Open numomcmc opened 6 months ago
BELONOVSKII
commented
5 months ago Hello @numomcmc! The problem is that pip installs an old version of LightAutoML (<0.3.0). By default, kaggle uses python3.10 and the latest stable versions of LightAutoML are not seen for pip3.10. We are working on this.
A temporary solution is to install LightAutoML from the lightautoml-0.3.8-py3-none-any.whl file provided by @alexmryzhkov. See the notebook for an example.
numomcmc
commented
5 months ago Hello Peter
Thank you for your reply. I am having all kinds of problems using LAMA. The same call made with TabularAutoML results the following err msg with TabularUtilizedAutoML. So I'd like to have a clean reinstall to make sure everything is current. The link to the notebook returns a "No results matched your search", could you check that? I am unfortunately running Conda on Windows 11, so I hope the new version and the installation play nice together.
Thank you and I look forward to hearing from you.
Stefan
AttributeError: 'TabularUtilizedAutoML' object has no attribute 'reader'
On Wed, Jan 10, 2024 at 2:04 AM PeterBel @.***> wrote:
Hello @numomcmc https://github.com/numomcmc! The problem is that pip installs an old version of LightAutoML (<0.3.0). By default, kaggle uses python3.10 and the latest stable versions of LightAutoML are not seen for pip3.10. We are working on this.
A temporary solution is to install LightAutoML from the lightautoml-0.3.8-py3-none-any.whl https://www.kaggle.com/datasets/alexryzhkov/lightautoml-v0-3-8 file provided by @alexmryzhkov https://github.com/alexmryzhkov. See the notebook https://github.com/sb-ai-lab/LightAutoML/issues/url for an example.
— Reply to this email directly, view it on GitHub https://github.com/sb-ai-lab/LightAutoML/issues/147#issuecomment-1884541261, or unsubscribe https://github.com/notifications/unsubscribe-auth/AX6JLHOXI7WK3HZED5EY7YDYNZRUJAVCNFSM6AAAAABBTLJMZ6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQOBUGU2DCMRWGE . You are receiving this because you were mentioned.Message ID: @.***>
BELONOVSKII
commented
5 months ago Stefan, hello once again! I can not figure out why url is not working, so I will provide the code snippet here.
Firstly download the lightautoml-0.3.8-py3-none-any.whl file from here.
Then run pip install lightautoml-0.3.8-py3-none-any.whl inside your environment.
BELONOVSKII
commented
5 months ago Now, you can install LightAutoML on python3.10 by directly specifying the version0.3.8.b1,
i.e pip install -U lightautoml==0.3.8b1
numomcmc
commented
5 months ago Hi Peter
The new LAMA doesn't play nice with python 3.9.x, so I ran my notebook on Kaggle w/ 3.10, and it was a success. Thank you for your swift response!
In the meantime, a question about GPU support. Both LightGBM and Catboost can utilize GPU via CUDA. However, LAMA seems to have its own version of Catboost and LightGBM packaged together, and the fact Github page indicated that GPU support is conditionally available, it suggests to me that I can not run LightGBM on GPU via parameters (same for Catboost) the usual way. Is that the correct understanding?
Thank you for your support.
S On Thu, Jan 11, 2024 at 12:25 AM PeterBel @.***> wrote:
Update
Now, you can install LightAutoML on python3.10 by directly specifying the version0.3.8.b1, i.e pip install -U lightautoml==0.3.8b1
— Reply to this email directly, view it on GitHub https://github.com/sb-ai-lab/LightAutoML/issues/147#issuecomment-1886604052, or unsubscribe https://github.com/notifications/unsubscribe-auth/AX6JLHPBSYFOFSQSHRXSVPLYN6OWPAVCNFSM6AAAAABBTLJMZ6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQOBWGYYDIMBVGI . You are receiving this because you were mentioned.Message ID: @.***>
BELONOVSKII
commented
5 months ago Hello, Stephen. LightGBM and Catboost automatically switch to GPU, if available.
Moreover, there is an explicit GPU implementation of LightAutoML. Check LightAutoML_GPU.
numomcmc
commented
5 months ago Hi Peter
A few follow up questions.
I would like to give LightAutoML_GPU a try, is the GPU version in synch with the regular LAMA? Meaning, it's now based 0.3.8b1?
Also, I like to learn more about how to use LAMA to tune hyperparameters. The YouTube video talked about how LAMA tunes hyperparameters, but there is no mention of how to actually do it. There is hardly anything about hyperparameter tuning in the documentation as well. Could you provide more information?
In the Kaggle Home Price regression example, as well as the documentation ( https://lightautoml.readthedocs.io/en/latest/_modules/lightautoml/automl/presets/tabular_presets.html#). There are mentions of "lgb_tuned" and "cb_tuned". So I assume that allows users to specify hyperparameter tuning as a part of the training. But how does LAMA actually tune them? Which parameters are tuned, to what range of values? Are the user supplied parameter values considered as the starting value? What are the final resulting hyperparameter values? How does the "optimized" validation compare to non-optimized results, where can I find the comparison? Are they available in the report?
Many thanks for your help.
Stefan
On Fri, Jan 12, 2024 at 7:44 AM PeterBel @.***> wrote:
Hello, Stephen. Yes, this LAMA is developed only for CPU computations.
However, there is an explicit GPU implementation of LightAutoML that could be run on a GPU with CUDA. Check LightAutoML_GPU https://github.com/sb-ai-lab/LightAutoML_GPU.
— Reply to this email directly, view it on GitHub https://github.com/sb-ai-lab/LightAutoML/issues/147#issuecomment-1889534714, or unsubscribe https://github.com/notifications/unsubscribe-auth/AX6JLHMAD7TDDWL53H6VW5DYOFK53AVCNFSM6AAAAABBTLJMZ6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQOBZGUZTINZRGQ . You are receiving this because you were mentioned.Message ID: @.***>
🐛 Bug
To Reproduce
Steps to reproduce the behavior:
!pip install -U LightAutoMLERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts. yellowbrick 1.3.post1 requires numpy<1.20,>=1.16.0, but you have numpy 1.21.6 which is incompatible. tensorflow 2.4.1 requires numpy~=1.19.2, but you have numpy 1.21.6 which is incompatible. pdpbox 0.2.1 requires matplotlib==3.1.1, but you have matplotlib 3.4.2 which is incompatible. matrixprofile 1.1.10 requires protobuf==3.11.2, but you have protobuf 3.17.3 which is incompatible. kornia 0.5.5 requires numpy<=1.19, but you have numpy 1.21.6 which is incompatible. imbalanced-learn 0.8.0 requires scikit-learn>=0.24, but you have scikit-learn 0.23.2 which is incompatible. Successfully installed CairoSVG-2.7.1 LightAutoML-0.3.7.3 Pyphen-0.14.0 StrEnum-0.4.15 alabaster-0.7.13 autowoe-1.3.2 cairocffi-1.6.1 catboost-1.2.2 cssselect2-0.7.0 dask-2022.2.0 distributed-2022.2.0 efficientnet-pytorch-0.7.1 featuretools-1.11.1 holidays-0.27.1 imagesize-1.4.1 importlib-metadata-1.7.0 json2html-1.3.0 numpy-1.21.6 opencv-python-4.5.2.52 pandas-1.3.5 poetry-core-1.6.1 snowballstemmer-2.2.0 sphinx-4.3.2 sphinxcontrib-applehelp-1.0.2 sphinxcontrib-devhelp-1.0.2 sphinxcontrib-htmlhelp-2.0.0 sphinxcontrib-jsmath-1.0.1 sphinxcontrib-qthelp-1.0.3 sphinxcontrib-serializinghtml-1.1.5 tinycss2-1.2.1 tqdm-4.66.1 weasyprint-52.5 woodwork-0.16.4 WARNING: Running pip as root will break packages and permissions. You should install packages reliably by using venv: https://pip.pypa.io/warnings/venv
from lightautoml.automl.presets.tabular_presets import TabularAutoML, TabularUtilizedAutoMLTypeError Traceback (most recent call last)