huggingface

/

time-series-transformer-tourism-monthly

like 17

Follow

Hugging Face 4,716

I'm attempting to modify the TimeSeriesTransformerForPrediction to accept my own multi variable dataset. My dataset contains ~70 runs. Each run contains a time series and three static environment variables that remain constant throughout the run and describe the conditions of the run. For example, one run may contain a time series, and the variables 1,250,5. Each time series contains values sampled every minute over 19 hours (1140 Samples).
I have taken the following steps to prepare the data following following Kashif's time series dataset example:

1. Create a list of dictionaries with each list item containing key value features for target, start, and feat_static_cat. For example, this is an item from the train list
   {'start': Timestamp('1970-01-01 00:00:00'), 'target': array([ 775.457207, 775.457207, 785.056306, ..., 1231.042793,
   1239.531532, 1239.598348]), 'feat_static_cat': [1, 250, 5]}
2. Define a feature schema and

features  = Features(
    {    
        "start": Value("timestamp[s]"),
        "target": Sequence(Value("float32")),
        "feat_static_cat": Sequence(Value("uint64")),
    }
)

3. Create training and test datasets from feature schema and list

train_dataset= Dataset.from_list(trainList,features)
test_dataset= Dataset.from_list(testList,features)

After preparing the dataset, I computed the cardinality of the static features

card= [len(depots.categories), len(tanks.categories), len(ships.categories)]

to get a list [5,3,5]

I then declare my transformer config as follows:

config = TimeSeriesTransformerConfig(
    prediction_length=prediction_length,
    context_length=30,
   lags_sequence=[1, 2, 3, 4,5,6,7],
    num_time_features=len(time_features) + 1, # we'll add 2 time features ("month of year" and "age", see further)
    num_static_categorical_features=3, # depot, shipnum, and tank size
    cardinality=card, 
    input_size=3,
    embedding_dimension=[1,1,1], 
    encoder_layers=4, 
    decoder_layers=4,
)

I then create a data loader and batch iterator:

train_dataloader = create_train_dataloader(
    config=config, 
    freq=freq, 
    data=train_dataset, 
    batch_size=256,
    num_batches_per_epoch=100,
)

batch = next(iter(train_dataloader))

When I run the train data loader, I get the following error:

 File "/usr/local/lib/python3.9/site-packages/gluonts/exceptions.py", line 95, in assert_gluonts
    raise exception_class(message.format(*args, **kwargs))
gluonts.exceptions.GluonTSDataError: Input for field "target" does not have the requireddimension (field: target, ndim observed: 1, expected ndim: 2)

I suspect the error is comping from when the data loader converts the target to an Numpy array.

            AsNumpyArray(
                field=FieldName.TARGET,
                # in the following line, we add 1 for the time dimension
                expected_ndim=1 if config.input_size==1 else 2,
            ),

When I change the data loader to expect n_dim=1, I get the following error

Batch Information:
static_categorical_features torch.Size([256, 3]) torch.LongTensor
static_real_features torch.Size([256, 1]) torch.FloatTensor
past_time_features torch.Size([256, 37, 6]) torch.FloatTensor
past_values torch.Size([256, 37]) torch.FloatTensor
past_observed_mask torch.Size([256, 37]) torch.FloatTensor
future_time_features torch.Size([256, 60, 6]) torch.FloatTensor
future_values torch.Size([256, 60]) torch.FloatTensor
future_observed_mask torch.Size([256, 60]) torch.FloatTensor
Error:
  File "/usr/local/lib/python3.9/site-packages/torch/nn/functional.py", line 2210, in embedding
    return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
IndexError: index out of range in self

Is there something I am not doing when adding static features to my time series?

Thanks