
Reputation
Badges 1
33 × Eureka!` from importlib.machinery import EXTENSION_SUFFIXES
import catboost
from clearml import Task, Logger, Dataset
import lightgbm as lgb
import numpy as np
import pandas as pd
import dask.dataframe as dd
import matplotlib.pyplot as plt
MODELS = {
'catboost': {
'model_class': catboost.CatBoostClassifier,
'file_extension': 'cbm'
},
'lgbm': {
'model_class': lgb.LGBMClassifier,
'file_extension': 'txt'
}
}
class ModelTrainer():
def init(sel...
Simplified a little bit and removed private parameters, but thats pretty much the code. We did not try with toy examples, since that was already done with the example pipelines when we implemented and the model training itself is quite simple basic there already (only few hyperparameters set)
Martin, if you want, feel free to add your answer in the stackoverflow so that I can mark it as a solution.
Could you supply any reference of this dataset containing other datasets? I might have skipped that when reading the documentation, but I do not recall seeing this functionality.
oooohhh.. you mean the key of the nested dict, that would make a lot of sense
I was checking here, and apparently if I use a parameter as suggested, together with a Task.init(task_name=f'{task name in this loop}')
for each of the loops it should work, right? Creating different tasks in the server
Apparently found out a solution:dataset_zip = dataset._task.artifacts['data'].get()
will return the path to the zip file containing all the files (that will be downloaded to the local machine)
after that:import zipfile zip_file = zipfile.ZipFile(d, 'r') files = zip_file.namelist()
retrieving the names of the files
unzip usingimport os os.system(f'unzip {dataset_zip}') # in this case to your script directory
and using the files
list one can them open them selectively
yes, variations of the data, using only a subset of the features
I saw regarding the chunks, but it is not clear how one can retrieve the dataset based on files
regarding (2), if use run_remote, does it also ignore the init?
Considering something along the lines of
https://github.com/allegroai/clearml/blob/master/examples/advanced/execute_remotely_example.py
That would make sense, although clearml, at least on UI, shows the deeper level of the nested dict as a int, as one would expect
I noticed that when a pipeline step returns an instance of a class, it tries to pickle. I am currently facing the issue with it not being able to pickle the output of the "load_baseline_model" function
` Traceback (most recent call last):
File "/tmp/tmpqr2zwiom.py", line 37, in <module>
task.upload_artifact(name=name, artifact_object=artifact)
File "/home/zanini/repo/RecSys/.venv/lib/python3.9/site-packages/clearml/task.py", line 1877, in upload_artifact
return self._artifacts_man...
Apparently the error comes when I try to access from get_model_and_features
the pipeline component load_model
. If it is not set as pipeline component and only as helper function (provided it is declared before the components that calls it (I already understood that and fixed, different from the code I sent above).
My code pretty much createas a dataset, uploads it, trains a model (thats where the current task starts), evaluates it and upload all the artifacts and metrics. The artifacts and configurations are upload alright, but the metrics and plots are not. As with Lavi, my code hangs on the task.close(), where it seems to be waiting for the metrics, etc but never finishes. No retry message is shown as well.
After a print I added for debug right before task.close() the only message I get in the consol...
I will try the suggested edit here
` all done
ClearML Monitor: Could not detect iteration reporting, falling back to iterations as seconds-from-start
^CTraceback (most recent call last):
File "/home/zanini/repo/RecSys/src/cli/retraining_script.py", line 710, in <module>
mr.retrain()
File "/home/zanini/repo/RecSys/src/cli/retraining_script.py", line 701, in retrain
self.task.close()
File "/home/zanini/repo/RecSys/.venv/lib/python3.9/site-packages/clearml/task.py", line 1783, in close
self.__shutdown()
File "...
It worked!
After commenting all the metric/plot reporting, we noticed the model was not uploading the artifacts to S3. A solution was to add wait_for_upload
in task.upload_artifact()
UnsightlyHorse88 , do you know?
Steps (pipeline components):
Load the model Infereces witht he model
Its equivalent tomodel = Step1(*args) preds = Step2(model, *args)
After step 1, I have the model loaded as a torch object, as expected. When this object is passed to step 2, inside of step 2, it is read as an object of class 'pathlib2.PosixPath'.
I assume that is because there is some kind of problem in the pickling/loading/dumping of the inputs from a step to another in the pipeline. Is it some kind of known issue or ...
Yes, seems indeed it was waiting for the uploads, which weren't happening ( I did give it quite a while to try to finish the process in my tests). I thought it was a problem with metrics, but apprently it was more like the artifacts before them. The artifacts were shown in the webui dashboard, but were not on S3
Looks quite good indeed! Thanks! Is there in the repository the experiment template used in this example? Just not fully sure how the parameters are used/connected in it. Could I just build it and log these parameters using task.set_parameters()
so that I call task.get_parameters()
later?
yes, but is there a way to generate multiple tasks like I mentioned using task.init in different points of a .py and and run each of them as a separate remote exercution? Didn you just say that once I trigger the task.execute_remotely it will ignore the task.init?
It is an instance of a custom class.