attr=fluid.initializer.NumpyArrayInitializer(dcn_kernel),

bias_attr=False)
return out_feature

model = DcnModel(name_scope='dcn',
in_feature_number=128,
out_feature_number=128,
kernel_size=3)

x0 = paddle.ones((1,128,160,180))
i = 0
while(i < 100):
start1 = time()
out = model(x0)
out = out.numpy()
end1 = time()
print('Processed in {:0.03f} sec.'.format(end1 - start1))
i += 1