Q: is the auto-saved model to be used instead of lag-llama.ckpt ?

time-series-foundation-models / lag-llama

Lag-Llama: Towards Foundation Models for Probabilistic Time Series Forecasting

Apache License 2.0

1.21k stars 147 forks source link

Q: is the auto-saved model to be used instead of lag-llama.ckpt ? #86

Open kenadianu opened 3 months ago

kenadianu commented 3 months ago

During fine-tuning the models are automatically saved, the best being, say \lag-llama-main\lightning_logs\version_12\checkpoints\epoch=34-step=5678.ckpt' (this is displayed on the console)

For subsequent prediction, with zero-shot, should I provide LagLlamaEstimator(...) with epoch=34-step=5678.ckpt instead of lag-llama.ckpt ? Thank you.

ashok-arjun commented 2 months ago

Hi, if you're using finetuning, then the evaluation is no longer "zero-shot". And yes, you should provide the right (latest) checkpoint if you want to evaluate results on that checkpoint.