That Job extends org.apache.mahout.common.AbstractJob, so it probably
will accept a -D argument to set "mapred.fairscheduler.pool=..." . Have
you tried this?
On 11/8/12 3:41 PM, Yazan Boshmaf wrote:
Hello,
I'm trying to run the ASF Email example here:
https://cwiki.apache.org/confluence/disp
See the response "Re: Clustering without hadoop" by Johannes Schulte two
postings earlier than yours on user@m.a.o. The driver functions can also
be run in sequential mode from a local file system and do not require
Hadoop. There are several examples of Java invocation in the unit tests.
TestCl
Canopy is very sensitive to the value of T2: Too small a value will
cause the creation of very many canopies in each mapper and these will
swamp the reducer. I suggest you begin with T1=T2=
until you get enough canopies. With CosineDistanceMeasure, a value of 1
ought to produce only a single
implementation of the code or i am
setting the way too off-set value for parameters?
is there any more info that i could provide to help you to help me analyze
the issue?
if i set t3,t4, would it help?
thanks
On Tue, Nov 13, 2012 at 10:01 PM, Jeff Eastman
wrote:
Canopy is very sensitive to the value
The classification phase of Dirichlet uses a most-likely assignment of
points to clusters by default. This means that, unlike the training
phase where points are assigned statistically to likely clusters, the
classification may result in empty clusters even though those clusters
have nonzero co
Hi Yunming,
The problem I see with what you are proposing is that Hadoop only gives
you a single input vector per call of CIMapper.map(). Using multiple
threads to perform the body of that method would not be of benefit. If
you want to experiment with thread-based concurrent execution of that
ism within a mapper,
Thanks again everyone!
Yunming
On Sat, Dec 22, 2012 at 12:48 AM, Jeff Eastman
wrote:
Hi Yunming,
The problem I see with what you are proposing is that Hadoop only gives
you a single input vector per call of CIMapper.map(). Using multiple
threads to perform the body of t
Here's a response to a similar question from a couple of months ago:
The classification phase of Dirichlet uses a most-likely assignment of
points to clusters by default. This means that, unlike the training
phase where points are assigned statistically to likely clusters, the
classification m
on.
I may be wrong but this is bug?
Thanks,
Yoshihiro.
2012/12/26 Jeff Eastman
Here's a response to a similar question from a couple of months ago:
The classification phase of Dirichlet uses a most-likely assignment of
points to clusters by default. This means that, unlike the trai
Depending upon your data, 0.7 Canopy can be extremely sensitive to the
value you specify for T2. Somewhere between the larger T2 value that
yields 1 canopy and the smaller T2 value that yields "the wrong number
of [i.e. too many] centroids" lies a value that will give you fewer
centroids. You c
To unsubscribe from this list, send an email to
user-unsubscr...@mahout.apache.org
On 1/10/13 11:14 AM, Walshe, Maurice (RBI-UK) wrote:
unsubscribe
-Original Message-
From: akshay shetye [mailto:akshay.she...@gmail.com]
Sent: 10 January 2013 14:40
To: user@mahout.apache.org
Subject: Re
I know of no reliable ways to avoid some iteration in setting the T
values for Canopy but T1 really has no impact on the number of clusters
so setting T1==T2 and experimenting with T2 will reduce your search space.
On 2/1/13 6:29 AM, Chris Harrington wrote:
Seems my lack of any clusters what
If you don't specify a -k value but specify a -ci directory that
contains clusters you want to use for the prior then the ClusterIterator
will use them for kmeans and fuzzyk. You will need to create one or more
sequence files containing ClusterWritables to do this.
On 2/1/13 9:08 AM, sri krish
) to Cluster type ?
From: Jeff Eastman
To: user@mahout.apache.org
Sent: Saturday, 2 February 2013 12:18 AM
Subject: Re: intial centriods for fuzzy k means algorithm
If you don't specify a -k value but specify a -ci directory that
contains clusters you
That's a really good question. Mahout does not have an "explain"
feature; however, you can use the ClusterDumper to print out the cluster
centers and vectors clustered within each cluster. Output is pretty
verbose and, with large text vectors being truncated, might not be that
useful. You might
Kinda looks like you didn't specify the right input file. That job
expects the delimited values from the synthetic control download,
converts them to vectors and clusters them. The vectors are of
cardinality 60 but somehow your input data generated 1151 elements. I'd
look there.
On 2/4/13 2:
Maybe a typo? I would expect folder 16 to follow folder 15. For many
reasons though, the cluster numbers may not be monotonic. Suggest you
just iterate over the directories that are presented, their names should
correspond to the clusterIds that exist in you clusters-final directory.
On 2/4/13
Note that the old clustering algorithms also run without Hadoop in
sequential execution mode from the local file system.
On 2/6/13 11:04 AM, Tanguy tlrx wrote:
Thanks!
-- Tanguy
2013/2/6 Ted Dunning
https://github.com/tdunning/knn/
especially the docs directory
On Wed, Feb 6, 2013 at 7:5
ingle process in memory.
On 2/7/13 4:32 AM, vivek bairathi wrote:
Hi Jeff,
Can you name some?
On Thu, Feb 7, 2013 at 12:33 AM, Jeff Eastman wrote:
Note that the old clustering algorithms also run without Hadoop in
sequential execution mode from the local file system.
On 2/6/13 11:04 AM, Tanguy
What kind of data are you clustering?
Which model distribution are you using?
How many iterations are you running?
How do the cluster n= values change as you increase the number of
iterations?
On 2/7/13 11:35 AM, Aysu Ezen wrote:
Hello,
I am having difficulty with Dirichlet process clusteri
You also need to specify a fully-qualified class name
On 2/12/13 11:48 AM, Dan Filimon wrote:
You need to add the JAR containing the distance measure you want to
the classpath.
By default the CLASSPATH is set in line 120 of the mahout script. (the
script itself is in the bin/ folder of your Maho
This is a common question and you can search the email archives for more
discussion. Start by setting t1 == t2 as t2 is the variable that
controls the number of clusters produced. Then iterate to find a value
that gives you the "right" number of clusters. Smaller values of t2
yield larger numbe
I don't know where the timeout is happening, but each mapper and each
reducer writes all its clusters out at the end of its run. With a large
number of clusters, and with the non-sparse center and radius vectors
that tend to accumulate, this could take a while...
On 3/8/13 9:46 AM, Colum Foley
Unfortunately, all attachments are stripped by the Apache mail server.
You will need to open a JIRA to get those attachments to us. You could;
however, also tell us a bit more about your example: which algorithm are
you running and what is your command line? I cannot think of any way
that the 0
On 3/22/13 10:39 AM, Sebastian Briesemeister wrote:
Dear all,
I am facing troubles when retrieving the cluster probabilities of instances:
I am clustering instances using the FuzzyKMeansDriver.
Afterwards, I am reading instances of WeightedVectorWritable from the
clusteredPoints file (e.g. part
Fuzzy KMeans will use a lot of heap memory because every vector is
observed (with weighting) by every cluster. This will make the cluster
centers (and other vectors) much more dense than with any of the other
clustering algorithms. Figure you are storing 90k doubles in each vector
and each clus
Hi Erinn,
The radius calculation in KMeans and other clustering algorithms uses a
running sums algorithm (see RunningSumsGaussianAccumulator) and the
radius is really the standard deviation produced by this method. In this
method (as you likely know) s0 is the number of points observed, s1 is
classification results.
*From:*Jeff Eastman [mailto:jeast...@windwardsolutions.com]
*Sent:* Tuesday, May 14, 2013 11:10 AM
*To:* Erinn Schorsch
*Cc:* user@mahout.apache.org
*Subject:* Re: Problems with KMeans Clustering - Radius calculation
returns incorrect ZERO value in some cases.
Hi Erinn,
The
Another option would be to add a new command line option to the
ClusterDumper to produce the abbreviated output you desire. Then you
could submit it as a patch and everybody could benefit. Off hand, this
seems like a useful output representation.
Jeff
On 5/24/13 6:57 AM, Rajesh Nikam wrote:
The clustering arguments are usually directories, not files. Try:
mahout clusterdump -d mahout/vectors/dictionary.file-0 -dt sequencefile -i
mahout/kmeans-clusters/clusters-1-final -n 20 -b 100 -o cdump.txt -p
mahout/kmeans-clusters/clusteredPoints
On 8/1/13 2:51 PM, Marco wrote:
mahout
Hi Pablo,
Look in the CDBw unit tests for examples of invoking it from Java code.
Jeff
On Sep 6, 2013, at 5:56 PM, Pablo Andretta Jaskowiak
wrote:
> Hello,
>
> I'm trying to use the CDBw implementation from Mahout. Given that I
> have a dataset in CSV format and a clustering solution how ca
I've not tried this particular use case personally but it should be possible.
NamedVectors flow transparently through k-means and Dirichlet, indeed they
should be transparent in all Vector clustering. When you run the final
clustering (-cl) step with Dirichlet, this will use the final iteration'
I can't think of a good reason not to do this, but I've not been in that code
for a while. If you make that change, do all of the unit tests still run? Can
you write another test that uses DenseVectors? If yes, then please submit a
JIRA patch.
Jeff
On Nov 3, 2013, at 1:30 PM, "DeBarr, Dave"
, August 31, 2011 12:16 PM
To: Jeff Eastman
Cc: user@mahout.apache.org
Subject: Re: Problems running examples
On 10 June 2011 18:34, Jeff Eastman wrote:
> I'm still trying to figure out why reuters-0.5 does not work on either of my
> clusters. The scripts themselves have no diff and the
You can use -Dmapred.reduce.tasks=n to set the number of reducers for most
Mahout CLI jobs. Just be sure it is the first argument.
-Original Message-
From: myn [mailto:m...@163.com]
Sent: Tuesday, September 13, 2011 9:15 AM
To: user@mahout.apache.org
Subject: why so many place does`t set
Looks like your input data is not numeric:
"2Yu�3.1_0�osLinux". The InputMapper is barfing
trying to convert this into a double.
-Original Message-
From: Varun Thacker [mailto:varunthacker1...@gmail.com]
Sent: Tuesday, September 13, 2011 10:09 AM
To: user@
See inline comments
-Original Message-
From: Nivlem Trahm [mailto:nivle...@yahoo.com]
Sent: Friday, September 16, 2011 12:32 PM
To: user@mahout.apache.org
Subject: Making Mahout cluster results more like Cluto's?
Hi,
I evaluated Cluto some time ago, and the results I was getting from
Actually, most of the clustering jobs (including DirichletDriver) accept the
-Dmapred.reduce.tasks=n argument as noted below. Canopy is the only job which
forces n=1 and this is so the reducer will see all of the mapper outputs.
Generally, by adjusting T2 & T1 to suitably-large values you can ge
write canopies in
a filesystem. This is done as a mapreduce job. Then the KMeansDriver
needs these canopy points as input to run KMeans.
On 20-09-2011 01:39, Jeff Eastman wrote:
> Actually, most of the clustering jobs (including DirichletDriver) accept the
> -Dmapred.reduce.tasks=n ar
nique would you suggest to cluster really big
data ( considering performance and big size as parameters )?
Thanks and Regards,
Paritosh Ranjan
On 20-09-2011 21:35, Jeff Eastman wrote:
> Well, while it is true that the CanopyDriver writes all its canopies to the
> file system, they are written at
g single vector ).
On 20-09-2011 22:56, Jeff Eastman wrote:
> I guess it depends upon what you expect from your HUGE data set: How many
> clusters do you believe it contains? A hundred? A thousand? A million? A
> billion? With the right T-values I believe Canopy can handle the first t
ersion of clustering to a "persisted"
one. The current implementation is not scalable. I have a valid business
scenario with 5 million clusters, and I think there would be more users
with bigger datasets/cluster numbers.
Thanks and Regards,
Paritosh Ranjan
On 20-09-2011 23:35, Jeff Eastman wr
would be more users with bigger
> datasets/cluster numbers.
>
>
> Thanks and Regards,
> Paritosh Ranjan
>
> On 20-09-2011 23:35, Jeff Eastman wrote:
>
>> As all the Mahout clustering implementations keep their clusters in
>> memory, I don't believe any of t
This is a common problem with canopy, since it is single-pass and uses a single
reducer that must see the outputs of all mappers. You can adjust T2 upward and
that will reduce the number of canopies produced by each mapper. T1 does not
affect the number of canopies, only their centroid calculati
>>> I think, two improvements, can be applied to the current algorithm.
>>>
>>> 1) To ask for minimum number of vectors to be inside a
>>> canopy/cluster, or
>>> the cluster is discarded.
>>> 2) To change this "in memory" version of clus
The spectral clustering uses similarity matrices. We have no hierarchical
implementation.
-Original Message-
From: prasenjit mukherjee [mailto:prasen@gmail.com]
Sent: Thursday, September 29, 2011 10:44 AM
To: user@mahout.apache.org
Subject: Clustering based on Similarity matrix
I ha
The sequential and mapreduce implementations do not produce the same results,
as the sequential implementation runs canopy once and the mapreduce
implementation twice: in each mapper and in the reducer. This is documented in
https://cwiki.apache.org/confluence/display/MAHOUT/Canopy+Clustering (s
Well, the default clusterFilter == 0, so this is not the difference between the
implementations. When you talk about distributing similar vectors to each
mapper, you are really moving into a hierarchical clustering method where you
cluster your input points into a few large clusters and then clu
Check out TestClusterDumper.testDirichlet2&3 for an example of text clustering
using DPC. It produces reasonable looking clusters when compared with k-means
and the other algorithms, but on a small vocabulary. Also check out
DisplayDirichlet, which does a great job of clustering some random 2-d
ative values are coming from.
This deserves further exploration, which I am doing...
-Original Message-
From: Jeff Eastman [mailto:jeast...@narus.com]
Sent: Tuesday, October 18, 2011 9:24 AM
To: user@mahout.apache.org
Subject: RE: Dirichlet Process Clustering
I agree something is amiss here, but it could be the model is just not suitable
for this problem. Running with the Reuters dataset, I see all the points being
assigned to C-0 in the very first iteration as you do. I think the problem is
with the pdf() calculations in the mapper for very wide vec
tQuick(i),
getCenter().getQuick(i), getRadius().getQuick(i) + 0.01);
}
return pdf;
}
-Original Message-
From: Jeff Eastman [mailto:jeast...@narus.com]
Sent: Wednesday, October 19, 2011 9:04 AM
To: user@mahout.apache.org
Subject: RE: Dirichlet Process Clustering not wo
-mp org.apache.mahout.math.DenseVector \
-dm org.apache.mahout.common.distance.CosineDistanceMeasure
-Original Message-
From: Jeff Eastman [mailto:jeast...@narus.com]
Sent: Wednesday, October 19, 2011 9:53 AM
To: user@mahout.apache.org
Subject: RE: Dirichlet Process Clustering not working
Th
ave played around with Reuters set.
Ed
ps. The runtime has indeed reduced significantly!!! Possibly 100 times
faster as you said. Loved it!!
2011/10/20 Jeff Eastman
> R1186452 commits two small changes that seem to do much better with Reuters
> than before:
> - fixed Dista
Unfortunately, the cluster dumper loads all the points into memory so it can
sort them by cluster for display. What are you trying to do with the 20M
points? Certainly not display them! A better step for subsequent processing
would be for you to write a short MR program to read in the clusteredP
+1 from me too. IIRC this all got added when we were annotating Vectors too and
there we ended up with NamedVector as a wrapper. If this Matrix annotation is
not being used then let's clean it up.
-Original Message-
From: Ted Dunning [mailto:ted.dunn...@gmail.com]
Sent: Wednesday, Novem
See inline,
Jeff
-Original Message-
From: gaurav redkar [mailto:gauravred...@gmail.com]
Sent: Wednesday, November 09, 2011 4:09 AM
To: user@mahout.apache.org
Subject: meanshift clustering
Hi.. I am unable to identify where is the clusterPoints() function in the
MeanShiftCanopyClusterer.j
Check out AbstractCluster.formatVector(). If the vector is sparse or if
bindings are present it uses the index:value notation, else it uses the more
compact notation.
-Original Message-
From: Grant Ingersoll [mailto:gsing...@apache.org]
Sent: Friday, November 11, 2011 7:22 AM
To: user@m
rt-m-* correct..??
On Wed, Nov 9, 2011 at 11:27 PM, Jeff Eastman wrote:
> See inline,
> Jeff
>
> -Original Message-
> From: gaurav redkar [mailto:gauravred...@gmail.com]
> Sent: Wednesday, November 09, 2011 4:09 AM
> To: user@mahout.apache.org
> Subject: me
Sorry for the delay in responding. By now you may have already figured this
out. If not:
1. Did you specify the -cl option on Dirichlet to emit the clusteredPoints
directory? The default is not to do so.
2. Did you specify the -p option on ClusterDumper to use that directory?
3. Which model are
Usually we see this error when the expected input vectors are not present.
Often it is a configuration issue. Verify your paths exist and that there are
vectors where you expect them.
-Original Message-
From: Ahmad Ammari [mailto:ammari...@gmail.com]
Sent: Wednesday, November 16, 2011 1
K-means is attempting to load your initial clusters and is not finding any.
Have you checked your -c path? You can also add -xm sequential so you can run
the sequential algorithm. This allows you to use a debugger to verify your
paths.
-Original Message-
From: Ahmad Ammari [mailto:ammar
How did you set the heap sizes? If you are running on a cluster you need to add
properties to your mapred-site.xml. Something like this:
mapred.map.child.java.opts
-Xmx1500m
Java opts for the map tasks.
MapR:
Default heapsize(-Xmx) is determined by memory reserved for mapreduc
Hi to all of you who may have been wondering what ever happened to Jeff.
I've unpacked most of my boxes in Northglenn, CO and am now, mostly,
back on line. Of course, there have been over a thousand Mahout postings
during my two week absence, so it may take me a while to wade through
all of the
Could you elaborate a bit on what you mean by "cluster a collection of
words and phrases by syntactic similarity over a distributed environment
"? If you can describe your collection in terms of a set of (sparse or
dense) term vectors then you should be able to use Mahout clustering
directly. T
I usually run it from Eclipse too, but
./bin/mahout org.apache.mahout.clustering.display.DisplayKMeans
just ran fine for me on trunk
On 12/2/11 3:47 PM, Grant Ingersoll wrote:
You should just be able to run it. It doesn't take in any input. "sample" and
"output" are just the names of the d
You also want to run locally.
unset HADOOP_HOME
unset HADOOP_CONF_DIR
On 12/3/11 11:03 AM, Jeff Eastman wrote:
I usually run it from Eclipse too, but
./bin/mahout org.apache.mahout.clustering.display.DisplayKMeans
just ran fine for me on trunk
On 12/2/11 3:47 PM, Grant Ingersoll wrote
You will need to wrap your input vectors in a NamedVector, using your
document ids as the names. These will pass through the clustering
process and you will be able to map each clustered vector back to your
input that way.
On 12/5/11 5:02 PM, Neil Chaudhuri wrote:
I am attempting to programma
The KMeansDriver has a method (clusterData) which you can invoke from a
Java program to cluster (classify) your new data with the old clusters.
You need to be sure the vectors are the same size (and the elements
denote the same attributes) for this to work. There is currently no CLI
to invoke t
Mahout in general uses sequence files for input and output. These are
binary encoded files that can only be read by a compatible program. If
you are trying to e.g. less .../part-xxx then you won't see much that is
human readable. You can run the ClusterDumper to get human readable
output from r
I'm almost certain there is no current way to do this from the command
line. You could write a small utility to do this (see
CanopyClusterer.buildClustersSeq() for a simple skeleton you could use).
But I would suggest trying CosineDistanceMeasure instead of Euclidean
for text. If you have a sma
The Synthetic Control examples use a similar (but space delimited) input
format and there is an InputDriver in integration/ which can convert
those files into Mahout Vector sequence files. You could easily modify
the InputMapper to be comma delimited or modify your own file formats to
use space
Even better, you might figure out how to pass the desired delimiter into
the InputDriver as an argument and submit a patch to make that a
permanent Mahout feature. It should be straightforward and it would
start you down the path to become a committer.
On 1/9/12 2:52 PM, Jeff Eastman wrote
It kinda looks to me like you have some inconsistent data in your input
data set. Here's the sequence of operations that occur in Synthetic Control:
1. The InputDriver reads your input text files and produces sequence
files of VectorWritable. Each vector is produced by the InputMapper
after pr
What you have read is correct. Mahout clustering (unsupervised
classification) can only deal with continuous, homogeneous vector
representations of the input data, where each vector element is weighted
the same as the other elements. Mahout (supervised) classification can
deal with continuous,
Just remember that Longitude is a spherical coordinate and +179 is
closer to -179 than their numeric difference. Latitude is spherical too
but +89 is indeed quite far from -89.
On 1/13/12 4:36 AM, StreetCat wrote:
The raw data had location expressed as strings such as "Paris, France" and
I tr
That's probably because you are not performing the clustering (vector
classification) step. The clusterer has a method
(emitPointToNearestCluster) which supports that to files, but you will
have to write your own method to do it all in memory. Suggest you look
at the driver's sequential cluster
Mean Shift accumulates the pointIds of every point assigned to a
cluster, so I would expect n= to be correct in the cluster dumper
output. It is most likely the postprocessor is misbehaving. Please
create a JIRA and attach your dataset and we will take a look at it.
It would also be useful for
+1 Congratulations to Shannon for a job well done. We now have a 0.6
release and can begin to concentrate on the plan and issues for a 0.7
release.
On 2/6/12 2:19 PM, Shannon Quinn wrote:
Apache Mahout has reached version 0.6. All developers are encouraged
to begin using version 0.6, as much h
2012 at 8:01 PM, Jeff Eastman
wrote:
Now that 0.6 is in the box, it seems a good time to start thinking about
0.7, from a high level goal perspective at least. Here are a couple that
come to mind:
Target code freeze date August 1, 2012
Get Jenkins working for us again
Complete clustering
dencies on Mahout and on an analyzer to
test things out.
Another thing would be adding or improving the integration tools. For
example adding a mysql2seq to cluster text from a SQL database.
On Sat, Feb 11, 2012 at 8:01 PM, Jeff Eastman
wrote:
Now that 0.6 is in the box, it seems a good time to sta
We have a couple JIRAs that relate here: We want to factor all the (-cl)
classification steps out of all of the driver classes (MAHOUT-930) and
into a separate job to remove duplicated code; MAHOUT-931 is to add a
pluggable outlier removal capability to this job; and MAHOUT-933 is
aimed at fact
+users@
Just to be clear, I'm not advocating replacing the JIRA process with a
new set of green-field goals. Rather, IMHO, having a small number of
overarching goals for a release *could* help us focus our efforts
(triage our feature JIRAs) and *might* suggest some missing JIRAs that
would gi
This is correct, and is actually what the documentation says though it
may not be completely clear. The reason for this is somewhat historical:
K-Means originally did not have a -k argument and required the user to
provide the prior cluster centers in the -c argument, using canopy for
example. Then
For human-readable output, yes.
On 2/17/12 6:09 AM, Tharindu Mathew wrote:
Or I can just use the cluster dump tool right...?
On Fri, Feb 17, 2012 at 5:55 PM, Paritosh Ranjan wrote:
Try logging in and updating.
Thanks...
On 17-02-2012 17:54, Tharindu Mathew wrote:
OffTopic: How would I con
of primitives for designing new serial and distributed machine
learning algorithms. And I think it has a high utility for integration
into highly visible commercial projects. But its high level public API
really is a barrier to entry when trying to design commercial
applications.
On Sun, Feb 12,
Sure, just look at Hadoop. But I'm not hung up on the port/starboard
numbering scheme either. Either 0.6.1 or 0.7 work for me.
On 2/22/12 11:46 AM, Jake Mannix wrote:
On Wed, Feb 22, 2012 at 10:40 AM, Dmitriy Lyubimovwrote:
I guess i still prefer 0.6.1 for maintenance releases (esp. given the
Grin. I broke it myself and I'm on it.
Jeff
On 2/22/12 5:45 PM, Lance Norskog wrote:
Please fix the Jenkins build.
You can cause a canopy to un-observe a vector by observing it with a -1
weight. This will have the effect of subtracting all influence of the
vector on observations of that canopy. But you can only do that before
you computeParameters, as all observation semantics is reset by that
operation and
+1 exactly correct Paritosh. When I wrote that I was using a physical
mental model of, perhaps, accretion of mass in a swarm of asteroids or
in deep space where stars accrete from irregularities in the
interstellar dust cloud. I've never been 100% sure this is truely
MeanShift, but it seems to
What's your Hadoop config in terms of the maximum number of reducers?
It's a function of your available RAM on each node and numbers of nodes.
On 3/10/12 8:55 PM, WangRamon wrote:
> Hi ParitoshI did the tests with 1 job and 5 jobs, they all have the same
> problem, the job i'm running is the
EuclideanDistance is not a great choice for document clustering,
especially with a lot of terms. Suggest you try CosineDistance which
will give you all distances between 0 and 1. If you still end up with
only one canopy it is because T2 is too large. T1 has no effect upon the
number of canopies
008. Every case, the
reducer quickly passed 67%, then very very slowly progress, for
example, it takes several minutes to finish 1% more.
Is that something wrong in my data?
Best
Baoqiang
On Tue, Mar 13, 2012 at 3:08 PM, Jeff Eastman
wrote:
EuclideanDistance is not a great choice for document
is the inevitable solution for my problem. Ironically,
I went to canopy in hope of getting better results out of kmeans.
Thanks again.
Baoqiang
On Tue, Mar 13, 2012 at 5:01 PM, Jeff Eastman
wrote:
No, Canopy only uses a single reducer, so what's happening is many mappers
are munching your
With Canopy this is a symptom of T2 being too large. This causes an
explosion of clusters - in the limit, one per input vector - and if
vector dimension is large too there is no amount of memory which can
hold them all for large datasets. Reduce T2 until you get a tractable
number of clusters, then
Good question. The only way I can think of an infinity in a Kluster
center is if there were some infinity values in the vectors it observed.
The center (centroid) is calculated in each iteration after all points
have been observed by dividing S1 by S0. If, for some reason, S0 was
zero this would ca
Empty? Note that the printouts of Mahout vectors prints only the
non-zero elements. It looks like you may have had many such zero vectors
and they were clustered into VL-1705919 which has zero for center and
radius. If your other clusters look differently, then I think this is
probably correct.
Most of the Mahout clustering algorithms have an -xm sequential CLI
option that runs locally in-memory from/to Hadoop-style sequence files.
And, as below, you can also call the Java driver methods directly from
your program.
On 3/22/12 9:22 AM, Ahmed Abdeen Hamed wrote:
Hi,
I think I can ans
Hi Oscar,
It would help a lot if you could provide a bit more information on the
data that you wish to cluster, particularly the dimensionality of each
record and the number of records. Also please note that Mahout's k-means
implementation runs in a batch mode on Hadoop so integrating this wit
And, finally, if you want to embed Mahout clustering in your web
application and the data all fits into memory or you can devise an
Iterator to feed data into it, the ClusterIterator.iterate(...)
method will do k-means (also fuzzyK and Dirichlet) clustering in memory.
On 4/16/12 11:34 AM, Manu
1 - 100 of 393 matches
Mail list logo