so other process can use it

This is why there is a model repository, so you can query the last model created, or by name or tag or query the Task that created it and then via the Task the model and it's location.
This is a stable way to make sure your application code (the one using the model) will get to use stable models regardless of the training processes.
I would add a Tag to the model and then search based on the project and the tag, wdyt?

It can also work by running on multiple known nodes.

Horovod sits on top of openmpi that needs ssh to open multiple nodes, I'm not sure how one would connect it without passing the SSH keys from one node to the other, and making sure they can directly communicate. (Not saying it is not possible, but just a few things to configure before it works, the enterprise edition remove the need for the direct SSH connection between the nodes)

How would i add a glue for multinode?

Basic...

FrustratingWalrus87 If you need active one, I think there is currently no alternative to TB tSNE 🙂 it is truly great 🙂
That said you can use plotly for the graph:
https://plotly.com/python/t-sne-and-umap-projections/#project-data-into-3d-with-tsne-and-pxscatter3d
and report it to ClearML with Logger report_plotly :
https://github.com/allegroai/clearml/blob/e9f8fc949db7f82b6a6f1c1ca64f94347196f4c0/examples/reporting/plotly_reporting.py#L20

Is this reproducible ?

Hi WickedElephant66

Setting the pipeline controller with pipeline_execution_queue as None

is actually launching the pipeline controller on your "dev" machine, not sure why you have two of them?

Hi UnsightlyShark53 , just a quick FYI, you can also log the entire config file config.json this will be stored as model configuration, and you can see it in the input/output models under the artifacts tab.
See example here you can path either the path to the configuration file, or the dictionary itself after you loaded the json, whatever is more convenient :)

UnsightlyShark53 See if this one solves the problem :)
BTW: the reasoning for the message is that when running the task with "trains-agent" if the parsing of the argparser happens before the the Task is initialized, the patching code doesn't know if it supposed to override the values. But this scenario was fixed a long time ago, and I think the error was mistakenly left behind...

Oh, did you try task.connect_configuration
?
https://allegro.ai/docs/examples/reporting/model_config/#using-a-configuration-file

Having the ability to pack jobs/tasks onto the same "resource" (underlying server/EC2 instance)

This is essentially a "queue". Basically a queue is a way to abstract a specific type of resource, so that you can achieve exactly what you descibed.

open up a streaming use case, wherein batch (offline) inference could be done directly inside of a ClearML pipeline in reaction to an event/trigger (like new data landing in your data lake).

Yes, that's exactly how clearml is designed, a...

When looking at the worker details, it says "No queues currently assigned to this worker"

Yes, I think we should have better information there, the "AWS service" is not directly pulling jobs from any specific queue, hence nothing there. It is "listening" to queues and launching machines, those machines will be listening to the queue. I wonder if it is just easier to also make sure it is listed as "assigned" to those queues . wdyt?

(i.e. importing the trains package is enough to patch the argparser, only when you call the task.init the arguments will be logged, before they are stored in memory)

Yes you can 🙂 (though not on the open-source version)

it was uploading fine for most of the day

What do you mean by uploading fine most of the day ? are you suggesting the upload stuck to the GS ? are you seeing the other metrics (scalars console logs etc) ?

Hi WorriedParrot51
Assuming you run the code "manually" once (i.e. without the agent). Then when you call Task.init it will register the argparser.
When running with the agent, the first time you will call parse, it will automatically override the argparse defaults with the values stored in the Task.
Make sesne?

am getting None for Task.current_task() at the beginning of my script.

Task.init() is doing the magic , only after this call you will have current_task (either running manua...

We are here if you need further help 🙂

Task.add_requirements does not handle it (traceback in the thread). Any suggestions?

Hmm that is a good point maybe we should fix that 🙂
I'm assuming someone already created this module? Or is it part of the repository?
(if it than the assume this executed from the git root)

understood, can you try
Task.add_requirements("-e path/to/folder/")

I think the main issue is running with python -m module.name --args
Which is a bit different, when trying to "understand" what is the actual repository.
Can you try to run it from the repository folder (same command, just to see if it will have any effect on the detected packages)

BTW: what happens if you pass the same s3://bucket to Task.init output_uri ? I assume you are getting the same access issue ?

Hi UnsightlyShark53 apologies for this delayed reply, slack doesn't alert users unless you add @ , so things sometimes get lost :(
I think you pointed at the correct culprit...
Did you manage to overcome the circular include?
BTW , how could I reproduce it? It will be nice if we could solve it

Martin, if you want, feel free to add your answer in the stackoverflow so that I can mark it as a solution.

Will do 🙂 give me 5

Hi GrittyKangaroo27
Maybe check the TriggerScheduler , and have a function trigger something on k8s every time you "publish" a model?
https://github.com/allegroai/clearml/blob/master/examples/scheduler/trigger_example.py

HI SubstantialElk6
Yes you are correct the glue only needs to change the yaml and it will work.
When you say "Dev end" , what do you mean? I was thinking adding additional glue for multi node and just adding queues , for example add 4nodes queue and attach a glue to it, wdyt?
Regrading horovod, horovod is spinning its own nodes so integration with k8s is not trivial (regardless of ClearML). That said I know that they do have support for horovod in the Enterprise edition, but I'm not sure ...

I think this was the issue: None
And that caused TF binding to skip logging the scalars and from that point it broke the iteration numbering and so on.

Hi SpicyOtter88
plt.plot([0, 1], [0, 1], 'r--', label='')ti cannot have a legend without a label, so it gives it "anonymous" label, I think it should just get "unlabeled 0" wdyt?

Hmm I see your point.
Any chance you can open a github issue with a small code snippet to make sure we can reproduce and fix it?

(I am not an expert on UI to be honest)

Same here 🙂 lol

we can implement this externally

What do you mean by that?

Hmm, it is not returned, it is inside the function....

HealthyStarfish45 you mean as in RestAPI ?

Yes exactly like a Task (pipeline is a type of task)
'''
clonedpipeline=Task.clone(pipeline_uid_here)
Task.enqueue(...)
'''