sunwei925
I received negative scores in the results.csv from the test. Is this expected?
I received negative scores in the results.csv from the test. Is this expected? What is the reason, please results_384p.csv Some of the contents are as follows: filename,score test/0001.mp4,-1.0951594911515712 test/0002.mp4,-1.145484207570553 test/0003.mp4,-1.1424250453710556 test/0004.mp4,-1.1390137687325477 test/0005.mp4,-1.1351750060915946 test/0006.mp4,-1.1392134383320809 test/0007.mp4,-1.1463412337005139 test/0008.mp4,-0.9099979531019926 test/0009.mp4,-0.8705645047128201 test/0010.mp4,-0.862069771438837 test/0011.mp4,-0.8733548872172833 test/0012.mp4,-0.9043583035469055 test/0013.mp4,-0.9076500739902258 test/0014.mp4,-0.9421681996434927 test/0015.mp4,-1.1354316420853139 test/0016.mp4,-1.139632887765765 test/0017.mp4,-1.1341129895299673 test/0018.mp4,-1.1247829228639603 test/0019.mp4,-1.1306168153882026 test/0020.mp4,-1.131969215720892 test/0021.mp4,-1.1665853392332792 test/0022.mp4,-1.0058448605239392 test/0023.mp4,-0.9575022861361504 test/0024.mp4,-0.9606049969792366 test/0025.mp4,-0.9736806564033031 test/0026.mp4,-0.9875168107450009 test/0027.mp4,-0.9978467017412186 test/0028.mp4,-1.0311993107199668 test/0029.mp4,-0.8786280214786529 test/0030.mp4,-0.8837987545877695 test/0031.mp4,-0.8964707858860492 test/0032.mp4,-0.8951154794543982 test/0033.mp4,-0.923457158729434 test/0034.mp4,-0.9513952970504761 test/0035.mp4,-0.9861461881548166 test/0036.mp4,-1.1022858384996652 test/0037.mp4,-1.1096711702644826 test/0038.mp4,-1.117858549952507 test/0039.mp4,-1.1239035919308662 test/0040.mp4,-1.1206100273877382 test/0041.mp4,-1.1333493940532207 test/0042.mp4,-1.1466195002198218
It is right. Since we only use PLCC loss to optimize the model, the trained model will aim to achieve a high PLCC value between the model outputs and the ground-truth scores. This approach does not ensure that individual video scores will closely match their respective ground-truth values. To make the model scores more understandable, you can use the four-parameter logistic fitting function [y_output_logistic = fit_function(y_label, y_output) in util.py] to map the model scores to the scale of the ground-truth scores.
这是正确的。由于我们仅使用PLCC损失来优化模型，因此训练的模型将旨在实现模型输出和地面实况分数之间的高PLCC值。这种方法并不能确保各个视频分数与其各自的地面实况值紧密匹配。为了使模型分数更易于理解，您可以使用四参数逻辑拟合函数 [y_output_logistic = fit_function（y_label， y_output） in util.py] 将模型分数映射到真值分数的尺度。
Hello author, since I'm using it for inference and I don't actually have y_label, how should I use fit_function(y_label, y_output)?