tft-torch
tft-torch copied to clipboard
PlaytikaOSS
A Python library that implements ״Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting״
Change log: 1. Allow usage of torch's built in attention implementation when attention scores are not required.
Change log: 1. Removed class arguments in TFT building blocks that use count of categorical features when cardinalities are available. Instead the amount is inferred from the number of carinalities,
The parameter `num_inputs` is redundant and can be inferred directly from the `cardinalities` parameter,
Consider wrapping the call to self.attention in InterpretableMultiHeadAttention with `with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=True):` In order to improve speed and memory efficiency.
You have mentioned in www.playtika-blog.com/playtika-ai/multi-horizon-forecasting-using-temporal-fusion-transformers-a-comprehensive-overview-part-2/, that "The different heads simply take care of the interactions between the Queries and the Keys, and the outputs of the heads are aggregated and...
Metadata
Owner
Metadata
A Python library that implements ״Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting״