Coder Social home page Coder Social logo

amrege / caffe-fast-rcnn Goto Github PK

View Code? Open in Web Editor NEW

This project forked from rbgirshick/caffe-fast-rcnn

0.0 1.0 0.0 34.14 MB

Caffe fork that supports Fast R-CNN

License: Other

CMake 2.48% Makefile 0.45% Shell 0.19% C++ 86.34% MATLAB 0.51% M 0.01% Python 4.30% Cuda 3.15% Protocol Buffer 1.85% C 0.59% M4 0.10% Objective-C 0.04%

caffe-fast-rcnn's Introduction

Caffe

Build Status License

Caffe is a deep learning framework made with expression, speed, and modularity in mind. It is developed by the Berkeley Vision and Learning Center (BVLC) and community contributors.

Check out the project site for all the details like

and step-by-step examples.

Join the chat at https://gitter.im/BVLC/caffe

Please join the caffe-users group or gitter chat to ask questions and talk about methods and models. Framework development discussions and thorough bug reports are collected on Issues.

Happy brewing!

Intel Caffe

This fork is dedicated to improving Caffe performance when running on CPU, in particular Xeon servers.

Performance Results

Time measures are average Forward-Backward as stated by caffe time. The speedup factor is calculated as (bvlc-caffe-master branch measure) / (intelcaffe-master branch measure).

Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz (36 threads, MKL 11.3, GCC 4.8.3)

Branch googlenet [ms] caffenet [ms] alexnet [ms] cifar10-bn [ms]
intelcaffe-master 493 1163 1260 20
bvlc-caffe-master 3990 6917 7307 270
speedup factor x8.1 x5.9 x5.8 x13.5

Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz (36 threads, OpenBLAS 0.2.14, GCC 4.8.3)

Branch googlenet [ms] caffenet [ms] alexnet [ms] cifar10-bn [ms]
intelcaffe-master 4066 2780 4011 45
bvlc-caffe-master 12992 25613 68415 494
speedup factor x3.1 x9.2 x17.1 x11.0

Tests were made using MKL and OpenBLAS. Please note that MKL is now available free of charge. The speedup factor highly depends on the amount of running caffe threads, system load and CPU temperature.

A following models are used to perform above performance tests:

./build/tools/caffe time --model=models/bvlc_alexnet/train_val.prototxt -iterations 1000
./build/tools/caffe time --model=models/bvlc_googlenet/train_val.prototxt -iterations 1000
./build/tools/caffe time --model=models/bvlc_reference_caffenet/train_val.prototxt -iterations 1000
./build/tools/caffe time --model=examples/cifar10/cifar10_full_sigmoid_train_test_bn.prototxt -iterations 10000

For lower amount of iterations (e.g. 10 times lower) to obtain relible results, all samples for about 15 seconds since test startup should be ignored when calculating average time.

Building

Build procedure is the same as on bvlc-caffe-master branch. Both Make and CMake can be used. When OpenMP is available will be used automatically.

Running

Run procedure is the same as on bvlc-caffe-master branch.

Current implementation uses OpenMP threads. By default the number of OpenMP threads is set to the number of CPU cores. Each one thread is bound to a single core to achieve best performance results. It is however possible to use own configuration by providing right one through OpenMP environmental variables like OMP_NUM_THREADS or GOMP_CPU_AFFINITY.

If some system tool like numactl is used to control CPU affinity, by default caffe will prevent to use more than one thread per core. When less than required cores are specified, caffe will limit execution of OpenMP threads to specified cores only.

Multinode Training

Please see the example how to run in examples/cifar10/train_full_multinode.sh. The script will run data server, synchronous parameter server and 4 clients. Prepared proto solvers should result in exactly the same behavior as single node full cifar training. The basic setup is to run parameter server with command like this: "$TOOLS/caffe param_server --solver=/path/to/proto --listen_address=tcp://:port" Than run clients on machines you want with: "$TOOLS/caffe train --solver=/path/to/proto --param_server=tcp://127.0.0.1:7777" The udp protocol can be used as well, for point to point communication and with multicast (i.e. "udp://127.0.0.1:7777;239.1.1.1:7778"). It is also possible to run the scheme with mpi with mpirun command, i.e: "mpirun
-host localhost -n 1
$TOOLS/caffe param_server --solver=/path/to/proto --listen_address=mpi://uid
:
-host localhost -n 1
$TOOLS/caffe train --solver=/path/to/proto --param_server=mpi://uid" You can run relay points, to accumulate/broadcast data in a tree structure: "$TOOLS/caffe param_server --solver=/path/to/proto --listen_address=tcp://
:port --param_server=tcp://127.0.0.1:7777" It only works with tcp protocol.

The mpi setup with explicit all reduce is with command like this: "mpirun -host 127.0.0.1 -n 5 $TOOLS/caffe train --solver=/path/to/proto --param_server=mpi" This will run 5 processes on hosts set with host, and each process will calculate it's own gradients, and propagate it up with a tree structure to the root, which will apply them and propagate parameters down also in a tree structure. This version is less configurable than one with param server, relay and client, however it uses less cpu resource per node and can get most of the mpi implementations.

Data server is for convenience. By the default you could use data shard prepared on each node separetely, either by shuffling the data uniquely or by creating a subset of your training data. The remote data layer can be used to get data from data server. It can also be used to cache data from the server in order to reduce the network traffic. Use only tcp protocol with data server. In the case of choosing caching policy USE_CACHE_WHEN_FULL, it will first download cache_size batches and then will randomized the cached data for actual training.

The proto files need to be set up manually at the time, although you can use model server to distribute some proto files among clients. To run model server use the caffe tool similar to data server: "$TOOLS/caffe model_server --solver=/path/to/proto --listen_address=tcp://*:6666" To use the model in clients, replace the path to solver with model server address: "$TOOLS/caffe train --solver=address"

Please see also prepared examples (for 2 nodes only) for googlenet in: models/bvlc_googlenet/solver_param_server.prototxt models/bvlc_googlenet/solver_client.prototxt The solver tries to offset the bigger batch size with bigger learning rate. According to paper @article{ Author = {Forrest N. Iandola, Khalid Ashraf, Matthew W. Moskewicz, Kurt Keutzer}, Journal = {arXiv preprint arXiv:1511.00175}, Title = {FireCaffe: near-linear acceleration of deep neural network training on compute clusters}, Year = {2016} } this should use 72 epochs to train googlenet.

License and Citation

Caffe is released under the BSD 2-Clause license. The BVLC reference models are released for unrestricted use.

Please cite Caffe in your publications if it helps your research:

@article{jia2014caffe,
  Author = {Jia, Yangqing and Shelhamer, Evan and Donahue, Jeff and Karayev, Sergey and Long, Jonathan and Girshick, Ross and Guadarrama, Sergio and Darrell, Trevor},
  Journal = {arXiv preprint arXiv:1408.5093},
  Title = {Caffe: Convolutional Architecture for Fast Feature Embedding},
  Year = {2014}
}

caffe-fast-rcnn's People

Watchers

 avatar

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    ๐Ÿ–– Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. ๐Ÿ“Š๐Ÿ“ˆ๐ŸŽ‰

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google โค๏ธ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.