Temporal Fusion Transformer predict error

1

Hello everyone,

I’m currently working with the Temporal Fusion Transformer (TFT) and facing a dimension mismatch issue. My dataset contains ~5,000 unique combinations (e.g., SKU, store, etc.) over 24 timesteps (2 years of monthly data). My goal is to forecast the next month for each combination.

When I calculate metrics like the Mean Absolute Error (MAE) using Baseline().predict() with return_y=True, I encounter the following error:

RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 1280 but got size 641 for tensor number 4 in the list.

Interestingly, if I set return_y=False, the code runs without issues. Below is the setup I’ve been using:

# Create validation set (predict=True) to predict the last max_prediction_length points in time for each series  
validation = TimeSeriesDataSet.from_dataset(training, df, predict=True, stop_randomization=True)  

# Create dataloaders for the model  
batch_size = 128  # Set this between 32 to 128  
train_dataloader = training.to_dataloader(train=True, batch_size=batch_size, num_workers=0)  
val_dataloader = validation.to_dataloader(train=False, batch_size=batch_size * 10, num_workers=0, drop_last=True)  

# Define baseline predictions  
baseline_predictions = Baseline().predict(val_dataloader, return_y=True)  

# Calculate MAE  
MAE()(baseline_predictions.output, baseline_predictions.y)

Here are the parameters I’ve configured:

  • max_prediction_length = 2
  • max_encoder_length = 12

It appears to be a dimension mismatch issue, but I’m not sure how to address it. Has anyone encountered a similar issue or can suggest a solution?

Thank you in advance for your help!

2

Do you know what the expectations are from MAE() for the shape of its inputs?

You mention that the code runs without issues when return_y=False, would baseline_predictions.y still passed to MAE() in that case?