GPU's defaulting to 0 in Timeseries forecasting

[alexkovaceski]

My windows machine is recognising in anaconda with AG installed that GPU=1, but when I run the forecasting model, it defaults to GPU=0. Any idea what's going on, I am now lost for ideas.

predictor = TimeSeriesPredictor(
... prediction_length=48,
... path="autogluon-m-minutes",
... target="target",
... eval_metric="MASE",
... freq="T",
... quantile_levels=[0.1 ,0.3, 0.7, 0.9],
... )
Warning: path already exists! This predictor may overwrite an existing predictor! path="autogluon-m-minutes"

predictor.fit(
... train_data,
... presets="best_quality",
... time_limit=60*120,
... verbosity=4,
... num_val_windows=3,
... )
Beginning AutoGluon training... Time limit = 7200s
AutoGluon will save models to 'autogluon-m-minutes'
=================== System Info ===================
AutoGluon Version: 1.0.0
Python Version: 3.10.13
Operating System: Windows
Platform Machine: AMD64
Platform Version: 10.0.19045
CPU Count: 12
GPU Count: 1
Memory Avail: 54.15 GB / 63.77 GB (84.9%)
Disk Space Avail: 637.21 GB / 1862.10 GB (34.2%)
===================================================
Setting presets to: best_quality

Fitting with arguments:
{'enable_ensemble': True,
'eval_metric': MASE,
'freq': 'T',
'hyperparameters': 'default',
'known_covariates_names': [],
'num_val_windows': 3,
'prediction_length': 48,
'quantile_levels': [0.1, 0.3, 0.7, 0.9],
'random_seed': 123,
'refit_every_n_windows': 1,
'refit_full': False,
'target': 'target',
'time_limit': 7200,
'verbosity': 4}

train_data contains missing values represented by NaN. They have been filled by carrying forward the last valid observation.
Provided train_data has 979999 rows, 1 time series. Median time series length is 979999 (min=979999, max=979999).
Setting random seed to 123

Provided dataset contains following columns:
target: 'target'
past covariates: ['aapl', 'amzn', 'googa', 'googb', 'msft', 'nvda', 'sqqq', 'tsla']
Saving autogluon-m-minutes\models\trainer.pkl
Saving autogluon-m-minutes\learner.pkl

AutoGluon will gauge predictive performance using evaluation metric: 'MASE'
This metric's sign has been flipped to adhere to being higher_is_better. The metric score can be multiplied by -1 to get the metric value.

Starting training. Start time is 2024-01-18 23:16:55
Saving autogluon-m-minutes\utils\data\train.pkl
Models that will be trained: ['SeasonalNaive', 'CrostonSBA', 'NPTS', 'AutoETS', 'DynamicOptimizedTheta', 'AutoARIMA', 'RecursiveTabular', 'DirectTabular', 'DeepAR', 'TemporalFusionTransformer', 'PatchTST']
Reserving 599.6s for ensemble
Training timeseries model SeasonalNaive. Training for up to 599.6s of the 7195.4s of remaining time.
Fitting SeasonalNaive with 'num_gpus': 0, 'num_cpus': 12
Window 0
Shortening all time series to at most 2500
Predicting with model SeasonalNaive\W0
-7.4844 = Validation score (-MASE)
0.000 s = Training runtime
1.416 s = Prediction runtime
Window 1

[shchur]

Hi, this is a mistake in the debug logging. If you see

GPU Count: 1

at the start of AutoGluon training, then all models that support GPU (=deep learning models like TFT, PatchTST, DeepAR) will use the GPU. You can verify this by checking the GPU utilization during training.

You can safely ignore the debug message Fitting SeasonalNaive with 'num_gpus': 0, 'num_cpus': 12, it's logged by mistake and we will try to fix it soon.

[alexkovaceski]

Thanks