search

hysts / pytorch_mpiigaze

An unofficial PyTorch implementation of MPIIGaze and MPIIFaceGaze
MIT License
346 stars 85 forks source link

Model choosing #36

Closed Kelly-ZH closed 3 years ago

Kelly-ZH commented 3 years ago

Hi, hysts I have a strange question, why do you choose model resnet_preact_8 for mpiigaze dataset and resnet_14 for mpiifacegaze? Besides, the cited paper used VGG16 for mpiigaze and got 5.4 mean angle error, used a CNN with spatial weights then got 4.8 angle error. Thank you very much.

Yours, Kelly

hysts commented 3 years ago

@Kelly-ZH

One reason was that I thought ResNet would perform better, and another reason was that AlexNet was too large to run on the CPU. As for the depth, I just stopped there because I wanted a fast model and increasing the depth further didn't improve performance much.

Kelly-ZH commented 3 years ago

@Kelly-ZH

One reason was that I thought ResNet would perform better, and another reason was that AlexNet was too large to run on the CPU. As for the depth, I just stopped there because I wanted a fast model and increasing the depth further didn't improve performance much.

So have you tried resnet_preact_10 or even more layers？ Thank you.

hysts commented 3 years ago

@Kelly-ZH

If you mean whether I've tried larger models, the answer is yes.