Atomo: Communication-efficient Learning via Atomic Sparsification

This repository contains source code for Atomo, a general framework for atomic sparsification of stochastic gradients. Please check the full paper for detailed information about this project.

Overview:

ATOMO is a general framework for atomic sparsification of stochastic gradients. Given a gradient, an atomic decomposition, and a sparsity budget, ATOMO gives a random unbiased sparsification of the atoms minimizing variance. ATOMO sets up and optimally solves a meta-optimization that minimizes the variance of the sparsified gradient, subject to the constraints that it is sparse on the atomic basis, and also is an unbiased estimator of the input.

Depdendencies:

Tested stable depdencises:

python 2.7 (Anaconda)
PyTorch 0.3.0 (please note that, we're moving to PyTorch 0.4.0, and 1.0.x)
torchvision 0.1.18
MPI4Py 0.3.0
python-blosc 1.5.0

We highly recommend installing an Anaconda environment. You will get a high-quality BLAS library (MKL) and you get a controlled compiler version regardless of your Linux distro.

We provide this script to help you with building all dependencies. To do that you can run:

bash ./tools/pre_run.sh

Cluster Setup:

For running on distributed cluster, the first thing you need do is to launch AWS EC2 instances.

Launching Instances:

This script helps you to launch EC2 instances automatically, but before running this script, you should follow the instruction to setup AWS CLI on your local machine. After that, please edit this part in ./tools/pytorch_ec2.py

cfg = Cfg({
    "name" : "PS_PYTORCH",      # Unique name for this specific configuration
    "key_name": "NameOfKeyFile",          # Necessary to ssh into created instances
    # Cluster topology
    "n_masters" : 1,                      # Should always be 1
    "n_workers" : 8,
    "num_replicas_to_aggregate" : "8", # deprecated, not necessary
    "method" : "spot",
    # Region speficiation
    "region" : "us-west-2",
    "availability_zone" : "us-west-2b",
    # Machine type - instance type configuration.
    "master_type" : "m4.2xlarge",
    "worker_type" : "m4.2xlarge",
    # please only use this AMI for pytorch
    "image_id": "ami-xxxxxxxx",            # id of AMI
    # Launch specifications
    "spot_price" : "0.15",                 # Has to be a string
    # SSH configuration
    "ssh_username" : "ubuntu",            # For sshing. E.G: ssh ssh_username@hostname
    "path_to_keyfile" : "/dir/to/NameOfKeyFile.pem",

    # NFS configuration
    # To set up these values, go to Services > ElasticFileSystem > Create new filesystem, and follow the directions.
    #"nfs_ip_address" : "172.31.3.173",         # us-west-2c
    #"nfs_ip_address" : "172.31.35.0",          # us-west-2a
    "nfs_ip_address" : "172.31.14.225",          # us-west-2b
    "nfs_mount_point" : "/home/ubuntu/shared",       # NFS base dir

For setting everything up on EC2 cluster, the easiest way is to setup one machine and create an AMI. Then use the AMI id for image_id in pytorch_ec2.py. Then, launch EC2 instances by running

python ./tools/pytorch_ec2.py launch

After all launched instances are ready (this may take a while), getting private ips of instances by

python ./tools/pytorch_ec2.py get_hosts

this will write ips into a file named hosts_address, which looks like

172.31.16.226 (${PS_IP})
172.31.27.245
172.31.29.131
172.31.18.108
172.31.18.174
172.31.17.228
172.31.16.25
172.31.30.61
172.31.29.30

After generating the hosts_address of all EC2 instances, running the following command will copy your keyfile to the parameter server (PS) instance whose address is always the first one in hosts_address. local_script.sh will also do some basic configurations e.g. clone this git repo

bash ./tool/local_script.sh ${PS_IP}

SSH related:

At this stage, you should ssh to the PS instance and all operation should happen on PS. In PS setting, PS should be able to ssh to any compute node, this part dose the job for you by running (after ssh to the PS)

bash ./tools/remote_script.sh

Prepare Datasets

We currently support MNIST and Cifar10 datasets. Download, split, and transform datasets by (and ./tools/remote_script.sh dose this for you)

bash ./src/data_prepare.sh

Job Launching

Since this project is built on MPI, tasks are required to be launched by PS (or master) instance. run_pytorch.sh wraps job-launching process up. Commonly used options (arguments) are listed as following:

Argument	Comments
`n`	Number of processes (size of cluster) e.g. if we have P compute node and 1 PS, n=P+1.
`hostfile`	A directory to the file that contains Private IPs of every node in the cluster, we use `hosts_address` here as mentioned before.
`lr`	Inital learning rate that will be use.
`momentum`	Value of momentum that will be use.
`network`	Types of deep neural nets, currently `LeNet`, `ResNet-18/32/50/110/152`, and `VGGs` are supported.
`dataset`	Datasets use for training.
`batch-size`	Batch size for optimization algorithms.
`test-batch-size`	Batch size used during model evaluation.
`comm-type`	A fake parameter, please always set it to be `Bcast`.
`num-aggregate`	Number of gradients required for the PS to aggregate.
`max-steps`	The maximum number of iterations to train.
`svd-rank`	The expected rank of gradients ATOMO used (which is the same as the sparsity budget `s` in our paper).
`epochs`	The maximal number of epochs to train (somehow redundant).
`eval-freq`	Frequency of iterations to evaluation the model.
`enable-gpu`	Training on CPU/GPU, if CPU please leave this argument empty.
`train-dir`	Directory to save model checkpoints for evaluation.

Model Evaluation

Distributed evaluator will fetch model checkpoints from the shared directory and evaluate model on validation set. To evaluate model, you can run

bash ./src/evaluate_pytorch.sh

with specified arguments.

Evaluation arguments are listed as following:

Argument	Comments
`eval-batch-size`	Batch size (on validation set) used during model evaluation.
`eval-freq`	Frequency of iterations to evaluation the model, should be set to the same value as run_pytorch.sh.
`network`	Types of deep neural nets, should be set to the same value as run_pytorch.sh.
`dataset`	Datasets use for training, should be set to the same value as run_pytorch.sh.
`model-dir`	Directory to save model checkpoints for evaluation, should be set to the same value as run_pytorch.sh.

Future Work

Those are potential directions we are actively working on, stay tuned!

Explore the use of Atomo with Fourier decompositions, due to its utility and prevalence in signal processing.
Examine how we can sparsify and compress gradients in a joint fashion to further reduce communication costs.
Explore jointly sparsification of the SVD and and its singular vectors.
Integrate ATOMO to state-of-the-art PS (or distributed) frameworks e.g. Ray.

Citation

@inproceedings{wang2018atomo,
  title={ATOMO: Communication-efficient Learning via Atomic Sparsification},
  author={Wang, Hongyi and Sievert, Scott and Liu, Shengchao and Charles, Zachary and Papailiopoulos, Dimitris and Wright, Stephen},
  booktitle={Advances in Neural Information Processing Systems},
  pages={9871--9882},
  year={2018}
}

##changelog ###2019-5-21

在LSS_SGD_cluster分支上测试cluster方法的效果，详细见该分支日志 ###2019-5-22
取resnet18在cifar10上的梯度，batchsize 128*8，每10个step取一次，在cifar100上的梯度，batchsize 128*8，每200个step取一次
安装pytorch0.4.0测试在qsgd上的速度，decode速度还是比pytorch0.3.0慢了10倍，而且在cifar10上SVD方法mpi仍然会报错
将梯度画成图，卷积层基本集中分布在0，bn和shotcut参数高斯分布，全连接层的分布比较分散
pytorch0.3.0imagenet训练需要测试Torchvision的函数是否可用
用真实的梯度数据测试，试着解决数据偏斜问题中发现的一些情况：类簇的entropy相似3.**；bucket增大压缩时间减少，cluster数量增大，压缩时间增大，总的bucketsize可以作为一个tradeoff；clustercenter可以作为调节对聚类中心数据大小自适应能力的参数，应该在getarraysize中考虑；真实的梯度中类簇之间的确是数据分布不均匀，但是这种不均匀和梯度的不均匀不成比例，猜测是因为把太大的数值舍弃了的缘故（大数值应该在梯度训练中被考虑，而不是被认为是离群点）
imagenet数据resnet应该是224，改用官方模型吧 ###2019-5-23
pytorch0.3.0训练imagenet遇到MPI Bcast Truncted错误
将每个数据画成图,全连接层数据变化范围较大
读sketchML论文
分别统计大数值和小数值的误差,大数值的误差比小数值少十倍。

big num error 0.20622697472572327, small num error 2.626340389251709

###2019-5-24

读sketchML论文，思考每个类簇中如何降低每个bucket的误差。目前来看，大数值相对误差还可以控制，但是小数值的相对误差就有些大了，见下表，包含小数值的最大误差都很大；但是根据图片来看，大误差的数据占比很小；而且如何降低positive error，尤其是大数值的positive error，目前来看，正负误差数量差不多；问题的关键在于hash到bucket之后每个桶里的数据是平均的，而且每个类簇的范围似乎还挺大的。
下一步可以把采样的数据聚类结果中，每一类的数据的直方图画出来,看看类簇范围
在实际做实验的时候应该注意设置聚类采样率，并且观察小数值的误差对梯度下降的影响

positive_error.max()	np.abs(negative_error).max()	small_score_error.max()	big_score_error.max()
115943.33	90833.9	115943.33	0.7291682

理论证明部分，需要论证解压后的梯度与随机梯度之间的误差平方的期望有个上界，这个上界小于某个和随机梯度相关的数值
深度压缩感知的论文应该仔细读懂

###2019-5-31

entropy计算方法
二分查找
归一化
离群点用大的arraysize ###2019-6-3
离群点entropy计算，用所有的数据点？
如果不将离群点单独计算，会造成entropy=0，部分bucket里为空，需要修改getlsstable方法才能保证程序正确
数据归一化之后效果并不明显，修改arraysize计算方法，去掉density，但是大数值的arraysize=1
如何设计getarraysize的方法，是的大数值能够分得bucket多一些，而不是仅仅考虑到其entropy很小，density有没有用？
二值网络与signSGD一起训练
自适应的梯度量化方法
压缩感知

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

README.md

README.md

Atomo: Communication-efficient Learning via Atomic Sparsification

Overview:

Depdendencies:

Cluster Setup:

Launching Instances:

SSH related:

Prepare Datasets

Job Launching

Model Evaluation

Future Work

Citation

Files

README.md

Latest commit

History

README.md

File metadata and controls

Atomo: Communication-efficient Learning via Atomic Sparsification

Overview:

Depdendencies:

Cluster Setup:

Launching Instances:

SSH related:

Prepare Datasets

Job Launching

Model Evaluation

Future Work

Citation