Skip to content

Internal Imagenet normalisation for pretrained Densenet models #782

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 3 commits into from
Closed

Internal Imagenet normalisation for pretrained Densenet models #782

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
af2cd76
Internal imagenet normalisation for pretrained models like InceptionV…
ekagra-ranjan Mar 8, 2019
4c262fd
fixed flake8
ekagra-ranjan Mar 8, 2019
9f3140d
Updated docs of densnets with transform_input argument
ekagra-ranjan Mar 9, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 19 additions & 1 deletion torchvision/models/densenet.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,10 +69,12 @@ class DenseNet(nn.Module):
"""

def __init__(self, growth_rate=32, block_config=(6, 12, 24, 16),
num_init_features=64, bn_size=4, drop_rate=0, num_classes=1000):
num_init_features=64, bn_size=4, drop_rate=0, num_classes=1000, transform_input=False):

super(DenseNet, self).__init__()

self.transform_input = transform_input

# First convolution
self.features = nn.Sequential(OrderedDict([
('conv0', nn.Conv2d(3, num_init_features, kernel_size=7, stride=2, padding=3, bias=False)),
Expand Down Expand Up @@ -110,6 +112,14 @@ def __init__(self, growth_rate=32, block_config=(6, 12, 24, 16),
nn.init.constant_(m.bias, 0)

def forward(self, x):

# imagenet normalisation
if self.transform_input:
x_ch0 = (torch.unsqueeze(x[:, 0], 1) - 0.485) / 0.229
x_ch1 = (torch.unsqueeze(x[:, 1], 1) - 0.456) / 0.224
x_ch2 = (torch.unsqueeze(x[:, 2], 1) - 0.406) / 0.225
x = torch.cat((x_ch0, x_ch1, x_ch2), 1)

features = self.features(x)
out = F.relu(features, inplace=True)
out = F.adaptive_avg_pool2d(out, (1, 1)).view(features.size(0), -1)
Expand All @@ -123,6 +133,8 @@ def densenet121(pretrained=False, **kwargs):

Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
transform_input (bool): If True, preprocesses the input according to the method with which it
was trained on ImageNet. Default: *False*
"""
model = DenseNet(num_init_features=64, growth_rate=32, block_config=(6, 12, 24, 16),
**kwargs)
Expand Down Expand Up @@ -150,6 +162,8 @@ def densenet169(pretrained=False, **kwargs):

Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
transform_input (bool): If True, preprocesses the input according to the method with which it
was trained on ImageNet. Default: *False*
"""
model = DenseNet(num_init_features=64, growth_rate=32, block_config=(6, 12, 32, 32),
**kwargs)
Expand Down Expand Up @@ -177,6 +191,8 @@ def densenet201(pretrained=False, **kwargs):

Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
transform_input (bool): If True, preprocesses the input according to the method with which it
was trained on ImageNet. Default: *False*
"""
model = DenseNet(num_init_features=64, growth_rate=32, block_config=(6, 12, 48, 32),
**kwargs)
Expand Down Expand Up @@ -204,6 +220,8 @@ def densenet161(pretrained=False, **kwargs):

Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
transform_input (bool): If True, preprocesses the input according to the method with which it
was trained on ImageNet. Default: *False*
"""
model = DenseNet(num_init_features=96, growth_rate=48, block_config=(6, 12, 36, 24),
**kwargs)
Expand Down