liuzhuang13/DenseNet
DenseNet
DenseNet is a computer vision model and convolutional neural network implementation designed for image recognition and classification tasks. It utilizes a densely connected network architecture where each layer is connected to every other layer to improve feature propagation.
The implementation reduces the number of parameters while maintaining accuracy through a dense-connectivity pattern and layer-aggregation concatenation. It supports model construction using both standard and bottleneck-compressed architectures, with configurable network depth and growth rates to balance inference time and memory usage.
The project covers convolutional neural network optimization and deep learning model training for visual tasks. It includes GPU memory optimization techniques, such as shared gradient inputs, to reduce the memory footprint during training.
Features
- Dense Neural Networks - Implements a densely connected network architecture where every layer connects to all subsequent layers for optimal feature propagation.
- Computer Vision Models - Provides a deep learning architecture tailored for image recognition and classification tasks.
- Computer Vision Research - Provides an implementation of DenseNet for evaluating performance on standard image recognition benchmarks.
- Convolutional Neural Network Architectures - Implements a deep convolutional architecture specifically designed to improve feature reuse for image classification.
- Growth Rate Controls - Controls the number of feature maps added by each layer to manage the model size and memory usage.
- CNN Architectures - Provides a complete convolutional neural network implementation based on the DenseNet architecture.
- Feature Map Concatenations - Combines feature maps from all preceding layers using concatenation instead of summation.
- CNN Hyperparameter Optimization - Allows tuning of network depth and growth rates to optimize memory usage and inference speed.
- Model Construction - Enables the construction of densely connected convolutional layers using both standard and bottleneck-compressed designs.
- Bottleneck Layers - Implements bottleneck layers that use 1x1 convolutions to compress feature maps and reduce computational overhead.
- Model Width Configurations - Deno DenseNet allows tuning the depth and growth rate of the network to balance inference time, memory usage, and accuracy.
- PyTorch Implementations - Implements the DenseNet model using the PyTorch framework for image recognition.
- Training Memory Optimizers - Employs shared gradient inputs to reduce the GPU memory footprint during the backpropagation process.
- Deep Neural Network Training Optimization - Optimizes the memory footprint during the training of densely connected layers on GPUs.
- Computer Vision - Densely connected convolutional networks.
- Image Classification Architectures - Original implementation of densely connected convolutional networks.
