When predicting mesh-xlinear with only_topk the value of 20 is always used, regardless of which value is set up. Therefore we can only output 20 predictions for each datapoint, which is problematic in high recall set-ups.
The reason for that is because this parameter should be passed to prediction.
When predicting
mesh-xlinearwithonly_topkthe value of 20 is always used, regardless of which value is set up. Therefore we can only output 20 predictions for each datapoint, which is problematic in high recall set-ups.The reason for that is because this parameter should be passed to prediction.