@<1580367723722969088:profile|SmoothDuck83> Not every plot is trivially be formed as a table (i.e. CSV), that's why the JSON export is available for all plots.
What were you considering?
DefeatedCrab47 For the most part, mlflow can serve basic ML models using scikit-learn. In contrast, Trains was designed with more general purpose ML/DL workflows in mind, for which there's no "generic" way to serve models as different scenarios can use different input encoding, models results would be represented in a variety of forms, etc.
Consider also, that creating an HTTP endpoint for model inference is quite a breeze: there are multiple examples of Flask on top of any DL/ML framework w...
DefeatedCrab47 Happy you're finding Trains useful 🙂
but it definitely has it's advantages if TRAINS would support it (early stage Data Science infrastructure).
No doubt, and I definitely see such usable example in the cards for Trains' upcoming versions...
MelancholyElk85 Thanks for calling this to attention. What do you think would have made it easier for you to notice the available extended list content?
I would assume that a "type to match" option would also have helped?
Appreciate if you could https://github.com/allegroai/clearml/issues/new/choose so this can be pushed forward.
ScrawnyLion96 Looks like a case of broken links - Check out https://clear.ml/docs/latest/docs/references/api/definitions#tasksexecution and https://clear.ml/docs/latest/docs/references/api/definitions#tasksconfiguration_item
UnevenDolphin73 I think it'd be easier to track as a separate one.
ItchyJellyfish73 Have you looked at the https://clear.ml/docs/latest/docs/clearml_agent#dynamic-gpu-allocation ?
DepressedChimpanzee34a filter similar to one in the scalars page where you can display a subset of the reported debug images can be useful
The scalars page provides a metric hide/show control - Is this the one you mean? The debug images page also provides a filter by metric - Depending on your naming policy this can easily be used to focus on more sparsely appearing images.
Else, an example of the filter you were thinking of would be appreciated.
Regardless, direct iteration access cou...
DepressedChimpanzee34 Thanks for clarifying where the current debug images display falls short for your use case - Extending the filtering to liken the behaviour of the scalars sound like a great idea 🙂
DepressedChimpanzee34 Apologies for missing your previous comment.
Totally agree that the global selection indicator should maintain its 'clear selection' behaviour even if some/all of the selection is off-screen.
DepressedChimpanzee34 Have you noticed the "Show n experiments selected" button on the bottom bar? This effectively toggles your view between whatever is currently sorted/filtered and the current item selection.
To address the scenario you describe: Switch to "Show selected experiments", remove the redundant items, and switch back to the original view: "Show all experiments"
Thoughts?
HappyDove3 Notice that in https://github.com/allegroai/clearml/issues/400 the goal is to see a table plot in the UI scalars tab for a specific experiment (with additional discussions on how these will be addressed when comparing experiments).
Note that once you take the approach you suggested of logging your metrics single values, you can configure your experiment comparison scalars view to show single values instead of the time-series graph which I think will provide you with the matrix c...
If the credentials don't provide access, the calls should fail (there's no fallback - just default values in place of empty configuration).
Notice you explicitly configure all hosts values, so you don't end up using a specific server for API access, and the default demo server for File server access...
SharpDove45 you can programmatically control the configured server using https://allegro.ai/clearml/docs/rst/references/clearml_python_ref/task_module/task_task.html?highlight=set_credentials#clearml.task.Task.set_credentials
@<1559349204206227456:profile|BeefyStarfish55> try checking out the general overview on pipelines here , and info on the pipelines UI here .
Each step's arguments (and results) should appear in the steps details panel (which you could then follow to the underlying task for complete, in-depth, details).
RotundHedgehog76 Thanks for the spot - seems like docs are wrong, and CLI help is correct: '--skip-docker-network' will NOT pass '--network host' to the docker.
HappyDove3 you can get some more insight on the different configuration methods and how to use theme https://clear.ml/docs/latest/docs/fundamentals/hyperparameters
Hi DefeatedCrab47 ,
The examples folder has just been restructured: Find the example here:
https://github.com/allegroai/trains/blob/master/examples/services/hyper-parameter-optimization/hyper_parameter_optimizer.py
DefeatedCrab47 Thanks for pointing it out.
We'll get in touch with the PyTorch Lightning team to better understand the code restructure they're effecting (see https://github.com/PyTorchLightning/pytorch-lightning/pull/2384 ).
In the mean time, you can look at the prior version: https://github.com/PyTorchLightning/pytorch-lightning/blob/0.8.1/pytorch_lightning/loggers/trains.py
DepressedChimpanzee34 Always appreciated
@<1523701157564780544:profile|TenseOstrich47> The storage in question here is what's available on the machine hosting the ClearML server's docker containers (specifically, the ES one).
IrateDolphin19 ClearML provides for saving files generated as part of your code execution through the https://clear.ml/docs/latest/docs/references/sdk/task#upload_artifact . For your use case, you can have your code thus create the artifact as it runs, you can set the specific storage location when you edit your configuration, through the task's output_uri field.
Does this help?
From the https://github.com/allegroai/trains-server/releases/tag/0.13.0 :
Reports average load metrics per day (CPU/memory) Reports average workload per day (amount and average duration of queues, agents and experiments)
MysteriousBee56 would providing Trains with an "import mode" (say, via environment or command line variable), which means that it should create a draft server entry, populate all the execution/environment info and exit before it actually starts employing the ML infrastructure address your use case?
KindGiraffe71 Have you checked out the https://github.com/allegroai/clearml/blob/master/examples/frameworks/pytorch-lightning/pytorch_lightning_example.py ? https://clearml.slack.com/archives/CTK20V944/p1616070536033700 previous discussion provides some insight into how it works under the hood.
GreasyPenguin14 That's an annoying bug indeed - Thanks for spotting it. If you need to circumvent it before a fix comes out in one of the near releases, you can programatically use the https://clear.ml/docs/latest/docs/references/api/endpoints#post-projectsupdate e.g.from clearml.backend_api.session.client import APIClient client = APIClient() client.projects.update(project='<project ID>', description='My new description')
Note you can get your project's ID either from the webapp URL...
@<1523705301990117376:profile|WickedCat12> ClearML Scalars explicitly show metrics time progression (you can display iteration/wall-time).
Plotting one metric against another is a feature that lies further down ClearML's roadmap.
If your metric is reported only once per epoch you can make use of the existing scalars functionality by making use of the iteration parameter when reporting your metric to reflect the epoch instead.
Does this make sense?
BattyLion34 Adding to AgitatedDove14 hint. See the following docs page: https://allegro.ai/clearml/docs/docs/deploying_clearml/clearml_config_for_clearml_server.html
Thanks for clarifying @<1523705301990117376:profile|WickedCat12> .
As I mentioned originally, plotting an arbitrary metric against another is further down the ClearML roadmap.
It'd be great if you use a github issue to help push it through :)
TightElk12 This makes a lot of sense - should make it into one of the coming releases