SuccessfulKoala55 just in case you have any more thoughts, but we could also continue as is 😊

Hi @<1523703436166565888:profile|DeterminedCrab71> and @<1523701070390366208:profile|CostlyOstrich36> , coming back to this after a while. It actually seems to be related to Google Cloud permissions:

• The images in the ClearML dashboard to not show as discussed above
• If I copy the image url (coming out as something like None and open it in another tab where I’m logged into my Google Account, the image loads
• If I do t...

It was related, special characters also made prevented some access.
But it was and is also related to some authentication problem with Google: If you open the dashboard in Chrome, go to the developer console, you see a bunch of failed links to authenticate to. If you click one of them in another tab, it shows the Google signin screen and afterwards you can see the Debug Samples in Dashboard.
That all does not work in safari though for some reason 🙂

Ok I see, that is what I thought. But do you have any idea why I am not seeing these images? I am logged into my Gmail account and into the Google Cloud Console and can access both in another tab of the same browser. Am I missing something here?

Ok so actually if I run task.flush(wait_for_uploads=True) at the end of the script it works ✔

👍 👍

Hi John, thanks for getting back to me!
So it shows up in the UI like shown below. It happens both when “recording” the local run on Mac and on Linux.

Happy to and thanks!

Yes makes sense, it sounded like that from the start. Luckily, the task.flush(...) way seems to work for now 🙂

Hey guys, really appreciating the help here!
So what I meant by “it does work” is that the environment variables go through to the container, I can use them there, everything runs.

The remaining problem is that this way, they are visible in the ClearML web UI which is potentially unsafe / bad practice, see screenshot below.

Hi SuccessfulKoala55 , thanks for getting back to me!
In the docs of Task.set_base_docker() it says “When running remotely the call is ignored”. Does that mean that this function call is executed when running locally to “record” the arguments and then when I duplicate the experiment and clone it remote, the call is ignored and the recorded values are used?

Ok great! I will debug starting with a simpler training script.
Just as a last question, is torchrun also supported rather than the (now deprecated but still usable) torch.distributed.launch ?

Ok, I re-checked and saw that the data was indeed cached and re-loaded - maybe I waited a little too long last time and it was already a new instance. Awesome implementation guys!

Sorry that these issues go quite deep and chaotic - we would appreciate any help or ideas you can think of!

Won’t they be printed out as well in the Web UI? That shows the full Docker command for running the task right…

Results of a bit more investigation:

The ClearML example does use the Pytorch dist package but none of the DistributedDataParallel functionality, instead, it reduces gradients “manually”. This script is also not prepared for torchrun as it launches more processes itself (w/o using the multiprocessing of Python or Pytorch.)

When running a simple example (code attached...

So my own repo I’m launching with either
torchrun --nproc_per_node 2 --standalone --master_addr 127.0.0.1 --master_port 29500 -m http://my_folder.my _script --some_option
or
python3 -m torch.distributed.launch --nproc_per_node 2 --master_addr 127.0.0.1 --master_port 29500 -m http://my_folder.my _script --some_option

@<1523701070390366208:profile|CostlyOstrich36> thank you, now everything works so far!
Last thing: Is there any way to change all the links in the new ClearML server such that an artifact that was previous under s3://… is now taken from gs://… ? The actual data is already available under the gs:// link of course

Hi @<1523701087100473344:profile|SuccessfulKoala55> , sorry there was a mistake on my end - clearml.conf pointed to the wrong URL 🙈

AgitatedDove14 maybe to come at this from a broader angle:
Is ClearML combined with DataParallel or DistributedDataParallel officially supported / should that work without many adjustments? If so, would it be started via python ... or via torchrun ... ? What about remote runs, how will they support the parallel execution? To go even deeper, what about the machines started via ClearML Autoscaler? Can they either run multiple agents on them and/or start remote distribu...

Hi Jake, yes I’d love to! Just a question: how clean and complete does the example need to be? For example, this code relies on you building a correct Machine Image on GCP (which is somewhat unrelated to ClearML) and it does not get the logs from the agent instances - is that still good enough?

Yes, when the WebUI prompted me for them. They also seem to work since images in Debug Samples (also in S3) show up after I entered them.
Also, I can see that the plot is also saved in Debug Samples after explicit reporting, even though I don’t set report_interactive=False

When running on our bigger research repository which includes saving checkpoints and uploading to S3, the training ends with errors as shown below and a Killed message for the main process (I do not abort the main process manually):

2023-01-26 17:37:17,527 INFO: Save the latest model.
2023-01-26 17:37:19,158 - clearml.storage - INFO - Starting upload: /tmp/.clearml.upload_model_cvqpor8r.tmp => glass-clearml/RealESR/Glass-ClearML Demo/[Lambda] FMEN distributed check, v10 fileserver u...

More stack trace:

clearml-elastic   | ElasticsearchException[failed to bind service]; nested: AccessDeniedException[/usr/share/elasticsearch/data/nodes];
clearml-elastic   | Likely root cause: java.nio.file.AccessDeniedException: /usr/share/elasticsearch/data/nodes
clearml-elastic   |     at java.base/sun.nio.fs.UnixException.translateToIOException(UnixException.java:90)
clearml-elastic   |     at java.base/sun.nio.fs.UnixException.rethrowAsIOException(UnixException.java:106)
clearml-el...

I actually wanted to load a specific artifact, but didn’t think of looking through the tasks output models. I have now changed to that approach which feels much safer, so we should be all done here. Thanks!

Sorry to ask again, but the values are still showing up in the WebUI console logs this way (see screenshot.)
Here is the config that I paste into the EC2 Autoscaler Setup:
` agent {
extra_docker_arguments: ["-e AWS_ACCESS_KEY_ID=XXXXXX", "-e AWS_SECRET_ACCESS_KEY=XXXXXX"]

hide_docker_command_env_vars {
    enabled: true
    extra_keys: ["AWS_SECRET_ACCESS_KEY"]
    parse_embedded_urls: true
}    

} `Never mind, it came from setting the options wrong, it has to be ...

So the container itself gets deleted but everything is still cached because the cache directory is mounted to the host machine in the same place? Makes absolute sense and is what I was hoping for, but I can’t confirm this currently - I can see that the data is reloaded each time, even if the machine was not shut down in between. I’ll check again to find the cached data on the machine

Although, some correction here: While the secret is indeed hidden in the logs, it is still visible in the “execution” tab of the experiment, see two screenshots below.
One again I set them with
task.set_base_docker(docker_arguments=["..."])

That was the missing piece - thank you!
Awesome to all the details you have considered in ClearML 😉

Well duh, now it makes total sense! Should have checked docs or examples more closely 🙏
Yes if that works reliably then I think that option could make sense, it would have made things somewhat easier in my case - but this is just as good.