You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 
 
 
daquexian d93aa95078 Improve err msg in onnx2bnn 4 years ago
.daq_pm/configs Add run_net_x86 configuration 4 years ago
benchmark Add quick benchmark 4 years ago
binaries Fix the wrong new operator 4 years ago
ci Fix wrong .so name in Net.java 4 years ago
cmake Only specify libcxx for google benchamrk on Android 4 years ago
common Add support for prelu 4 years ago
dabnn Add support for prelu 4 years ago
docs FIX bash git --recursive 4 years ago
images Update imgs 4 years ago
tests move some missing net->read after setting optimize field 4 years ago
third_party Eliminate dropout by custom onnx optimizer 4 years ago
tools Improve err msg in onnx2bnn 4 years ago
.clang-format init 4 years ago
.gitignore Add quick benchmark 4 years ago
.gitmodules Move flatbuffers to submodule 4 years ago
CMakeLists.txt Move flatbuffers to submodule 4 years ago
LICENSE init 4 years ago
README.md Add qq group in english readme 4 years ago
README_CN.md Move model conversion from wiki to docs, move benchmark and comparison to a separate doc 4 years ago

我们推出了 dabnn,一个用 ARM 汇编重写了卷积,高度优化的二值网络 inference 框架。实验显示 dabnn 相对于现有的二值网络和浮点网络 inference 框架均能带来极大的加速。在运行单个二值卷积时,我们的速度是 BMXNet (除 dabnn 外唯一一个二值网络 inference 框架) 的 800%~2400%,在运行二值化的 ResNet-18 时,我们的速度是 BMXNet 的约 700%。

C++ C Markdown Text Shell other