TankardYegga/sgr-code

loss

Opened this issue · 1 comments

你好,我这边跑的loss非常大Epoch 2 :training loss 51703894016.0000,[Epoch=2 iter=1] iter_loss is tensor(2.9257e+10, device='cuda:0', grad_fn=),每个类的精度很低甚至为0,score is{'Pixel Accuracy': 0.03415111488005151, 'Mean Accuracy': 0.07123155002253284, 'Frequency Weighted IoU': 0.008247191628445588, 'Mean IoU': 0.018683431210509416, 'Class IoU': {0: 0.0, 1: 0.0, 2: nan, 3: 0.03504468961664688, 4: nan, 5: nan, 6: 0.0, 7: 0.0, 8: 0.06110863047776998, 9: 0.0, 10: nan, 11: 0.0, 12: 0.0, 13: 0.0, 14: 0.05331412958965847, 15: 0.0, 16: nan, 17: 0.0, 18: nan, 19: 0.0, 20: nan}}。请问这种原因是什么呢?

不好意思,这个我也没有解决,我一直怀疑是fasttext的文本向量数据有问题,我当时也遇到这个问题了,一直没有解决。