Reproducibility issue in TimeGrad with ver-0.7.0

[nonconvexopt]

I installed pytorch-ts by git clone and branched to ver-0.7.0.
I debugged all the issues in using timegrad model on electricity dataset, resolving differences in using diffusers instead self-implemented diffusion models.
However, the train loss (or validation loss too) does not record around 0.07 reported at the timegrad-electricity example. I get 0.2 mininum even with known hyperparameter setting. Even if I tune the hyperparameters extensively, I get almost similar result (increased the number of training steps (diffusion steps), tuned learing rate, and so on).
I assume there are some issues in adapting diffusers library currently.

Can you update the timegrad-electricity example with ver-0.7.0?

[nonconvexopt]

It looks like that the predictions are gathered around 0

[nonconvexopt]

I applied DEISMultistepScheduler following the #145,
However, valid loss does no go down under 0.3.
@ProRedCat Can you share the details of your setting? Below is my code:

estimator = TimeGradEstimator(
    num_layers = 2,
    hidden_size = 40,
    lr=1e-3,
    weight_decay = 0,
    batch_size=32,
    num_batches_per_epoch=100,
    prediction_length=elec_dataset.metadata.prediction_length,
    context_length=elec_dataset.metadata.prediction_length,
    input_size=370,
    freq=elec_dataset.metadata.freq,
    scaling=True,
    scheduler = DEISMultistepScheduler(
        num_train_timesteps=150,
        beta_end=0.1,
    ),
    trainer_kwargs={
        "max_epochs": 200,
    }
)

[ProRedCat]

Here's my parameters, the loss gets to 0.277 but the paper does not use loss for its evaluation, they use CPRS-Sum and I get CRPS-Sum: 0.018305275885916936 for this model. The code for that should be at the bottom of the example.

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=200, accelerator="gpu", devices="1"),
)

[ProRedCat]

For reference here is the graphs from that model

[nonconvexopt]

Your result looks completely fine. Thank you for sharing. I will try to reproduce it.

[coding-loong]

Hi, I would like to ask if you know how to set the right hyperparameters of TimeGrad on the solar and Wikipedia datasets to get results consistent with the paper.

[nonconvexopt]

@ProRedCat If we use DEISMultistepScheduler, does it mean that we are using a little bit advanced version of ScoreGrad: Multivariate Probabilistic Time Series Forecasting with Continuous Energy-based Generative Models?

[kashif]

sorry folks i am traveling this week... I will try to have a look next week