convnet-benchmarks

ConvNet Benchmarks is an open-source benchmarking suite designed to compare the computational performance of convolutional neural network implementations across machine learning frameworks. It provides reproducible tests for measuring forward and backward propagation times. The repository includes benchmarks for frameworks such as Caffe, Torch, TensorFlow, Theano, Chainer, and MXNet. It evaluates convolution operations using architectures associated with prominent ImageNet models, including AlexNet and VGG. Results distinguish forward execution time, backward execution time, and combined processing performance. Benchmark configurations document hardware specifications, batch sizes, and tensor dimensions. The project also compares CPU, GPU, cuDNN, and different numerical precision implementations. Its historical results help researchers study framework performance and optimization approaches.