Closed newExplore-hash closed 3 years ago
GoGoDuck912
commented
3 years ago Hi @newExplore-hash ,
In brief, you don't need to install the latest InPlaceABNSync by yourself. The cuda files in ./modules are compiled by Pytorch automatically.
As your own environmental problems, please refer to https://github.com/mapillary/inplace_abn/tree/v0.1.1 for some needed packages.
GoGoDuck912
commented
3 years ago Moreover, if you only need inference, you can replace the InPlaceABNSync layer with one BatchNorm2D layer and one LeakyRelu layer in PyTorch.
newExplore-hash
commented
3 years ago Moreover, if you only need inference, you can replace the InPlaceABNSync layer with one BatchNorm2D layer and one LeakyRelu layer in PyTorch.
yes, i just use this for inference, i know the role of InPlaceABNSync is to reduce the memory required for training deep networks. So i use nn.BatchNorm2d and nn.LeakyReLU instead of InPlaceABNSync in the script of AugmentCE2P.py for inference, but there is a problem as following:
Traceback (most recent call last):
File "simple_extractor.py", line 166, in
this error just happened in Decoder_Module for self.conv3, and there is normal for PSPModule and Edge_Module and other operations(self.conv1 and self.conv2) for Decoder_Module .
this is code: class Decoder_Module(nn.Module): """ Parsing Branch Decoder Module. """
def __init__(self, num_classes):
super(Decoder_Module, self).__init__()
self.conv1 = nn.Sequential(
nn.Conv2d(512, 256, kernel_size=1, padding=0, dilation=1, bias=False),
InPlaceABNSync(256),
#######
nn.LeakyReLU()
)
self.conv2 = nn.Sequential(
nn.Conv2d(256, 48, kernel_size=1, stride=1, padding=0, dilation=1, bias=False),
InPlaceABNSync(48),
########
nn.LeakyReLU()
)
self.conv3 = nn.Sequential(
nn.Conv2d(304, 256, kernel_size=1, padding=0, dilation=1, bias=False),
InPlaceABNSync(256),
########
nn.LeakyReLU(),
nn.Conv2d(256, 256, kernel_size=1, padding=0, dilation=1, bias=False),
InPlaceABNSync(256),
#########
nn.LeakyReLU()
)
self.conv4 = nn.Conv2d(256, num_classes, kernel_size=1, padding=0, dilation=1, bias=True)tks
fuyawangye
commented
3 years ago I modified "networks/AugmentCE2P.py" which replace C++ extended “InPlaceABNSync” for inferencing on cpu successful！
# from ..modules import InPlaceABNSync
# BatchNorm2d = functools.partial(InPlaceABNSync, activation='none')
BatchNorm2d = nn.BatchNorm2d
class ReplacePlaceABNSync(BatchNorm2d):
def __init__(self, *args, **kwargs):
super(ReplacePlaceABNSync, self).__init__(*args, **kwargs)
self.act = nn.LeakyReLU()
def forward(self, input):
output = super(ReplacePlaceABNSync, self).forward(input)
output = self.act(output)
return output
@newExplore-hash
modifiy "networks/AugmentCE2P.py" as follow is better
# from ..modules import InPlaceABNSync
# BatchNorm2d = functools.partial(InPlaceABNSync, activation='none')
BatchNorm2d = nn.BatchNorm2d
class InPlaceABNSync(BatchNorm2d):
def __init__(self, *args, **kwargs):
super(InPlaceABNSync, self).__init__(*args, **kwargs)
self.act = nn.LeakyReLU()
def forward(self, input):
output = super(InPlaceABNSync, self).forward(input)
output = self.act(output)
return output
very appreciate your awesome work, but according to your brief description of the installation requirements, i found that this program can't run at all, and various bugs will appear. just like the error of Segmentation fault (core dumped) caused by InPlaceABNSync. i use batchnorm instead of InPlaceABNSync, but i found the performance is very bad.
tks