Model can't find encoding_available variable [BUG]

[Joe-TheBro]

Describe the bug
After fitting TCN model, when attempting to backtest program spits out error regarding self.encoders.encoding_available

Error

AttributeError                            Traceback (most recent call last)
/tmp/ipykernel_1955/95313221.py in <module>
      7     retrain=False,
      8     num_samples=1,
----> 9     verbose=True,
     10 )

/lib/python3.7/site-packages/darts/utils/utils.py in sanitized_method(self, *args, **kwargs)
    170 
    171                 getattr(self, sanity_check_method)(*only_args.values(), **only_kwargs)
--> 172             return method_to_sanitize(self, *only_args.values(), **only_kwargs)
    173 
    174         return sanitized_method

/lib/python3.7/site-packages/darts/models/forecasting/forecasting_model.py in historical_forecasts(self, series, past_covariates, future_covariates, num_samples, start, forecast_horizon, stride, retrain, overlap_end, last_points_only, verbose)
    421                 past_covariates=past_covariates,
    422                 future_covariates=future_covariates,
--> 423                 num_samples=num_samples,
    424             )
    425 

/lib/python3.7/site-packages/darts/models/forecasting/forecasting_model.py in _predict_wrapper(self, n, series, past_covariates, future_covariates, num_samples)
   1008             past_covariates=past_covariates,
   1009             future_covariates=future_covariates,
-> 1010             num_samples=num_samples,
   1011         )
   1012 

/lib/python3.7/site-packages/darts/utils/torch.py in decorator(self, *args, **kwargs)
     68         with fork_rng():
     69             manual_seed(self._random_instance.randint(0, high=MAX_TORCH_SEED_VALUE))
---> 70             return decorated(self, *args, **kwargs)
     71 
     72     return decorator

/lib/python3.7/site-packages/darts/models/forecasting/torch_forecasting_model.py in predict(self, n, series, past_covariates, future_covariates, trainer, batch_size, verbose, n_jobs, roll_size, num_samples, num_loader_workers)
   1076         )
   1077 
-> 1078         if self.encoders.encoding_available:
   1079             past_covariates, future_covariates = self.encoders.encode_inference(
   1080                 n=n,

AttributeError: 'NoneType' object has no attribute 'encoding_available'

To Reproduce

backtest_series = model.historical_forecasts(
    series_transformed,
    start=series_transformed.get_timestamp_at_point(0.6),
    forecast_horizon=7,
    stride=5,
    retrain=False,
    num_samples=1, 
    verbose=True,
)

Expected behavior
I expect to backtest on fitted model

System

• Python version: 3.7.11
• darts version 0.17.1

[dennisbader]

Hi @Joe-TheBro, I couldn't reproduce this error in version 0.17.1. Could you provide a code snippet to do so?

[Joe-TheBro]

Relevant code

series = TimeSeries.from_dataframe(data)

series = series.astype(np.float32)
train, val = series.split_after(0.6)
scaler = Scaler()
train_transformed = scaler.fit_transform(train)
val_transformed = scaler.transform(val)
series_transformed = scaler.transform(series)

trainer_kwargs={"accelerator": "gpu", "gpus": 1, "auto_select_gpus": True}

model = TCNModel(
    n_epochs=1, 
    input_chunk_length=400,
    output_chunk_length=50, 
    dropout=0, 
    dilation_base=2, 
    weight_norm=True,
    kernel_size=7,
    num_filters=4,
    random_state=RANDOM_STATE,
    save_checkpoints=True,
    pl_trainer_kwargs=trainer_kwargs
)

model.fit(series=train_transformed, val_series=val_transformed)

backtest_series = model.historical_forecasts(
    series_transformed,
    start=series_transformed.get_timestamp_at_point(0.6),
    # Split data 60 train / 40 validate-test --- value
    forecast_horizon=7,
    stride=5,
    retrain=False,
    num_samples=1, 
    verbose=True,
)

Then error is thrown out

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
/tmp/ipykernel_1955/95313221.py in <module>
      7     retrain=False,
      8     num_samples=1,
----> 9     verbose=True,
     10 )

/pdrive/dt-env/lib/python3.7/site-packages/darts/utils/utils.py in sanitized_method(self, *args, **kwargs)
    170 
    171                 getattr(self, sanity_check_method)(*only_args.values(), **only_kwargs)
--> 172             return method_to_sanitize(self, *only_args.values(), **only_kwargs)
    173 
    174         return sanitized_method

/pdrive/dt-env/lib/python3.7/site-packages/darts/models/forecasting/forecasting_model.py in historical_forecasts(self, series, past_covariates, future_covariates, num_samples, start, forecast_horizon, stride, retrain, overlap_end, last_points_only, verbose)
    421                 past_covariates=past_covariates,
    422                 future_covariates=future_covariates,
--> 423                 num_samples=num_samples,
    424             )
    425 

/pdrive/dt-env/lib/python3.7/site-packages/darts/models/forecasting/forecasting_model.py in _predict_wrapper(self, n, series, past_covariates, future_covariates, num_samples)
   1008             past_covariates=past_covariates,
   1009             future_covariates=future_covariates,
-> 1010             num_samples=num_samples,
   1011         )
   1012 

/pdrive/dt-env/lib/python3.7/site-packages/darts/utils/torch.py in decorator(self, *args, **kwargs)
     68         with fork_rng():
     69             manual_seed(self._random_instance.randint(0, high=MAX_TORCH_SEED_VALUE))
---> 70             return decorated(self, *args, **kwargs)
     71 
     72     return decorator

/pdrive/dt-env/lib/python3.7/site-packages/darts/models/forecasting/torch_forecasting_model.py in predict(self, n, series, past_covariates, future_covariates, trainer, batch_size, verbose, n_jobs, roll_size, num_samples, num_loader_workers)
   1076         )
   1077 
-> 1078         if self.encoders.encoding_available:
   1079             past_covariates, future_covariates = self.encoders.encode_inference(
   1080                 n=n,

AttributeError: 'NoneType' object has no attribute 'encoding_available'

[dennisbader]

Hey @Joe-TheBro, I tried your code snippet and it runs without issues..

This seems strange to me. The encoders should be initialized when model.fit() is called.

Here are some questions:

• Can you verify your darts version by printing print(darts.__version__)
• As I can see, you used a jupyter notebook to run the code (is that correct?).
• Is it possible that your model.fit() failed and that your model is still in its initial state?
• Can you print print(model.epochs_trained) and check if this is > 0?

[Joe-TheBro]

Questions in order

• Can you verify your darts version by printing print(darts.__version__)
  Result

  >>> import darts
  >>> print(darts.__version__)
  0.17.1
  >>>
  

• As I can see, you used a jupyter notebook to run the code (is that correct?).
  Yes a jupyter notebook is being used

• Is it possible that your model.fit() failed and that your model is still in its initial state?
  I'm not positive but I would assume model.fit() did not fail, while fitting I kept track of its val loss and saw it reduce in
  value so I assume that the model is not in its initial state. After training the model I did run
  model.save_model('dtmodel.pth.tar'). I don't know if saving the model and running it after would affect it though.

• Can you print print(model.epochs_trained) and check if this is > 0?
  Result

  model.load_model('dt-model.pth.tar')
  print(model.epochs_trained)
  0
  

  That is so weird, I don't know if it has to do with me saving the model and then loading it later on, but it says that epochs_trained is 0. If you want I can train and then see what epochs_trained says before saving or loading.

Thanks for the help!

[dennisbader]

Can you try model = TCNModel.load_model('dt-model.pth.tar') and see if this resolves the issue?
load_model() is a class method and returns the loaded model as a new TCNModel object.

[Joe-TheBro]

After running

model = TCNModel.load_model('dt-model.pth.tar')
print(model.epochs_trained)
1

so it seems this does the trick to reinitialize model. Then after this trying to backtest succeeds!

So, in summary, by not specifying TCNModel when reloading model, the program defaults to the torch forecasting object instead of the TCNModel object.

I appreciate all your help @dennisbader and commend your patience. As a suggestion going forward I would either update the documentation to let people know they have to specify the specific model when loading, or integrate into codebase with load_model() method that checks to see which model object fits and instantiates as such.

[Joe-TheBro]

Unfortunately, I have another bug that has arisen. Although I do not know if it is related. After running backtest code.

backtest_series = model.historical_forecasts(
    series=series_transformed,
    start=series_transformed.get_timestamp_at_point(0.9995),
    # Split data 60 train / 40 validate-test --- value
    forecast_horizon=7,
    stride=5,
    retrain=False,
    num_samples=1, 
    verbose=True
)

An error occurs after predicting the entire specified length occurs

ValueError                                Traceback (most recent call last)
/tmp/ipykernel_2006/609873824.py in <module>
      7     retrain=False,
      8     num_samples=1,
----> 9     verbose=True
     10 )
     11 print(backtest_series)

/pdrive/dt-env/lib/python3.7/site-packages/darts/utils/utils.py in sanitized_method(self, *args, **kwargs)
    170 
    171                 getattr(self, sanity_check_method)(*only_args.values(), **only_kwargs)
--> 172             return method_to_sanitize(self, *only_args.values(), **only_kwargs)
    173 
    174         return sanitized_method

/pdrive/dt-env/lib/python3.7/site-packages/darts/models/forecasting/forecasting_model.py in historical_forecasts(self, series, past_covariates, future_covariates, num_samples, start, forecast_horizon, stride, retrain, overlap_end, last_points_only, verbose)
    443                         step=1,
    444                     ),
--> 445                     np.array(last_points_values),
    446                 )
    447 

/pdrive/dt-env/lib/python3.7/site-packages/darts/timeseries.py in from_times_and_values(cls, times, values, fill_missing_dates, freq, columns, fillna_value)
    604             coords[DIMS[1]] = columns
    605 
--> 606         xa = xr.DataArray(values, dims=(times_name,) + DIMS[-2:], coords=coords)
    607 
    608         return cls.from_xarray(

/pdrive/dt-env/lib/python3.7/site-packages/xarray/core/dataarray.py in __init__(self, data, coords, dims, name, attrs, indexes, fastpath)
    404             data = _check_data_shape(data, coords, dims)
    405             data = as_compatible_data(data)
--> 406             coords, dims = _infer_coords_and_dims(data.shape, coords, dims)
    407             variable = Variable(dims, data, attrs, fastpath=True)
    408             indexes = dict(

/pdrive/dt-env/lib/python3.7/site-packages/xarray/core/dataarray.py in _infer_coords_and_dims(shape, coords, dims)
    153             if s != sizes[d]:
    154                 raise ValueError(
--> 155                     f"conflicting sizes for dimension {d!r}: "
    156                     f"length {sizes[d]} on the data but length {s} on "
    157                     f"coordinate {k!r}"

ValueError: conflicting sizes for dimension 'time': length 972 on the data but length 4856 on coordinate 'time'