Exgoneous features in Automodels

[kkckk1110]

Description

Hello! I am trying to use Automodels to automatically tune the parameters. However, I found the document confusing because I do not know how to specify the exogenous features in Automodels. I tried to specify in config as follows, but I have no idea whether I did it right. I also wonder what will happen if I add exogenous features in the config. Will the model use all the features, or, like tuning other parameters, selectively incorporating some features? I am looking forward to responses and I think a more detailed tutorial on how to combine Automodels and exogenous will be really helpful! Thanks!

config_nhits = {
"input_size": tune.choice([6, 62, 63]), # Length of input window
"start_padding_enabled": True,
"n_blocks": 5*[1], # Length of input window
"mlp_units": 5 * [[64, 64]], # Length of input window
"n_pool_kernel_size": tune.choice([5*[1], 5*[2], 5*[4],
[8, 4, 2, 1, 1]]), # MaxPooling Kernel size
"n_freq_downsample": tune.choice([[8, 4, 2, 1, 1],
[1, 1, 1, 1, 1]]), # Interpolation expressivity ratios
"learning_rate": tune.loguniform(1e-4, 1e-2), # Initial Learning rate
"scaler_type": tune.choice([None]), # Scaler type
"max_steps": tune.choice([1000]), # Max number of training iterations
"batch_size": tune.choice([1, 4, 10]), # Number of series in batch
"windows_batch_size": tune.choice([128, 256, 512]), # Number of windows in batch
"random_seed": tune.randint(1, 20), # Random seed
"futr_exog_list":futr_exog_list
}

Link

No response

[jmoralez]

Hey @kkckk1110, thanks for using neuralforecast. The exogenous features are used in the same way as the other hyperparameters, you can set them to a fixed value e.g. futr_exog_list: ['a', 'b'] or tune them as well, e.g. futr_exog_list: tune.choice([['a'], ['b'], ['a', 'b']])

[vidarsumo]

I tried this using hist_exog_list:

nhits_config = {
       "max_steps": 100,                                                         # Number of SGD steps
       "input_size": 24,                                                         # Size of input window
       "learning_rate": tune.loguniform(1e-5, 1e-1),                             # Initial Learning rate
       "n_pool_kernel_size": tune.choice([[2, 2, 2], [16, 8, 1]]),               # MaxPool's Kernelsize
       "n_freq_downsample": tune.choice([[168, 24, 1], [24, 12, 1], [1, 1, 1]]), # Interpolation expressivity ratios
       "val_check_steps": 50,                                                    # Compute validation every 50 steps
       "random_seed": tune.randint(1, 10),                                       # Random seed
       "hist_exog_list": ['feat_1', 'feat_2', 'feat_3', 'feat_4', 'feat_5', 'feat_6']
    }
    
    model = AutoNHITS(h=21,
                  config=nhits_config,
                  search_alg=HyperOptSearch(),
                  backend='ray',
                  num_samples=3)

nf = NeuralForecast(models=[model], freq='D')
nf.fit(df=df_train, val_size=21)

where feat_x are the names of the historical exogen features. I keep getting this error:

set(temporal_cols.tolist()) & set(self.hist_exog_list + self.futr_exog_list)
TypeError: can only concatenate tuple (not "list") to tuple

I could get this to work by also passing "futr_exog_list": [] to nhits_config.

[jmoralez]

That's a bug with the HyperOptSearch, which was fixed in #851 but hasn't made it into a release. If you can install from github it should work.

[noahvand]

I got the simular problem when using TFT models, and get error

"set(temporal_cols.tolist()) & set(self.hist_exog_list + self.futr_exog_list)
TypeError: can only concatenate tuple (not "list") to tuple"

and I fixed it by have modify the code in _base_model.py as below:

def _get_temporal_exogenous_cols(self, temporal_cols):
        self.hist_exog_list = list(self.hist_exog_list) # this is added 
        self.futr_exog_list = list(self.futr_exog_list) # this is added 
        return list(
            set(temporal_cols.tolist()) & set(self.hist_exog_list + self.futr_exog_list)