How many training iterations needed?

[ahtchow]

Just wondering how many training iterations are needed for respectable performance on average for single class? The max-iterations is set to inf, so I assume we run it until satisfaction. Do I look at the loss value?

[GunnerStone]

The publication does not specify the number of training iterations required to achieve the reported performance results. This omission complicates direct comparisons between their original method and subsequent improvements.

For an improvement paper, I'm deciding to make a trade-off between model performance and training time. Training the model for 30 minutes achieves good enough proof of concept results. To double that model's accuracy on validation data, it needs to train for 4 days. I have around 50 classes, and I'm not planning on training this thing for 200 days.

The question is, did the authors train for 4 days or 30 minutes? We'll never know.