Hi, thank you for providing the version of PyTorch. But I find that at the beginning of the training, the accuracy of the three validation datasets was very low. (I use MobileFaceNet, ArcFace loss and train on CASIA-WebFace, and the validation accuracy on lfw at the beginning epoch of training only 70%).
Then I find the code img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR) in data_pipe.py transforms the RGB to BGR. But in PIL, the image has the mode RGB and transforms.ToTensor() can process PIL image directly. So I think this transform is useless.
And I add one line of code(carray = carray[:, ::-1, :, :].copy()) in function evaluate to change the BRG to RGB. The accuracy of validation at the beginning of the train is higher.
So I think maybe should not do the transform RGB to BGR.
Hi, thank you for providing the version of PyTorch. But I find that at the beginning of the training, the accuracy of the three validation datasets was very low. (I use
MobileFaceNet,ArcFaceloss and train onCASIA-WebFace, and the validation accuracy onlfwat the beginning epoch of training only70%).Then I find the code
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)indata_pipe.pytransforms theRGBtoBGR. But inPIL, the image has the modeRGBandtransforms.ToTensor()can processPIL imagedirectly. So I think this transform is useless.And I add one line of code(
carray = carray[:, ::-1, :, :].copy()) in functionevaluateto change theBRGtoRGB. The accuracy of validation at the beginning of the train is higher.So I think maybe should not do the transform
RGBtoBGR.