Validation Accuracy Much Lower than Instructors
Hey Team,
My accuracy is not nearly as good as the instructor's. Has anyone else had this behavior? At first, I thought I made a mistake. But then I ran the code provided with the course and got the same behavior. Can anyone explain why that is?
Additionally, I noticed that the loss is not decreasing consistently. By my last epoch, the loss has actually increased past that of the 1st epoch. This is quite odd to me given what I saw in the lecture.
Can anyone explain if the tensorflow algorithms have changed over the years, which is leading to this discrepency? Thank you.
My accuracy is not nearly as good as the instructor's. Has anyone else had this behavior? At first, I thought I made a mistake. But then I ran the code provided with the course and got the same behavior. Can anyone explain why that is?
Additionally, I noticed that the loss is not decreasing consistently. By my last epoch, the loss has actually increased past that of the 1st epoch. This is quite odd to me given what I saw in the lecture.
Can anyone explain if the tensorflow algorithms have changed over the years, which is leading to this discrepency? Thank you.
1 answers ( 0 marked as helpful)
Okay, so the reason why is because of running the calculations on the M1 GPU. Apparently there is a bug in the tensorflow-metal plugin regarding the Adam optimizer that causes the gradien to explode. You need to run this on the CPU instead.
Can anyone with some deep knowledge of this shed some light on the stable version of tensorflow, tensorflow-metal, and python to use here to run this stuff on the GPU? Once we get to CNN's we won't be able to run this on the CPU any more because it will be too slow.
I'd really appreciate some instructor help here. It seems as though this entire course forum is dead and no one is supporting it any longer. I would really appreciate some feedback here.
-Justin
Can anyone with some deep knowledge of this shed some light on the stable version of tensorflow, tensorflow-metal, and python to use here to run this stuff on the GPU? Once we get to CNN's we won't be able to run this on the CPU any more because it will be too slow.
I'd really appreciate some instructor help here. It seems as though this entire course forum is dead and no one is supporting it any longer. I would really appreciate some feedback here.
-Justin
