PFLD-pytorch
PFLD-pytorch copied to clipboard
polarisZhao
PFLD pytorch Implementation
How to use SetPreparation.py to test on 300w dataset?
Hello, I trained PFLD according to your code. The loss on the validation dataset is unchanged after about 200 epochs, and finally nme is 0.074. Why is this? Looking forward...
作者您好，谢谢您的工作，我这里有基于300w数据集的PFLD论文复现的问题： 1.300W数据集无六个姿态属性，这里就无法对齐使用这个数据训练，原论文未提到 2.严格按照论文说300w数据增强方法，人脸旋转角度-30°至30°每间隔5°旋转一次，并翻转，随机遮挡百分之20人脸区域（具体如何遮挡没说，我采用的是马赛克和灰色随机遮挡（根据RandomErasing论文）），但是发现训练损失正常下降而验证损失完全不降反而会升高 3.使用与不使用辅助网络（按照准备数据时就计算姿态角度作为标签），基本没有影响 请问关于这些问题有实验过吗，期待您的回复
The paper says that PFLD 1X and PFLD 0.25X has 12.5 Mb and 2.1 Mb, respectively. As far as I confirmed with `thop.profile` and `torchsummary.summary`, the number of parameters of...
 In each row the first tensor is pred and the second tensor is gt. Why are the Euler angles so different? Thanks!
``` def train(train_loader, pfld_backbone, auxiliarynet, criterion, optimizer, epoch): losses = AverageMeter() weighted_loss, loss = None, None for img, landmark_gt, attribute_gt, euler_angle_gt in train_loader: img = img.to(device) attribute_gt = attribute_gt.to(device) landmark_gt...
