Reduce Deadlock Probability #84
Conversation
franchuterivera
force-pushed
the
refactor_development_loggermsg
branch
from
e0f8884
to
436d941
Compare
franchuterivera
force-pushed
the
refactor_development_loggermsg
branch
from
436d941
to
39b7f09
Compare
autoPyTorch/pipeline/components/setup/traditional_ml/classifier_models/classifiers.py
Outdated
Show resolved
Hide resolved
autoPyTorch/pipeline/components/setup/traditional_ml/classifier_models/classifiers.py
Outdated
Show resolved
Hide resolved
autoPyTorch/pipeline/components/setup/traditional_ml/classifier_models/classifiers.py
Outdated
Show resolved
Hide resolved
autoPyTorch/pipeline/components/setup/traditional_ml/classifier_models/classifiers.py
Outdated
Show resolved
Hide resolved
autoPyTorch/pipeline/components/setup/traditional_ml/classifier_models/classifiers.py
Outdated
Show resolved
Hide resolved
| """ | ||
| preprocessing = [] | ||
| estimator = [] | ||
| skip_steps = ['data_loader', 'trainer', 'lr_scheduler', 'optimizer', 'network_init', |
There was a problem hiding this comment.
Maybe we can have a verbose option where we include- trainer, lr_scheduler, optimizer and network_init?
There was a problem hiding this comment.
I was coding this and seeing the outcome, and it just doesn't add much information to 'what the best estimator is' because it's part of the construction of the model, but not part of the model itself.
I have a better proposal -- I would like to create a command like this.
Tpot is able to print the python code on how to train the model that you did. So if your goal were to see what happened (and for debug purposes) it would be great to produce a file that contains PyTorch commands with not only the scheduler but also the config (like it is nice to know we know more that the fact that we picked adam optimizers). If this is better, then I would like to disentangle this export_pipeline command from this PR and create an issue for this.
What do you think?
There was a problem hiding this comment.
yes this would be ideal to use. Could you add an issue so we don't forget?
| """ | ||
| preprocessing = [] | ||
| estimator = [] | ||
| skip_steps = ['data_loader', 'trainer', 'lr_scheduler', 'optimizer', 'network_init', |
franchuterivera
force-pushed
the
refactor_development_loggermsg
branch
from
3c3797c
to
e336373
Compare
…elopment_loggermsg
…or_development_loggermsg
Use the logger port instead of the logger for the TAE execution.
Add show_models() for debug purposes
Remove tensorboard output because killing a run in the process of writing to disk halts the complete search process and python does not handle the recovery nice. This is something we should look fixing in pynisher.
Minor fixes like empty cuda in when not needed.