Does anyone know what's the mechanism behind hyperparameter tuning job in AWS Sagemaker?
In specific, I am trying to do the following:
Bring my own container
Minimize cross entropy loss (this will be the objective metric of the tuner)
My question is when we define the hyper parameter in HyperParameterTuner class, does that get copied into /opt/ml/input/config/hyperparameters.json?
If so, should one adjust the training image so that it uses the hyper parameters from /opt/ml/input/config/hyperparameters.json?
Edit: I've looked into some sample HPO notebooks that AWS provides and they seem to confuse me more. Sometimes they'd use argparser to pass in the HPs. How is that passed into the training code?
So i finally figured it out and had it wrong all the time.
The file /opt/ml/input/config/hyperparameters.json is there. It just has slightly different content compared to a regular training-job. The params to be tuned as well as static params are contained there. As well as the metric-name.
So here is the structure, i hope it helps:
{
'_tuning_objective_metric': 'your-metric',
'dynamic-param1': '0.3',
'dynamic-param2': '1',
'static-param1': 'some-value',
'static-paramN': 'another-value'
}
If you bring your own container, you should consider pip installing SageMaker Training Toolkit. This will allow you to receive hyperparameters as command line arguments to your training script (to be processed with argparser). This will save you the need to Read and parse the /opt/ml/input/config/hyperparameters.json yourself.
Related
Currently i am using the REST interface to query vespa, which seems to work great but something tells me that i should be using searchers in the application to make the client(server side code) a bit lighter (bundle the jar file in the application package) to make it a bit smoother. I have managed to do some simple searcher/processor applications. But this is a bit overwhelming.
So are there any readily available examples ?
Basicially i want to:
Send to /search?query=someId
Do a ordinary search for the weighted set on this documentID (I guess this one can be handy: https://docs.vespa.ai/documentation/reference/inspecting-structured-data.html)
Take those items in the response and add it to a wand item(s) and query for a wand with wandsearcher on a given field. Similar to the yql:
"select * from sources * where wand(interest, some weightedsets));","ranking":"combined_score" and return the matches.
Just curious also, apart from the trouble of string building with the http request i am doing at the moment are there any performance gains of using a searcher or go the java route vs rest?
thanks for any insight or code help i can start with.
There is an example of using the WandItem (YQL wand)here https://docs.vespa.ai/documentation/advanced-ranking.html and see also https://docs.vespa.ai/documentation/using-wand-with-vespa.html as there are two wand implementations available in Vespa, it sounds from the description that the wand() is what you want to use for this use case. For the first call you probably want to have a dedicated document summary to reduce the amount of data fetched for your first query and also the option of serving it out of memory only (See https://docs.vespa.ai/documentation/document-summaries.html)
Also see https://docs.vespa.ai/documentation/searcher-development.html as a general resource on writing searchers.
For your use case it makes a lot of sense to write a searcher to perform these two queries as your second query depends on the first and you avoid the cost of rendering/http/yql parsing which might matter if your client is remote with high network latency.
We have a need for "Blending of hits from different sources", as per your documentation it is recommended to write a custom-searcher in JAVA. Is there a demo of this written somewhere on Github ? I wouldn't even know where to start :( I understand I can create search "chains" , preferably Asynchronous, and then blend results in JAVA before returning them...but then how would I handle paginations, limits...etc ? This all seems very complicated, for someone who doesn't even know JAVA that much. So, I am hoping someone has already written a demo for this ? Please ? Anyone ?
Thank you so much
EDIT to make my quesion clearer:
We are writing a search engine that fetches data from various websites. Some websites have 10mil indexable items, other websites only 100,000. When we present the results to end user, we want to include results from all our sources ( when match applies ). Let's say 10 results from each of the websites we crawl, so that they all get equal amount of attention on page. If we don't do custom blending, what happens is that the largest website with most items wins all our traffic.
I understand that we can send 10 separate queries to VESPA, and blend the results in our front end, but that seems very inefficient. Thus, the quesion of "Custome Searcher". Thank you so much !
That documentation covers some very advanced use cases which you do not have. Are your sources different Vespa schemas or content clusters? If so Vespa will by default blend the hits returned from each according to their relevance scores so there's nothing you need to do.
The two other most common use-cases are:
Some (or all) the data sources are external, so you need to write a Searcher component to fetch the external data and turn it into a Result.
You want the data to be blended in some custom way (rather than by relevance score). If so you need to exclude the default blending Searcher (com.yahoo.prelude.searcher.BlendingSearcher) and write your own.
If you provide some more information about your use cases I can give you some code examples.
EDIT: Use grouping to solve the need explained under "EDIT" in the question:
Create a "siteid" field when feeding (e.g in document processing).
Use the grouping expression all(group(siteid) each(max(10) output(summary())))
See http://docs.vespa.ai/documentation/grouping.html
I am using asterisk 11.9.0 and i want to generate an outgoing call.I found that for outgoing i have to make a .call file and place it in a var/spool/asterisk/outgoing.I am following the link below
http://the-asterisk-book.com/1.6/call-file.html#call-file-parameter
my code is same as given in the above link,the above example uses only single fixed number to call.
My problem is that
i have to generate an outgoing to a number fetched from database(outgoing to new number everytime),so how to write the code of .call file for multiple numbers outgoing and how to pass these numbers fetched from database to .call file from my extensions.conf
Is there any way to do that.
I am new to asterisk.
Any help would be appreciated.
You can use vicidial.org software to do that things.
Note, it is very bad idea do outboudn dialler-like app in asterisk without understanding asterisk logic and very-hi skills in programming/database.
For more info you also can use this page
http://www.voip-info.org/wiki/view/Asterisk+auto-dial+out
Might be easier using WombatDialer as it has a plain API where you can tell it what you want it to do and it will take care of the rest. We have a plain set up for outbound and it took maybe a couple of days from zero to what we have now. ViciDial would have been overkill.
On why rolling your own is not a great idea, the Wombat manual is quite clear: http://manuals.loway.ch/WD_UserManual-chunked/ch01.html#_why_was_wombatdialer_created
You could also use the AMI (Asterisk Manager Interface), would be easier to program with a deamon running in the back to control what gets dialed and the responses to those dials. Mora info here https://wiki.asterisk.org/wiki/pages/viewpage.action?pageId=4817239.
As my question was remained unanswered in Rapidminer forum via this address :
my original question
I would like to propose it here again, please inform me if by any chance you know the answer.
I have noticed no matter what type of extensions we use for getting export (i.e. .pdf, .ps .jpg , etc), the export engine do like a screenshot of the current view instead of the whole output space, this has caused lots of issues for me exporting gigantic decision trees of my research.
Please someone advise me on this issue.
Best,
Amir
Use the log operator at the end of your process and configure it by the information you need. The log operator will save all results in a text file. Based on this file you can generate all your graphs at any time using rapidminer or gnuplot...
I find this way the best to save your time and memory space and also very easy to share in my opinion.
how much time does it take to create a simple dance performing bot in second life using Linden Scripting Language ?? , i have no prior knowledge of lsl , but i know various object oriented and event driven programming languages .
Well, it's pretty simple to animate avatar: you'll need a dance animation (those are pretty easy to find, or you could create your own), put it in a prim (which is basic building object in SL), and then create a simple script which first requests permission to animate desired avatar:
llRequestPermissions(agent_key, PERMISSION_TRIGGER_ANIMATION);
You receive response in run_time_permissions event and trigger your animation by name:
run_time_permissions(integer param)
{
llStartAnimation("my animation");
}
Those are the essentials; you'll probably want to request permissions when an avatar touches your object, and stop animation on second touch... or you could request for permissions for each avatar which is in certain range.
As for the "bot" part, Second Life viewer code is open source; if you don't want to build it yourself, there are several customizable bots available. Or you could simply run an official SL viewer and leave it on; there is a way to run several instances of SL viewer at the same time. It all depends on what exactly you need it for.
Official LSL portal can be found here, though I prefer slightly outdated LSL wiki.
Slight language mismatch: To make an object perform a dance is currently known as "puppetry" in a SecondLife context. The term "bot" currently means control of an avatar by an external script api.
Anyway in one instance, it took me about two hours to write, when I did one for a teddy bear a few weeks back, but that was after learning alot from taking apart some old ones, and i never did finish making the dance, it just waggles the legs or hugs with the arms, but the script is capable for however many steps and parts you can cram in memory.
Puppetry of objects has not improved much in the past decade. It is highly restricted by movement update rates and script limitations. Movement is often delayed under server load and the client doesn't always get updates, which produces pauses and skips in varying measure. Scripts have a maximum size of 64k which should be plenty but in actuality runs out fast with the convolutions necessary in lsl. Moving each linked prim in an object seperately used to need a script in each prim, until new fuctions were introduced to apply attributes by linknumber, still many objects use old scripts which may never be updated. Laggy puppets make for a pitiful show, but most users would not know how to identify a good puppetry script.
The popular way to start making a puppet script is to find an older open source puppet script online, and update it to work from one script. Some archane versions are given as 'master' and 'slave' scripts which need to be merged placing the slave actions as a function into the master, changing llMessageLinked( ) for the function name. Others use an identical script for each prim. I said popular, not the simplest or easiest, and it will never be the best.
Scripting from scratch, the active flow needs to go in a timer event with nothing else in it. Use a different state for animating if a timer is needed when waiting because it's a heavy activity and you don't need any more ifs or buts.
The most crucial task is create a loop to build parameters from a list of linknumbers, positions, and rotations into a parameter list before the llSetLinkPrimitiveParamsFast( ). Yes, that's what they called it because it's a heavy list based function, you may just call it SLPPF inworld but not in the script. Because SLPPF requires the call to have certain constants for each parameter, the parameter list for each step will need to include PRIM_LINK_TARGET, linknumber, PRIM_POS, position, PRIM_ROT, rotation for each linked part in the animation step.
Here's an example of putting a list of a single puppetry step into SLPPF.
list parameters;
integer index;
while ( index < total ) { // total number of elements in each list
parameters += [
PRIM_LINK_TARGET, llList2Integer( currentstep, index ),
PRIM_POS, llList2Vector( currentstep, index+1 ),
PRIM_ROT, llList2Rotation( currentstep, index+2 )
];
index += 3;
}
llSetLinkPrimitiveParamsFast( 0, parameters );
How you create your currentstep list is another matter but the smoothest movement over many linked parts will only be possible if the script isn't moving lists around. So if running the timer at 0.2 seconds isn't any improvement over 0.3, it's because you've told lsl to shovel too many lists. This loop with three list calls may handle about 20 links at 0.1 seconds if the weather's good.
That's actually the worst of it over, just be careful of memory if cramming too many step lists into memory at once. Oh and if your object just vanishes completely, it's going to be hanging around near <0,0,0> because a 1 landed in the PRIM_LINK_TARGET hole.