Hello,I use cuda 10.0 and meet the question
//---------------------------------------------------------------
File "/home/hjl/object-detection/DNL-Object-Detection/mmdet/utils/flops_counter.py", line 35, in
from mmdet.ops.dcn import DeformConv, ModulatedDeformConv
File "/home/hjl/object-detection/DNL-Object-Detection/mmdet/ops/init.py", line 3, in
from .dcn import (DeformConv, DeformConvPack, DeformRoIPooling,
File "/home/hjl/object-detection/DNL-Object-Detection/mmdet/ops/dcn/init.py", line 1, in
from .deform_conv import (DeformConv, DeformConvPack, ModulatedDeformConv,
File "/home/hjl/object-detection/DNL-Object-Detection/mmdet/ops/dcn/deform_conv.py", line 9, in
from . import deform_conv_cuda
ImportError: libcudart.so.9.2: cannot open shared object file: No such file or directory
//--------------------------------------------------------------------------------------------------------------

is it means I should back cuda to 9.0 ?while the inplace_abn cuda should >10.0

//---------------------------------------------------------------------------------------
NOTE 1: our code requires PyTorch v1.1 or later

NOTE 2: we are only able to provide support for Linux platforms and CUDA versions >= 10.0

NOTE 3: in general, it is not possible to load weights from a network trained with standard BN into an InPlace-ABN network without severe performance degradation, due to the different handling of BN scaling parameters

To install the package containing the iABN layers:

pip install inplace-abn

Hello:
Thanks for your perfect work. I have a question about that which py do you compute unary term in?

Thank you very much!

Hello,thanks for your perfect work.
I have some question about Eq.2 ,μq is the mean of all query samples ,but i don't know which dimension it is acculate mean on ,channel dimension or the height*width weigth ?

thank you very much!

I cannot access your model and log in the SharePoint because I am not the internal user, will you open it ?

Thank your for great project. I have not find unary branch in your code. This DNL modules non-complete?

Hello,
I have a question about your paper 'Disentangled Non-local Neural Networks' .

I tried to derive the formula equation (17), equation (18), but I failed. Could you explain why diagonal matrix of A^T*A is derieved to equation (17)??

Thank you.

Hello,can I change the model in faster-rcnn resnet instead of mask rcnn only by change this demo?
//---------------------------------------------------------------------------------------------------------------------

model settings

model = dict(
type='FasterRCNN',
pretrained='torchvision://resnet50',
backbone=dict(
type='ResNet',
depth=50,
num_stages=4,
out_indices=(0, 1, 2, 3),
frozen_stages=1,
#norm_cfg=dict(type='BN', requires_grad=True),
#norm_eval=True,
nlgcb=dict(ratio=1. / 4., downsample=False, whiten_type=['channel'], temp=0.05, with_gc=True, use_out=False,
out_bn=False),
stage_with_nlgcb=[[], [], [-2], [-2, -1, 0]],
style='pytorch'),
neck=dict(
type='FPN',
in_channels=[256, 512, 1024, 2048],
out_channels=256,
num_outs=5),
rpn_head=dict(
type='RPNHead',
in_channels=256,
feat_channels=256,
anchor_scales=[8],
anchor_ratios=[0.5, 1.0, 2.0],
anchor_strides=[4, 8, 16, 32, 64],
target_means=[.0, .0, .0, .0],
target_stds=[1.0, 1.0, 1.0, 1.0],
loss_cls=dict(
type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
loss_bbox=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.0)),
bbox_roi_extractor=dict(
type='SingleRoIExtractor',
roi_layer=dict(type='RoIAlign', out_size=7, sample_num=2),
out_channels=256,
featmap_strides=[4, 8, 16, 32]),

hello,thanks for your fantastic work!
I have some questions in the Eq(12), the shape of the former is(b,THW,THW),the shape of the latter is (b,1,THW),how about the addation. sorry， I don't know whether my understanding is correct or not .
thank you very much!

Hello，thanks for your fantastic work.
i want to know why the output dimension of the unary Wm is 1 . why not 2 or 3?

thank you very much!

1, Eq (17): How the CauchySchwarz inequality can be applied to different vectors?
2, Eq (18): The orders of /alpha seems to be wrong.
3, Is there any definitions for the normalized differences between query and key pixels?

1. How does the first equal sign in Equation 9 hold?
2. Why can $\lambda_i$ be omitted in subsequent discussions?

there have not setup.py in mmdet/ops/roi_align thx