search

xiwenc1 / TimeMIL

The code for the paper: TimeMIL: Advancing Multivariate Time Series Classification via a Time-aware Multiple Instance Learning
40 stars 5 forks source link

Request for Environment Details, Attention Map Visualization, and Fine-tuned Parameters for Reproducing Results #5

Closed Evan-Kun closed 4 weeks ago

Evan-Kun commented 2 months ago

Hi,

Your work is truly impressive and insightful, and I’m genuinely interested in it! I’ve successfully implemented the code, but I haven't been able to reproduce the exact same results for each dataset.

Could you please share the environment requirements as a reference? Additionally, if you could provide details on the attention map visualization part and the fine-tuned parameters for each dataset, such as batch size, that would be extremely helpful.

Thank you in advance!

Best regards,

xiwenc1 commented 1 month ago

Thank you for your interest in our work!

Given the large volume of dataset logs in multiple machines, it may take us a little time to organize them.

As noted in the README, we performed a grid search for each dataset, and the hyperparameter tuning strategy is explained there. You’re welcome to try different hyperparameters based on that. Due to the small size of datasets, you may take multiple times to obtain a good result. But don't worry, We will also try to upload the checkpoints after a critical upcoming deadline.

For your convenience, we’ve introduced a tool result_extractor.py that efficiently summarizes all experimental results into a single CSV file. This file contains both the results and hyperparameters for each experiment, with one experiment per row.

Evan-Kun commented 1 month ago

Thank you so much for your helpful response! I really appreciate the notes on fine-tuning the parameters, and I’ll definitely give it a try. The result_extractor.py script sounds super useful, so I’ll be sure to use that to streamline my analysis.

Good luck with your upcoming deadline! I’m looking forward to the checkpoints whenever they’re ready.

Thanks again for your support!

Best regards,

xiwenc1 commented 1 month ago

Hi,

we provide the weights here: Link It includes some possible good hyper-parameters for each datasets.

Thanks for checking it.