what is ops_thresh and how do you arrive at the hard coded value for each model?

[fanhui3]

I understand the predicted values before applying op_norm returns probabilities in very low decimal points, sometimes as low as 0.00XXX. The ops_threshs is there to normalise the value between 0 and 1. What is the approach used to attained the selected ops_thresh?

I am currently doing finetuning on the model. Thank you.

p.s. Thank you for releasing such useful work to the world, by the way.

[ieee8023]

The ops_threshs is there to normalise the value between 0 and 1.

No, the ops_thresh is to normalize the values so that 0.5 is the decision boundary. So values above 0.5 can be considered positive and less than 0.5 considered negative. See Section 2.5 Model Calibration here: https://arxiv.org/pdf/2111.00595.pdf

Also check out Section 4 Transfer Learning Example where "model.op_threshs = None" to remove this calibration so it doesn't get in the way of fine tuning.

You can see how the thresholds are computed here:

torchxrayvision/scripts/model_calibrate.py

Lines 198 to 202 in f4505b8

	all_outputs = 1.0/(1.0 + np.exp(-results[2][i]))
	fpr, tpr, thres = sklearn.metrics.roc_curve(results[3][i], all_outputs)
	pente = tpr - fpr
	opt_thres = thres[np.argmax(pente)]

p.s. Thank you for releasing such useful work to the world, by the way.

No problem!