TimeGrad Notebook version 0.7.0 -> predicts all nans

[StefanStanisor]

Hello, I am writing my master thesis and while trying to make timegrad work for at least an example, I tried the example in the notebook. I looked at all the other issues people mentioned, but I think I got a new one.

The predictor outputs all nans. This is the code that I am using:

device = torch.device('mps')

dataset = get_dataset("electricity_nips", regenerate=False)
train_grouper = MultivariateGrouper(max_target_dim=min(2000, int(dataset.metadata.feat_static_cat[0].cardinality)))

test_grouper = MultivariateGrouper(num_test_dates=int(len(dataset.test)/len(dataset.train)),
                                   max_target_dim=min(2000, int(dataset.metadata.feat_static_cat[0].cardinality)))

dataset_train = train_grouper(dataset.train)
dataset_test = test_grouper(dataset.test)
scheduler = DEISMultistepScheduler(
    num_train_timesteps=150,
    beta_end=0.1,
)
estimator = TimeGradEstimator(
    input_size=int(dataset.metadata.feat_static_cat[0].cardinality),
    hidden_size=64,
    num_layers=2,
    dropout_rate=0.1,
    lags_seq=[1],
    scheduler=scheduler,
    num_inference_steps=150,
    prediction_length=dataset.metadata.prediction_length,
    context_length=dataset.metadata.prediction_length,
    freq=dataset.metadata.freq,
    scaling="mean",
    trainer_kwargs=dict(max_epochs=1, accelerator="gpu", devices="1"),
)

predictor = estimator.train(dataset_train)

s = time.time()
forecast_it, ts_it = make_evaluation_predictions(dataset=dataset_test,
                                                 predictor=predictor,
                                                 num_samples=100)

forecasts = list(forecast_it)
targets = list(ts_it)

[fbehnia]

I am trying to run the the Time-Grad-Electricity notebook, and from the following Iine I get an error.
line:
dataset_test = test_grouper(dataset.test)

error:
ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (2590,) + inhomogeneous part.

Do you know how I can fix this issue?

[Wjie-Z]

I am trying to run the the Time-Grad-Electricity notebook, and from the following Iine I get an error.
line:
dataset_test = test_grouper(dataset.test)

error:
ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (2590,) + inhomogeneous part.
Have you solved it?

[fbehnia]

I am trying to run the the Time-Grad-Electricity notebook, and from the following Iine I get an error. line: dataset_test = test_grouper(dataset.test)

error: ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (2590,) + inhomogeneous part. Have you solved it?

I had the same issue I used version 0.7.0 and edit the notebook based on #152

[fbehnia]

Hello, I am writing my master thesis and while trying to make timegrad work for at least an example, I tried the example in the notebook. I looked at all the other issues people mentioned, but I think I got a new one.

The predictor outputs all nans. This is the code that I am using:

device = torch.device('mps')

dataset = get_dataset("electricity_nips", regenerate=False)
train_grouper = MultivariateGrouper(max_target_dim=min(2000, int(dataset.metadata.feat_static_cat[0].cardinality)))

test_grouper = MultivariateGrouper(num_test_dates=int(len(dataset.test)/len(dataset.train)),
                                   max_target_dim=min(2000, int(dataset.metadata.feat_static_cat[0].cardinality)))

dataset_train = train_grouper(dataset.train)
dataset_test = test_grouper(dataset.test)
scheduler = DEISMultistepScheduler(
    num_train_timesteps=150,
    beta_end=0.1,
)
estimator = TimeGradEstimator(
    input_size=int(dataset.metadata.feat_static_cat[0].cardinality),
    hidden_size=64,
    num_layers=2,
    dropout_rate=0.1,
    lags_seq=[1],
    scheduler=scheduler,
    num_inference_steps=150,
    prediction_length=dataset.metadata.prediction_length,
    context_length=dataset.metadata.prediction_length,
    freq=dataset.metadata.freq,
    scaling="mean",
    trainer_kwargs=dict(max_epochs=1, accelerator="gpu", devices="1"),
)

predictor = estimator.train(dataset_train)

s = time.time()
forecast_it, ts_it = make_evaluation_predictions(dataset=dataset_test,
                                                 predictor=predictor,
                                                 num_samples=100)

forecasts = list(forecast_it)
targets = list(ts_it)

I am also getting all NaNs did you solve it?

[2ySong]

I also got the all NaNs, I can't solve it.

[jiajinghu19]

Did anyone solve the all NaN's problem? I also got all the NaNs

[2ySong]

you can set the num_inference_steps=149 that will solve the problem.

[jiajinghu19]

you can set the num_inference_steps=149 that will solve the problem.

Thank you! It worked!