An implementation of SRM block, proposed in "SRM : A Style-based Recalibration Module for Convolutional Neural Networks".

• Python >= 3.6
• PyTorch >= 1.1
• torchvision
• back > 0.0.3

back is PyTorch backbone for training loop.

For implementing channel-wise fully connected (CFC) layer I used Conv1d layer which is equal to CFC with next parameters:

Conv1d(channels, channels, kernel_size=2, groups=channels)

It turns out the use of depthwise 1d convolution.

# Cifar10
python cifar10_train.py --model_name srmnet

# ImageNet
python imagenet_train.py --model_name srmnet

# Logs
tensorboard --logdir=logs --host=0.0.0.0 --port=8080

batch_size = 128
epochs_count = 100
optimizer = optim.SGD(model.parameters(), lr=0.1, momentum=0.9,
                      weight_decay=1e-4)
scheduler = MultiStepLR(optimizer, [70, 80], 0.1)

batch_size = 64
epochs_count = 100
optimizer = optim.SGD(model.parameters(), lr=0.1, momentum=0.9,
                      weight_decay=1e-4)
scheduler = StepLR(optimizer, 30, 0.1)

Dark blue - ResNet

Blue - Se-ResNet

Green - SRM-ResNet

Weights for best models.