search

aicaffeinelife / Pytorch-STN

Spatial Transformer Networks in Pytorch.
MIT License
155 stars 33 forks source link

can't reproduce cifar10 #1

Open twmht opened 5 years ago

twmht commented 5 years ago

Can't reproduce cifar10 result.

Here is mine

without stn: 78.3% with stn: 76.7%

Any idea?

aicaffeinelife commented 5 years ago

What optimizer are you using?

twmht commented 5 years ago

sgd

twmht commented 5 years ago

Does this code can reproduce the result?

aicaffeinelife commented 5 years ago

yeah, this code uses Adam optimizer. Can you try using that? If not I'll take a closer look again

Thanks

twmht commented 5 years ago

@aicaffeinelife

I use adam and still can't repoduce the result

without stn: 78.4 with stn: 76.1

aicaffeinelife commented 5 years ago

Interesting, I'm traveling at the moment, I'll check the code again and update you about it in a couple of days. In the mean time, if you find any bug in the code, let me know.

world2025 commented 5 years ago

hello,when the epoch=82,the value is 77.3,after epoch=82,the value is very low,i don't know how to interpret it,can you help me ?ths

world2025 commented 5 years ago

can you teach me how to use the STN with vgg16 or resnet50?ths

aicaffeinelife commented 5 years ago

@tkone2018 open a separate issue for this one. Using STN in any CNN is easy, but it's not relevant to this thread.

aicaffeinelife commented 5 years ago

@twmht I just cloned the repo and re-ran experiments on the hardware mentioned in README.md. I got 78.19% accuracy. What hardware were you using?

world2025 commented 5 years ago

@aicaffeinelife ok

twmht commented 5 years ago

@aicaffeinelife

GTX-1080

aicaffeinelife commented 5 years ago

@twmht that is strange. Can you share your params.json here?