search

snwagh / falcon-public

Implementation of protocols in Falcon
89 stars 45 forks source link

How to load CIFAR-10 and ImageNet dataset? #42

Open llCurious opened 2 years ago

llCurious commented 2 years ago

Hey @snwagh .

I try to test the AlexNet and VGG16 which takes CIFAR-10 and ImageNet datasets as training data. Is there data loader for these two datasets?

And i notice that the input size of CIFAR-10 for AlexNet is modified to 33x33, which is here. I wonder why this modification is made and how does this affect the dataloader, zero padding the raw input from 32x32 to 33x33?

llCurious commented 2 years ago

I checked this blog . Currently, (32+29-11)/4 = 9.75 and (33+29-11)/4 = 10. Therefore, the input shall be resizes to 33*33?

I think for CIFAR-10 over ALexNet, the model architecture shall be modified? The first CNNLayer should use filter_size=7, strides=2, padding=2?

The model architecture for AlexNet is different from CryptGPU. For example, the second CNNLayer is 5596 as input, while in CryptGPU, it is 4496. Is there any mistakes here?

snwagh commented 2 years ago

I haven't written an dataset loader for CIFAR-10 or ImageNet but I would imagine it is not too hard to write one (using either the C++ reference code for MNIST or using a few lines of PyTorch/TensorFlow).

You're right about the reason for the input size to be set to 33. That is because C++ requires sizes to match up precisely and when I was looking at references to AlexNet, that was the best way I could reconcile this gap.

There are a number of references for the head architecture of AlexNet when working with CIFAR-10. As far as I understand there is no right or wrong architecture. The different head architectures do make a bit of a difference from the ML side (change the overall training/inference accuracy) but for the focus of this project a certain choice was made.

CryptGPU is a more recent work and thus it is reasonable to say that their choice might be more thought out.

llCurious commented 2 years ago

Yep, i have implemented the data loader for CIFAR-10 and ImageNet similar to MNIST.

And you are right that the model architecture does not influence the MPC side efficiency. I try to obtain reasonable accuracy for the private training of AlexNet over CIFAR10. Hence i have to figure out the reason for the mismatch of input shape and make sure the training over the given architecture can converge.

Currently, I have modified the code and model architecture according to the impl of torch, so that the AlexNet model load CIFAR-10 with 32x32x3 and ImageNet with 64x64x3. And there is no need to change input data shape. Thanks anyway!!