Hi Krishna,
I just gave Kmeans a try with your parameters and it runs fine with
hybrid_spark (default) through spark submit. However, I'm able to reproduce
the issue when forcing it into singlenode. Thanks for catching this - I'll
take care of it.
Regards,
Matthias
On Sat, Jul 29, 2017 at 9:34
well, there are different flavors to this problem:
1) If you're only interested in integer values and in testing if all values
of A appear in B, you can do the following:
A = seq(3,4)
B = seq(2,11)
N = min(nrow(A),nrow(B))
M = min(max(A),max(B))
I1 = table(A,1,M,1)!=0;
I2 = table(B,1,M,1)!=0;
without any specifics of scripts or datasets, it's unfortunately, hard
if not impossible to help you here. However, note that the memory
configuration seems wrong. Why would you configure the driver and
executors with 2TB if you only have 256GB per node. Maybe you observe an
issue of swapping.
Well, we explicitly cleanup all intermediates that are no longer used. You
can use -explain to output the runtime plan, which includes rmvar (remove
variable), cpvar (copy variable) and mvvar (move variable) instructions
that internally cleanup intermediates. This cleanup removes data from
memory,
well, at a high-level, resource negotiation and distributed storage are
orthogonal concepts. Yarn, Mesos, Standalone, and Kubernetes are resource
schedulers, which you can configure via master and a separate deploy mode
(client/cluster). Under the covers of the HDFS API, you can also use
various
here is the list of open tasks (disregarding open bugs), that I think would
be good to have in SystemML 1.0:
SYSTEMML-1284 Code generation
-> all remaining subtasks
SYSTEMML-1299 Language features 1.0
-> 1301, 1304/1307, 1306, 1316, 1426, 1444
SYSTEMML-1308 Runtime reatures 1.0
-> 1313, 1637
yes, I absolutely agree - we should add special handling of functions with
single outputs.
Quick background: Functions are initially split into separate statement
blocks so they can bind their outputs (potentially many) directly to
logical variable names because Hop DAGs cannot represent multiple
ok this has been fixed in master - it was essentially an issue of size
propagation for ctable with two sequence inputs.
Regards,
Matthias
On Sat, Jul 29, 2017 at 1:57 PM, Matthias Boehm <mboe...@googlemail.com>
wrote:
> Hi Krishna,
>
> I just gave Kmeans a try with your paramet
Hi all,
we see a lot of scripts where conditional statement blocks split DAGs of
operations. After constant folding of if predicates, unnecessary branches
are already removed (which is important for size propagation) but we don't
merge sequences of statement blocks yet. Consider the following
Good question - let me separate the somewhat orthogonal aspects to it.
First, for descriptive statistics such as sum, mean, skewness, or kurtosis,
we already use numerically stable implementations based on Kahan Plus (see
org.apache.sysml.runtime.functionobjects.KahanPlus if your interested).
s”.
>
> Mahout Samsara is based on Scala. PredictionIO (predictionio.incubator.
> apache.org) algorithms are based on Mahout Samsara and Scala. I asked
> Mr. Matthias Boehm at a conference how one could compare Mahout Samsara to
> SystemML. From what I understood, Samsara needs "
kul Jindal <naku...@gmail.com> wrote:
> >
> > Hi Matthias,
> >
> > If its not too much trouble, could you please create a design document
> for
> > this change.
> > This will help the rest of the contributors work on this component as
> well.
>
Generally, the parfor dependency analysis applies a series of tests
including traditional techniques from high-performance compilers combined
with additional rules for common cases. This dependency analysis tries to
proof that there are no loop-carried dependencies - so yes, false positives
can
or
> committing reserved memory.
>
>
> I can send my Spark and YARN configurations as well if that would be
> useful. Thanks a lot for your help.
>
>
> Best,
>
>
> Anthony
>
> On Thu, Jun 15, 2017 at 3:00 PM, Anthony Thomas <ahtho...@eng.ucsd.edu>
> w
github.com/dusenberrymw
> LinkedIn: linkedin.com/in/mikedusenberry
>
> Sent from my iPhone.
>
>
> > On Jun 14, 2017, at 9:08 PM, Matthias Boehm <mboe...@googlemail.com>
> wrote:
> >
> > Generally, the parfor dependency analysis applies a se
just FYI: in the process of making SystemML work smoothly w/ object stores,
we fixed our handling of file URI schemes. So now, you can read/write
from/to different file systems, independent of the configured default fs
implementations. For example you can now do something like this
A =
Well, at a high-level, we could emulate synchronous model parallelism via
our existing parfor construct out of the box. If this is sufficient from an
algorithm perspective, I would be in favor of making any necessary
improvements there instead of introducing a new construct for parameter
servers.
a closer look.
Regards,
Matthias
On Fri, Sep 15, 2017 at 11:52 PM, Krishna Kalyan <krishnakaly...@gmail.com>
wrote:
> Thank you so much for trying this Matthias. I will try this again with
> absolute path.
>
> Regards,
> Krishna
>
> On Sat, Sep 16, 2017 at 12:
>
> On Sat, Sep 16, 2017 at 7:24 AM, Matthias Boehm <mboe...@googlemail.com>
> wrote:
>
> > well, I don't think any HDFS fs implementation resolves '~' - so it has
> > probably created a directory called '~/open-source/scripts/PCA_data' in
> > your user path
Currently, our SystemML configuration properties use an inconsistent prefix
scheme. For example, some properties use the prefix 'dml' (e.g.,
dml.yarn.appmaster), others 'systemml' (e.g., systemml.stats.finegrained),
and yet others no prefix at all (e.g., localtmpdir).
We discussed this before but
I pretty much agree with Niketan and Deron. In general, it would be useful
to provide an R API as well. However, I'm a bit concerned for two reasons:
* Looking over the github repo, apparently R4ML is not under active
development/maintenance anymore (last commit Jul 20). So who would be
willing
I guess this is a just a delayed message, right? Ted was right, it's a
generated class, and hence not in the repo. However, when you build
SystemML, the generated source ends up in the src directory of the
org.apache.sysml.parser.dml package as well.
Note that for new builtin functions such as
just FYI: using the exact parameters of the failing test (randomly
generated), I was able to reproduce this now locally - so we should be able
to fix this flaky test once and for all.
Regards,
Matthias
On Wed, Sep 20, 2017 at 2:21 PM, wrote:
> See
as mentioned on PR-673, I'm probably not the right person to comment on
algorithm or API-related changes, but I'll try to have a look tomorrow.
Regards,
Matthias
On Tue, Oct 3, 2017 at 6:52 AM, Janardhan Pulivarthi <
janardhan.pulivar...@gmail.com> wrote:
> Hi Matthias,
>
> Based on your
thanks Ted - this was my fault. A recent change toward a more aggressive
use of CSR revealed a number of hidden issues, which I already fixed last
night - so, the next build should be fine.
Regards,
Matthias
On Sun, Oct 15, 2017 at 6:36 AM, Ted Yu wrote:
> *There seems to
hich of the type` *pow(a, 2) `*),
> to be consistent with the other symbols of dml.
>
> In this simple case:
> 1. ` *a (+) b (+) c (+) d(+)...* `
> 2. ` *xor(xor(a, b), c)..) ` (*sorry, if I written this syntax wrongly)
>
> Your word will be final.
>
> Thanks,
> Janardhan
&
thanks again for catching these issues Rajarshi. I'd like to briefly
summarize their resolutions.
ad 1) Most likely this was caused by a configuration issue - specifically,
the default parallelism being set to less than the number of executors,
leading to the number of cores per executor getting
that's a fair point and we should all help to move these forward. I'll make
another pass over 3 and 4 today.
Regards,
Matthias
On Wed, Aug 23, 2017 at 9:30 AM, Janardhan Pulivarthi <
janardhan.pulivar...@gmail.com> wrote:
> Dear committers,
>
> I am feeling that my contributions are not in an
ot;
}, new String[] { "predicted_y" }, false): is the code I posted and I don't
understand how it differs from what you said on point 2:
"So please change it as follows:
conn.prepareScript(conn.readScript(dml), new String[] {"w", "X"}, new String[]
{"predicted_
As it turns out, SystemML master computes correct results, whereas SystemML
0.14 compiled invalid instructions for the initialization of centroids. The
details are in the jira and we can go ahead with the 0.15 release.
Regards,
Matthias
On Fri, Sep 1, 2017 at 11:40 PM, Matthias Boehm <m
Hi Nantia,
that's a good question - if your input data is very small, there is nothing
to worry about. For the explain output, the memory estimates are rounded to
MB. Therefore, whenever your input is less than 65,536 cells (dense), you
will see a memory estimate of 0 MB (e.g.,
try for the XOR operator, with caret ` ^ ` symbol.
> But, this have been reserved for exponentiation. So, another alternative
> would be
>
> 1. ` (+) `
> 2. ` >< `
> 3. ` >-< `
>
> Thanks,
> Janardhan
>
> On Thu, Aug 31, 2017 at 7:38 PM, Matthia
it as a blocker.
Regards,
Matthias
On Fri, Sep 1, 2017 at 7:37 PM, Matthias Boehm <mboe...@googlemail.com>
wrote:
> +1 - I also think that a 0.15 release is a good idea. Many improvements
> and fixes have been made since 0.14 but we're not quite ready for a 1.0
> release.
>
>
--
>
>
>
>
> +1 for the change. This will give us enough time to test CLA with different
> setting/algorithms/data characteristics before the release.
>
> > On May 28, 2017, at 2:42 PM, Matthias Boehm <mboe...@googlemail.com>
> wrote:
> >
> &
>From a scalar operation perspective, you could of course emulate XOR via
AND, OR, and negation. However, you might want to write anyway a java-based
UDF to efficiently implement this recursive operator.
Down the road, we can think about a generalization of our existing
cumulative operations such
would
recommend to defer the RC1 for a couple of days until these issues are
fixed as well.
Regards,
Matthias
On Tue, Nov 14, 2017 at 1:31 PM, Krishna Kalyan <krishnakaly...@gmail.com>
wrote:
> +1
>
> Regards,
> Krishna
>
> On Sun, Nov 12, 2017 at 5:53 AM, Matthias Boehm &
<http://researcher.watson.ibm.com/researcher/view.php?
> person=us-npansar>
> > >
> > > "Glenn Weidner" ---12/11/2017 09:49:48 AM---+1 I ran Linear Regression,
> > > Logistic Regression, SVM, Naive Bayes Python tests
> > >
> > > From:
yes, that's indeed the case. The original PR was fine but I introduced
this issue during some final cleanups while merging this PR.
Regards,
Matthias
On 12/17/2017 11:49 PM, Ted Yu wrote:
XorTest failure seems to be related to:
[SYSTEMML-1883] New xor builtin functions over scalars
On Sun,
unfortunately, our cdf and invcdf currently only support the distributions
normal, exp, chisq, f, and t and scalar inputs. So you would have to
emulate this at script level. Extending the list of distribution functions
and adding matrix support would be a good addition though.
Regards,
Matthias
Hi all,
we made some good progress regarding deep learning support, code
generation, and low-latency scoring - so, I'm looking forward to our
upcoming 1.0 release. Since it's our first stable release, I think it would
be a good idea to allocate some extra time for QA. How about we shoot for a
Hi all,
we made some good progress regarding deep learning support, code
generation, and low-latency scoring - so, I'm looking forward to our
upcoming 1.0 release. Since it's our first stable release, I think it would
be a good idea to allocate some extra time for QA. How about we shoot for a
-1 due to the issue mentioned by Niketan, as well as additional correctness
and performance issues fixed in the last couple of days.
Regards,
Matthias
On Tue, Dec 5, 2017 at 6:25 PM, Niketan Pansare wrote:
> Soft -1 as GPU backend is in experimental mode. GPU matrix
Hi all,
with our SystemML 1.0 release around the corner, I think we should start
the discussion on the roadmap for SystemML 1.1 and beyond. Below is an
initial list as a starting point, but please help to add relevant items,
especially for algorithms and APIs, which are barely covered so far.
1)
+1
I ran the perftest suite with the artifact on Spark 2.2 up to 80GB without
any failures or performance issues. On earlier versions, I also ran the
perftest suite with Spark 2.1 and 2.2, w/ and w/o codegen, and w/ auto
compression up to 800GB without remaining issues.
As a minor nitpick (to be
cannot find anything relevant to 'org.apache.sysml.hops.rewrite'. Is
> there
> another file I should check?
>
> Thanks again,
> Nantia
>
> 2017-10-13 23:29 GMT+03:00 Matthias Boehm <mboe...@googlemail.com>:
>
> > Hi Nantia,
> >
> > in optimization lev
Janardhan, could you please elaborate a little what issues you faced?
SystemML itself does not require any specific installation. Also to be
productive, you might want to setup a dev environment, where you can run
tests locally directly from your IDE.
Regards,
Matthias
On Mon, Oct 30, 2017 at
Hi Guobao,
that sounds very good. In general, the "model" refers to the
collection of all weights and bias matrices of a given architecture.
Similar to a classic regression model, we can view the weights as the
"slope", i.e., multiplicative terms, while the biases are the
"intercept", i.e.,
This particular JIRA is only partially related. Niketan and Nakul
worked out the details - the only reason I show up as the reporter is
that, if I remember correctly, we split a larger scoped JIRA for
low-level optimizations (GPU, codegen, compression) into individual
JIRAs and created the
entries via l1[7] or l2['g'] accordingly. We're still
working on additional features to make the integration with IPA,
functions, and size/type propagation smoother, but the basic
functionality is already available.
Regards,
Matthias
On Sun, May 6, 2018 at 1:08 PM, Matthias Boehm <mboe...@gmail.
Well, in general this can be interesting. Apart from our default
testsuite, we occasionally ran static code analysis tools. Having
additional tests for partially valid scripts and inputs can help to
find more issues.
That being said, I don't think we currently qualify as a project with
;
>
>
> From: Krishna Kalyan
> To: dev@systemml.apache.org
> Date: 06/05/2018 10:09 PM
> Subject:Re: Release Planning SystemML 1.2
>
>
>
> +1
>
> I am completely available to help with the QA cycle and help with
> switching
> to new perf test suite.
given the current status of open tasks and the delay with regard to
QA, I think we need to push this release out by a couple of weeks.
Does mid to end July sound good to everyone?
Regards,
Matthias
On Wed, Jun 6, 2018 at 11:28 PM, Matthias Boehm wrote:
> thanks Berthold - that sounds g
Hi all,
please join me in welcoming Guobao Li as a GSoC 2018 student, who will
be working on SYSTEMML-2083 (Language and runtime for parameter
servers) this summer. We're currently in the community bonding phase,
but the project will start May 14.
Krishna already kindly volunteered (on the dev
at org.apache.sysml.runtime.controlprogram.Program.
execute(Program.java:130)
at org.apache.sysml.api.mlcontext.ScriptExecutor.executeRuntimeProgram(
ScriptExecutor.java:388)
... 16 more
...
On Fri, Dec 22, 2017 at 5:48 AM, Matthias Boehm <mboe...@gmail.com> wrote:
well, let's do the following to figure this out:
1) If the schema is
r.java:624)
at java.lang.Thread.run(Thread.java:748)
Anthony
On Sat, Dec 23, 2017 at 4:27 AM, Matthias Boehm <mboe...@gmail.com> wrote:
Given the line numbers from the stacktrace, it seems that you use a rather
old version of SystemML. Hence, I would recommend to upgrade to SystemML
:624)
at java.lang.Thread.run(Thread.java:748)
Best,
Anthony
On Sun, Dec 24, 2017 at 3:14 AM, Matthias Boehm <mboe...@gmail.com> wrote:
Thanks again for catching this issue Anthony - this IJV reblock issue with
large ultra-sparse matrices is now fixed in master. It likely
s
> a dataframe of sparse vectors to a DML script without issue. Sorry for the
> slow confirmation on this - I've been out of the office for the last couple
> weeks. Thanks for your help debugging this!
>
> Best,
>
> Anthony
>
> On Mon, Dec 25, 2017 at 5:35 AM, Matthias
,
Matthias
On Sat, Jan 27, 2018 at 1:57 PM, Nakul Jindal <naku...@gmail.com> wrote:
> This is awesome!
> I am guessing the goal is to have this epic be a summer worth of
> mini-projects for a single GSoC student, isthat correct?
>
>
>
> On Fri, Jan 26, 2018 at 7:
Berthold Reinwald
> > IBM Almaden Research Center
> > office: (408) 927 2208; T/L: 457 2208
> > e-mail: reinw...@us.ibm.com
> >
> >
> >
> > From: Matthias Boehm <mboe...@gmail.com>
> > To: dev@systemml.apache.org
> > Date: 02/
Hi Janardhan,
sure - adding such bitwise operations is a nice addition.
There is still an open task (SYSTEMML-1931) to generalize the existing NOT,
AND, OR, and XOR to matrix arguments which should be straightforward as is
seamlessly fits into the existing binary operator. In a separate task, we
well, let's do the following to figure this out:
1) If the schema is indeed [label: Integer, features: SparseVector],
please change the third line to val y = input_data.select("label").
2) For debugging, I would recommend to use a simple script like
"print(sum(X));" and try converting X and
gt; Subject:Re: Release Planning SystemML 1.2
>>
>>
>>
>> One more thing, I am already building a docker image. But, which image do
>> you prefer
>>
>> 1. CentOS 7 or
>> 2. Ubuntu - Later extensible to GPU very easily.
>>
>> This decrea
+1
I ran the perftest suite multiple times up to 80GB with and without
codegen. After fixing all the issues and regressions, the entire suite
ran successfully against Spark 2.2 and 2.3 and all use cases showed
equal or better performance compared to SystemML 1.1.
Regards,
Matthias
On Fri, Aug
Could someone with access to our logos please commit them into the
mentioned repo? @Deron: I remember you gave me once the archive with
all versions of our logos. Thanks.
Regards,
Matthias
-- Forwarded message --
From: Daniel Gruno
Date: Thu, Aug 23, 2018 at 1:52 PM
Subject:
can discuss here. If I don't
here back in 3 days, I'll recommend to close the issue at
google/oss-fuzz.
Regards,
Matthias
On Mon, May 21, 2018 at 5:29 PM, Matthias Boehm wrote:
> Well, in general this can be interesting. Apart from our default
> testsuite, we occasionally ran static code an
The Project Management Committee (PMC) for Apache SystemML has invited
Guobao Li to become a committer and we are pleased to announce that he
has accepted.
Guobao was instrumental in creating language and runtime support for
parameter servers in SystemML. This includes local and distributed
-- Forwarded message --
From: Matthias Boehm <mboe...@gmail.com>
Date: Tue, Mar 13, 2018 at 1:00 PM
Subject: Re: Extending Codegen algorithm tests for heuristics
To: Chamath Abeysinghe <abeysinghecham...@gmail.com>
without debugging it's hard to tell, but usually so
Hi Janardhan,
in general, we prefer to compare against R because it helps detecting
issues that are common across different optimizers and execution modes. So
for small scripts like PCA, I would recommend to simply create an R script,
which should be very similar to the dml script.
However, for
Hi Chamath,
ad 1: Yes, this is absolutely correct. However, it is important to realize
that within the workers, we want to run dml functions, and for these we'll
reuse our existing compiler, runtime, operations, and data structures.
ad 2: Yes, this is also correct. Indeed we can use an existing
Pu <qifan...@gmail.com> wrote:
> Matthias,
>
> Thanks so much for taking time to fix. Really appreciated it.
> Does the same reasoning apply to the cholesky script? The recursive approach
> also looks inherently sequential.
>
> Best,
> Qifan
>
> On Sat, Apr 21, 20
ach operator will generate a Spark task converting things
> into RDD operators.
> Thanks so much for the patience and detailed instructions. I have a much
> better understanding of the system now.
>
> On Sun, Apr 22, 2018 at 7:47 PM, Matthias Boehm <mboe...@gmail.com> wrote:
>>
>&
Well, SYSTEMML-2084 aims to integrate the new paramserv builtin function -
as described in the JIRA - into SystemML's semantic validation and
compiler. This would entail to first finalize the design of this builtin
function (e.g., function signature and semantics) and then integrate it
into the
thanks for the initial draft and extensions - I would remove internals
#2/#3 because they are still open, move the other internals to performance,
and include (or extend) the following:
* codegen extensions (operation support, extended optimizer, see
SYSTEMML-2065)
* new accumulator operator +=
-- Forwarded message --
From: Matthias Boehm <mboe...@gmail.com>
Date: Wed, Mar 28, 2018 at 9:34 PM
Subject: Re: Contribution to SystemML
To: Govinda Malavipathirana <mp.govi...@gmail.com>
well, first of all sorry that you wasted one of your proposals because it
wou
+1, I reran all tests mentioned before for RC2 as well without any issues.
Regards,
Matthias
On Fri, Mar 23, 2018 at 5:26 PM, Berthold Reinwald
wrote:
> Please vote on releasing the following candidate as Apache SystemML
> version 1.1.0
>
> The vote is open for at least 72
> select one?
> regards,
> Govinda
>
>
> On Thu, Mar 29, 2018 at 10:06 AM Matthias Boehm <mboe...@gmail.com> wrote:
>
>> -- Forwarded message --
>> From: Matthias Boehm <mboe...@gmail.com>
>> Date: Wed, Mar 28, 2018 at 9:34 PM
>>
Thanks for your interest Daiki. I created two JIRAs SYSTEMML-2233 and
SYSTEMML-2232 that might me a good starting point. I would recommend
to begin with 2233 as a basic cleanup task, which is meant to get you
comfortable. The other task is then a bit more involved but would
improve our function
-- Forwarded message --
From: Matthias Boehm <mboe...@gmail.com>
Date: Sat, Mar 17, 2018 at 5:41 PM
Subject: Re: Sub projects in Language and run time for parameter servers
[SYSTEMML-2083]
To: Chamath Abeysinghe <abeysinghecham...@gmail.com>
great to see that y
-1, sorry but I have to change my vote due to SYSTEMML-2201. The change
will be in master tomorrow - once this is done, we can cut another RC.
Regards,
Matthias
On Wed, Mar 21, 2018 at 4:12 PM, Matthias Boehm <mboe...@gmail.com> wrote:
> +1
>
> I ran the perftest s
awesome - thanks for sharing Krishna.
Regards,
Matthias
On Wed, Mar 21, 2018 at 2:10 AM, Krishna Kalyan
wrote:
> Hello All,
> Sharing a small Shiny App to visualize the runtime performance for System
> ML algorithms. This is work still in progress. Any feedback would
yes, you're absolutely right - in this form, results would always
differ. Even if we feed a seed to both R and DML scripts, the
implementation of our rand is very different as we need to ensure that,
given a seed, we generate the same data in local and distributed
operations. Accordingly, we
Raising the minimum version to 2.7.x is a good idea and fine by me. I
would suggest updating the pom to version 2.7.7 but only mention 2.7 as
a minimum requirement in the README because as far as I know no APIs
changed that wouldn't allow running SystemML on older versions as well.
Regards,
great - thank you so much for the summary. It would be awesome to get it
for a longer history as well and aggregate it with the release/maven
downloads.
Regards,
Matthias
On 10/08/2019 16:23, Janardhan wrote:
Hi,
The following are the queried SystemML download statistics to understand
the
Hi Remy,
generally these scripts are not much different from the ones found in
scripts/algorithms. However, the staging scripts were either in
development, or did not receive enough testing to move to algorithms yet.
If you have a specific algorithm in mind, let us know and we help
,
Janardhan
On Tue, Mar 24, 2020 at 6:28 PM Matthias Boehm wrote:
that's a good point Henry. Yes, with SystemDS 0.1.0, we removed the
MapReduce compiler and runtime backend, the pydml parser and language
support, the Java-UDF framework, and the script-level debugger. We are
concentrating on local
The Project Management Committee (PMC) for Apache SystemML (SystemDS)
has invited Arnab Phani, Mark Dokter, Shafaq Siddiqi, and Kevin
Innerebner to become committers, and we are pleased to announce that all
four have accepted.
The new committers cover many important areas of current and
yes this is the classic blocking issue - we only held back because you
commented you want to review it in detail. We'll take care of it now.
Regards,
Matthias
On 5/12/2020 8:05 AM, Janardhan wrote:
Hi,
@lukas-jkl have implemented ONNX support for SystemDS, also well
documented (in code and
thanks for following-up on this. @Berthold: do I remember correctly that
you looked into a similar setup a while ago?
Regards,
Matthias
On 5/15/2020 7:46 PM, Baunsgaard, Sebastian wrote:
Hi SystemDS developers.
Currently we have some docker containers that are associated with my personal
of the very different objectives and because SystemDS reflects
both the origin from SystemML and its new focus on data science pipelines.
[1]
https://issues.apache.org/jira/projects/PODLINGNAMESEARCH/issues/PODLINGNAMESEARCH-179?filter=allissues
Regards,
Matthias
On 3/9/2020 6:37 PM, Matthias
SystemDS being merged to SystemML repository?
- Henry
On Sat, Mar 21, 2020 at 2:47 PM Matthias Boehm wrote:
just FYI, we created a ticket for the suitable name search, and shared
the related results [1]. So from my perspective, it really boils down to
the question if we accept the closeness
thanks for the initiative and moving this discussion to the dev list. I
think this would be very valuable. Regarding how to start, I would
recommend to focus on external functionality first by documenting all
dml-bodied and native builtin functions (see
org.apache.sysds.common.Builtins) and
Janardhan, thanks for the initiative but please refrain from sending
such advertisements to our dev mailing list and making any promises in
the name of SystemDS/SystemML on potential inclusion of such
dependencies into future releases.
Right now we did not yet decide what our target
Thanks for the question and the detailed inputs - this is an effect of
simplification rewrites that only apply in one of the cases.
Specifically, (t(N)%*%t(M))[1,1] is rewritten to t(N)[1,] %*% t(M)[,1],
which is a form of selection pushdown. You can do the following, for
benchmarking*:
trying to
merge something else. That said nothing bad happened except you did not get a
chance to look through the PR, and it is progress in the direction discussed in
this thread.
Best regards
Sebastian
From: Matthias Boehm
Sent: Saturday, June 6, 2020 7:03:19
from my perspective it would be very important to have all builtin
functions in a single markdown file to allow users to search for things
and users don't care how a builtin function is internally implemented.
So we might want to use this opportunity to consolidate the already
documented
95 matches
Mail list logo