.. _sec_tabulargpu:

Training models with GPU support
================================


Training with GPU can significantly speed up base algorithms, and is a
necessity for text and vision models where training without GPU is
infeasibly slow. CUDA toolkit is required for GPU training. Please refer
to the `official documentation <https://docs.nvidia.com/cuda/>`__ for
the installation instructions.

.. code:: python

   predictor = TabularPredictor(label=label).fit(
       train_data,
       num_gpus=1,  # Grant 1 gpu for the entire Tabular Predictor
   )

To enable GPU acceleration on only specific models, the same parameter
can be passed into model ``hyperparameters``:

.. code:: python

   hyperparameters = {
       'GBM': [
           {'ag_args_fit': {'num_gpus': 0}},  # Train with CPU
           {'ag_args_fit': {'num_gpus': 1}}   # Train with GPU. This amount needs to be <= total num_gpus granted to TabularPredictor
       ]
   }
   predictor = TabularPredictor(label=label).fit(
       train_data, 
       num_gpus=1,
       hyperparameters=hyperparameters, 
   )

Multi-modal
-----------

In :ref:`sec_tabularprediction_multimodal` tutorial we presented how
to train an ensemble which can utilize tabular, text and images. If
available GPUs don’t have enough VRAM to fit the default model, or it is
needed to speedup testing, different backends can be used:

Regular configuration is retrieved like this:

.. code:: python

    from autogluon.tabular.configs.hyperparameter_configs import get_hyperparameter_config
    hyperparameters = get_hyperparameter_config('multimodal')
    hyperparameters




.. parsed-literal::
    :class: output

    {'NN_TORCH': {},
     'GBM': [{},
      {'extra_trees': True, 'ag_args': {'name_suffix': 'XT'}},
      'GBMLarge'],
     'CAT': {},
     'XGB': {},
     'AG_TEXT_NN': {'presets': 'medium_quality_faster_train'},
     'AG_IMAGE_NN': {},
     'VW': {}}



Text models
~~~~~~~~~~~

Text model preset to use can be set via:

.. code:: python

   hyperparameters['AG_TEXT_NN'] = ['<preset>']

Available text model presets:

.. code:: python

    from autogluon.text.text_prediction.presets import list_text_presets
    list_text_presets()




.. parsed-literal::
    :class: output

    ['default',
     'medium_quality_faster_train',
     'high_quality',
     'best_quality',
     'multilingual']



Vision models
~~~~~~~~~~~~~

Vision model preset to use can be set via:

.. code:: python

   hyperparameters['AG_IMAGE_NN'] = {'model': '<model>'}

The list of available text model presets is:

.. code:: python

    from autogluon.vision.predictor.predictor import _get_supported_models
    _get_supported_models()[:10]  # there're more, we just show a few




.. parsed-literal::
    :class: output

    ['adv_inception_v3',
     'bat_resnext26ts',
     'beit_base_patch16_224',
     'beit_base_patch16_224_in22k',
     'beit_base_patch16_384',
     'beit_large_patch16_224',
     'beit_large_patch16_224_in22k',
     'beit_large_patch16_384',
     'beit_large_patch16_512',
     'beitv2_base_patch16_224']



Enabling GPU for LightGBM
-------------------------

The default installation of LightGBM does not support GPU training,
however GPU support can be enabled via a special install. If
``num_gpus`` is set, the following warning will be displayed:

::

   Warning: GPU mode might not be installed for LightGBM, GPU training raised an exception. Falling back to CPU training...Refer to LightGBM GPU documentation: https://github.com/Microsoft/LightGBM/tree/master/python-package#build-gpu-versionOne possible method is:  pip uninstall lightgbm -y   pip install lightgbm --install-option=--gpu

If the suggested commands do not work, uninstall existing lightgbm
``pip uninstall -y lightgbm`` and install from sources following the
instructions in the `official
guide <https://lightgbm.readthedocs.io/en/latest/GPU-Tutorial.html>`__.
The optional `Install Python
Interface <https://lightgbm.readthedocs.io/en/latest/GPU-Tutorial.html#install-python-interface-optional>`__
section is also required to make it work with AutoGluon.

Advanced Resource Allocation
----------------------------

Most of the time, you would only need to set ``num_cpus`` and
``num_gpus`` at the predictor ``fit`` level to control the total
resources you granted to the TabularPredictor. However, if you want to
have more detailed control, we offer the following options.

``ag_args_ensemble: ag_args_fit: { RESOURCES }`` allows you to control
the total resources granted to a bagged model. If using parallel folding
strategy, individual base model’s resources will be calculated
respectively. This value needs to be <= total resources granted to
TabularPredictor This parameter will be ignored if bagging model is not
enabled.

``ag_args_fit: { RESOURCES }`` allows you to control the total resources
granted to a single base model. This value needs to be <= total
resources granted to TabularPredictor and <= total resources granted to
a bagged model if applicable.

As an example, consider the following scenario

.. code:: python

   predictor.fit(
       num_cpus=32,
       num_gpus=4,
       hyperparameters={
           'NN_TORCH': {},
       },
       num_bag_folds=2,
       ag_args_ensemble={
           'ag_args_fit': {
               'num_cpus': 10,
               'num_gpus': 2,
           }
       },
       'ag_args_fit': {
           'num_cpus': 4,
           'num_gpus': 0.5,
       }
       hyperparameter_tune_kwargs={
           'searcher': 'random',
           'scheduler': 'local',
           'num_trials: 2
       }
   )

We train 2 HPO trials, which trains 2 folds in parallel at the same
time. The total resources granted to the TabularPredictor is 32 cpus and
4 gpus.

For a bagged model, we grant 10 cpus and 2 gpus. This means we would run
two HPO trials in parallel, each granted 10 cpus and 2 gpus -> 20 cpus
and 4 gpus in total.

We also specified that for an individual model base we want 4 cpus and
0.5 gpus and we can train two folds in parallel according to the bagged
level resources -> 8 cpus and 1 gpus for a bagged model -> 16 cpus and 2
gpus when two trials running in parallel.

Therefore, we will use 16 cpus and 2 gpus in total and have two trials
of bagged model running in parallel each running two folds in parallel
-> 4 models training in parallel.

Troubleshooting
---------------

::

   OSError: libcudnn.so.X: cannot open shared object file: No such file or directory
   OSError: libcudart.so.XX.Y: cannot open shared object file: No such file or directory

This might happen when installed cuda is not matching MXNet library. To
resolve it, get the cuda version installed in system: ``nvcc --version``
or ``nvidia-smi``. Then install matching ``mxnet-cuXXX`` package (CUDA
``11.0`` -> ``mxnet-cu110``, etc.)

::

   pip install 'mxnet-cu110<2.0.0'