A Decentralized Distributed DL Architecture in the Multi-interface Network.
There are three main folders:
model
: the models used in DL training, incluing common CNN, AlexNet, VGG19, Inception-V1/V3 and ResNeXt101/152.8bit_quantization
: data quantization by transferring numbers from FP32 into INT8 format.QAT_for_FP_BP
: The 8-bit Quantization-aware-Training (QAT) for both forward (papameter quantization) and backward (gradient sketch) propagation stages.topology
: PS on FatTree and AllReduce on BCube.
Three classical datasets are supported: MNIST, Fashion MNIT, and CIFAR-10.
- MNIST: The demos have already contained MNIST in the directory
data
, you can also download it from http://yann.lecun.com/exdb/mnist/ - Fashion MNIST: you can download it from https://github.com/zalandoresearch/fashion-mnist.
- CIFAR-10: You can download it from https://www.cs.toronto.edu/~kriz/cifar.html.