AutogluonTimeseries: Keep getting this error [ValueError: Trainer has no fit models that can predict.]

[mputra533]

Hi,
I was able to run this previously and here's my system info

=================== System Info ===================
AutoGluon Version:  1.0.0
Python Version:     3.11.8
Operating System:   Linux
Platform Machine:   x86_64
Platform Version:   #1 SMP Fri Jun 9 10:57:30 UTC 2023
CPU Count:          16
GPU Count:          1
Memory Avail:       53.93 GB / 58.87 GB (91.6%)
Disk Space Avail:   66.22 GB / 94.25 GB (70.3%)

Now I have additional data around 100k rows (minute by minute data to predict 10min intervals), no missing values, no Nans.
The only difference is that this is now done in batches of a few item ids.
I have data that looks like this as a TimeSeriesDataframe loaded in a variable called data. This is the data just before being passed on fit below

item_id  timestamp value  
sr-14744 2024-02-11 20:28:01      9
         2024-02-11 20:28:10     71
         2024-02-11 20:29:01     27
         2024-02-11 20:29:10     83
         2024-02-11 20:30:01     32
...                             ...
sr-14760 2024-03-17 10:16:26     14
         2024-03-17 10:17:26     22
         2024-03-17 10:18:26     48
         2024-03-17 10:19:26     28
         2024-03-17 10:20:26     51

When I run this code on the first batch

predictor = TimeSeriesPredictor(
                prediction_length=prediction_length,
                path=forecast_folder,
                target="value",
                eval_metric="WQL",
                freq="10min",
            ).fit(
                data,
                presets="high_quality",
                time_limit=3600,
            )
predictions = predictor.predict(data)

I got the error below:

...
 Traceback (most recent call last):
  File "/path/forecast.py", line 132, in forecast
    predictions = predictor.predict(data)
                  ^^^^^^^^^^^^^^^^^^^^^^^
  File "/path/python3.11/site-packages/autogluon/timeseries/predictor.py", line 783, in predict
    predictions = self._learner.predict(data, known_covariates=known_covariates, model=model, use_cache=use_cache)
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/path/python3.11/site-packages/autogluon/timeseries/learner.py", line 176, in predict
    return self.load_trainer().predict(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/path/python3.11/site-packages/autogluon/timeseries/trainer/abstract_trainer.py", line 806, in predict
    model_name = self._get_model_for_prediction(model)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/path/python3.11/site-packages/autogluon/timeseries/trainer/abstract_trainer.py", line 785, in _get_model_for_prediction
    best_model_name: str = self.get_model_best()
                           ^^^^^^^^^^^^^^^^^^^^^
  File "/path/python3.11/site-packages/autogluon/timeseries/trainer/abstract_trainer.py", line 79, in get_model_best
    raise ValueError("Trainer has no fit models that can predict.")
ValueError: Trainer has no fit models that can predict.

Any help is appreciated, thank you!

[shchur]

Hi @mputra533, thank you for reporting the problem. Can you please call predictor.fit(..., verbosity=4) and share the error logs for the specific model failures?

[mputra533]

Hi @shchur, thank you for your reply. Turns out the prediction_length that's passed on the TimeSeriesPredictor was a float. After converting it to an integer it works.

[shchur]

Nice, thank you for sharing the solution @mputra533! I have submitted a fix to make sure that this case is handled automatically #4160.