Overview of GhostNet
GhostNet is a type of convolutional neural network that utilizes Ghost modules, resulting in greater efficiency and increased features with fewer parameters. GhostNet is mainly made up of a stack of Ghost bottlenecks, which are grouped into different stages based on the size of their input feature maps. The final stage uses a global average pooling and a convolutional layer to transform the feature maps to a 1280-dimensional feature vector for final classification.
What are Ghost modules?
Ghost modules are a type of neural network module that aim to generate more features by using fewer parameters. They are used in the GhostNet architecture to create more efficient neural networks. Ghost modules accomplish this by using a ghost feature map that is connected to a lower-dimensional weight matrix. This allows the neural network to learn richer feature representations while using fewer parameters and being more computationally efficient.
How is GhostNet constructed?
The GhostNet architecture is primarily constructed using Ghost bottlenecks, which are used to create a more efficient network. The first layer of GhostNet is a standard convolutional layer with 16 filters, followed by a series of Ghost bottlenecks with gradually increased channels. These bottlenecks are grouped into different stages based on the size of their input feature maps. All the Ghost bottlenecks have a stride of 1, except for the last bottleneck in each stage, which has a stride of 2. Finally, a global average pooling and a convolutional layer are applied to transform the feature maps into a 1280-dimensional feature vector for final classification.
What is the squeeze and excite (SE) module?
The squeeze and excite (SE) module is an optional component that can be applied to the residual layer in some Ghost bottlenecks. The SE module is used to enhance the representation power of the network by explicitly modeling feature interdependencies. This can improve the performance of the network by allowing it to learn more discriminative features.
How does GhostNet compare to other neural networks?
GhostNet is unique because it does not use hard-swish nonlinearity function, unlike MobileNetV3. This decision was made due to the large latency that comes along with using hard-swish. Instead, GhostNet uses different techniques that allow for increased efficiency while still producing high-quality feature representations.
GhostNet is a type of convolutional neural network that utilizes Ghost modules to generate more features by using fewer parameters. The GhostNet architecture is primarily constructed using Ghost bottlenecks, which are grouped into different stages based on the size of their input feature maps. GhostNet also uses a squeeze and excite (SE) module to enhance feature interdependencies if necessary. It differs from MobileNetV3 in that it does not use hard-swish nonlinearity function to avoid latency issues.
