Shypd.aiThe largest open index of AI toolsDiracNets is an open-source deep learning project for training very deep neural networks without skip-connections. It simplifies network architecture to a chain of convolution-ReLU layers, matching ResNet performance.
Trending
About
DiracNets provides PyTorch code and models for training very deep neural networks without the need for skip-connections, a common feature in architectures like ResNet. By using a simple weight parameterization, DiracNets allows for the training of plain networks with hundreds of layers. The project demonstrates that DiracNet-18 and DiracNet-34 can closely match the performance of corresponding ResNet models on ImageNet, while simplifying the network structure to a VGG-like chain of convolution-ReLU layers. The proposed Dirac weight parameterization can also be folded into a single filter for inference, making the resulting network easily interpretable. Pretrained models are available for download.
Capabilities
Pricing & Plans
Open Source
Free
FAQs
Trending
