Version 1.2.0
Version 1.2.0 is now available, 6 June, 2018. Download SINGA v1.2.0
-diff --git a/content/404.html b/content/404.html index 0d4626b71..6111f46f5 100644 --- a/content/404.html +++ b/content/404.html @@ -4,7 +4,7 @@ {"zIndex":100} ) }); -


Version 1.2.0 is now available, 6 June, 2018. Download SINGA v1.2.0
-

Version 1.2.0 is now available, 6 June, 2018. Download SINGA v1.2.0
-

SINGA was presented at DIBRIS, University of Genoa, Italy on 16 July 2018.9.
-

SINGA was presented at DIBRIS, University of Genoa, Italy on 16 July 2018.9.
-

SINGA was presented at DISI, University of Trento, Italy on 14 December 2018.
-

SINGA was presented at DISI, University of Trento, Italy on 14 December 2018.
-

Version 2.0.0 is now available, 20 April, 2019. Download SINGA v2.0.0
-

Version 2.0.0 is now available, 20 April, 2019. Download SINGA v2.0.0
-

SINGA participated at EU FOSSA Apache Hackathon in Brussels on 4 and 5 May 2019.
-

SINGA participated at EU FOSSA Apache Hackathon in Brussels on 4 and 5 May 2019.
-

SINGA participated in mentoring the Google Summer of Code 2019 project SpamAssassin: Statistical Classifier Plugin.
-

SINGA participated in mentoring the Google Summer of Code 2019 project SpamAssassin: Statistical Classifier Plugin.
-

Snapshots of the tutorial:



Snapshots of the tutorial:











Please check the overview page and detailed guides.
-

Please check the overview page and detailed guides.
-

Version 1.2.0 is now available, 6 June, 2018. Download SINGA v1.2.0
-

Version 1.1.0 is now available, 12 Feb, 2017. Download SINGA v1.1.0
-





Hardware: we use p2.8xlarge instances from AWS, each of which has 8 Nvidia Tesla K80 GPUs, 96 GB GPU memory in total, 32 vCPU, 488 GB main memory, 10 Gbps network bandwidth.
Metric: we measure the time per iteration for different number of workers to evaluate the scalability of SINGA. The batch size is fixed to be 32 per GPU. Synchronous training scheme is applied. As a result, the effective batch size is $32N$, where N is the number of GPUs. We compare with a popular open source system which uses the parameter server topology. The first GPU is selected as the server.
Scalability test. Bars are for the throughput; lines are for the communication cost.


Hardware: we use p2.8xlarge instances from AWS, each of which has 8 Nvidia Tesla K80 GPUs, 96 GB GPU memory in total, 32 vCPU, 488 GB main memory, 10 Gbps network bandwidth.
Metric: we measure the time per iteration for different number of workers to evaluate the scalability of SINGA. The batch size is fixed to be 32 per GPU. Synchronous training scheme is applied. As a result, the effective batch size is $32N$, where N is the number of GPUs. We compare with a popular open source system which uses the parameter server topology. The first GPU is selected as the server.
Scalability test. Bars are for the throughput; lines are for the communication cost.


If installed PySINGA from binary packages, e.g. debian or wheel, then you need to change the python interpreter, e.g., reset the $PATH to put the correct path of Python at the front position.
-

If installed PySINGA from binary packages, e.g. debian or wheel, then you need to change the python interpreter, e.g., reset the $PATH to put the correct path of Python at the front position.
-

cp python/singa/xx.py <path to conda>/lib/python3.7/site-packages/singa/
-

cp python/singa/xx.py <path to conda>/lib/python3.7/site-packages/singa/
-





work in progress
-

work in progress
-

然后,python接口提供了一个DistOpt类来封装一个optimizer对象,以执行基于MPI或Multiprocess的分布式训练。在初始化过程中,它创建了一个NCCL communicator对象(来自于上面小节提到的C接口),然后,DistOpt中的每一次all-reduce操作都会用到这个communicator对象。
在MPI或Multiprocess中,每个进程都有一个独立的rank,它给出了各个进程使用的GPU的信息。训练数据是被拆分的,因此每个进程可以根据一部分训练数据来评估子梯度。一旦每个进程的子梯度被计算出来,就可以将所有进程计算出的子梯度做all-reduce,得到总体随机梯度。
-

然后,python接口提供了一个DistOpt类来封装一个optimizer对象,以执行基于MPI或Multiprocess的分布式训练。在初始化过程中,它创建了一个NCCL communicator对象(来自于上面小节提到的C接口),然后,DistOpt中的每一次all-reduce操作都会用到这个communicator对象。
在MPI或Multiprocess中,每个进程都有一个独立的rank,它给出了各个进程使用的GPU的信息。训练数据是被拆分的,因此每个进程可以根据一部分训练数据来评估子梯度。一旦每个进程的子梯度被计算出来,就可以将所有进程计算出的子梯度做all-reduce,得到总体随机梯度。
-

Apache SINGA is an effort undergoing incubation at The Apache Software Foundation (ASF), sponsored by the name of Apache Incubator PMC. Incubation is required of all newly accepted projects until a further review indicates that the infrastructure, communications, and decision making process have stabilized in a manner consistent with other successful ASF projects. While incubation status is not necessarily a reflection of the completeness or stability of the code, it does indicate that the project has yet to be fully endorsed by the ASF.
-

Apache SINGA is an effort undergoing incubation at The Apache Software Foundation (ASF), sponsored by the name of Apache Incubator PMC. Incubation is required of all newly accepted projects until a further review indicates that the infrastructure, communications, and decision making process have stabilized in a manner consistent with other successful ASF projects. While incubation status is not necessarily a reflection of the completeness or stability of the code, it does indicate that the project has yet to be fully endorsed by the ASF.
-

















SINGA is released under Apache License Version 2.0
-

SINGA is released under Apache License Version 2.0
-







-

-

Q: When I run SINGA in Mac OS X, I got the error "Fatal Python error: PyThreadState_Get: no current thread Abort trap: 6"
A: This error happens typically when you have multiple versions of Python in your system, e.g, the one comes with the OS and the one installed by Homebrew. The Python linked by SINGA must be the same as the Python interpreter. You can check your interpreter by which python and check the Python linked by SINGA via otool -L <path to _singa_wrap.so>. This problem should be resolved if SINGA is installation via conda.


Q: When I run SINGA in Mac OS X, I got the error "Fatal Python error: PyThreadState_Get: no current thread Abort trap: 6"
A: This error happens typically when you have multiple versions of Python in your system, e.g, the one comes with the OS and the one installed by Homebrew. The Python linked by SINGA must be the same as the Python interpreter. You can check your interpreter by which python and check the Python linked by SINGA via otool -L <path to _singa_wrap.so>. This problem should be resolved if SINGA is installation via conda.
















Please replace 'model.bin' with the path to one of the checkpoint paths.
-

Please replace 'model.bin' with the path to one of the checkpoint paths.
-





-nthreads <int>: the number of threads to load data which feed to the model.
-

-nthreads <int>: the number of threads to load data which feed to the model.
-

$ python convert.py -h
-

$ python convert.py -h
-

Then we construct the GoogleNet using SINGA's FeedForwardNet structure. Note that we added a EndPadding layer to resolve the issue from discrepancy of the rounding strategy of the pooling layer between Caffe (ceil) and cuDNN (floor). Only the MaxPooling layers outside inception blocks have this problem. Refer to this for more detials.
-

Then we construct the GoogleNet using SINGA's FeedForwardNet structure. Note that we added a EndPadding layer to resolve the issue from discrepancy of the rounding strategy of the pooling layer between Caffe (ceil) and cuDNN (floor). Only the MaxPooling layers outside inception blocks have this problem. Refer to this for more detials.
-

$ python convert.py --file_name=inception_v4.ckpt
-

$ python convert.py --file_name=inception_v4.ckpt
-

$ python convert.py -h
-

$ python convert.py -h
-

$ python convert.py -h
-

$ python convert.py -h
-







































Users can report security vulnerabilities to SINGA Security Team Mail List
-

Users can report security vulnerabilities to SINGA Security Team Mail List
-













work in progress
-

work in progress
-

























Scalability test. Bars
are for the throughput; lines are for the communication cost.


Scalability test. Bars
are for the throughput; lines are for the communication cost.





