no attribute 'HeaanContext'

[gillesoldano]

Hi

I'm following the XGBoost notebook and trying to reproduce the same results locally, the notebook on the docker image runs without any issue, however on my machine I get this error:
AttributeError: module 'pyhelayers' has no attribute 'HeaanContext'
This happens when running the line:
he_run_req.set_he_context_options([pyhelayers.HeaanContext()])

I'm using the last pyhelayers version which is 1.5.1.0 and I pulled the latest docker pylab image.

[dubek]

Thank you @stealthBanana for trying HElayers.

Indeed the pyhelayers version that is included inside the helayers-pylab docker image does include the HEaaN backend, but the pyhelayers released on PyPi (which you install with pip) doesn't yet include the HEaaN backend. We're working a release of the Python package that will include it.

[dubek]

Hi @stealthBanana , pyhelayers 1.5.2.0 was released yesterday. It should have HEaaN support built-in.
Can you try it? (pip install --upgrade pyhelayers)

[gillesoldano]

Hi, I'm trying it right now, however I encoutered another problem, when following the deep neural network example. Since I do not have access to the resnet18 model of the docker container, I tried to export a torchvision resnet18 myself as a .onnx, when doing nnp.inport_from_files I get the following error:

RuntimeError: Neural network architecture initialization from ONNX failed: Neural network initialization from ONNX encountered an operator type that is currently not supported: Relu

How could I retrieve that specific resnet implementation ? Is it pretrained on imagenet ? And could I apply transfer learning to it and fine tune it ?

[dubek]

FHE models don't support operations like ReLu which cannot be directly implemented in FHE.

The FHE-friendly resnet18 model that our notebook uses is generated by the script below. It has random weights -- just to demonstrate the timing of running the model on encrypted inputs.

#!/usr/bin/env python3
import torch
import torch.nn as nn
import os
import h5py


# define model
class MyNeuralNet(nn.Module):
    def __init__(self):
        super(MyNeuralNet, self).__init__()

        self.conv_0 = nn.Conv2d(in_channels=3, out_channels=64, kernel_size=7, stride=2, padding=3)
        self.bn_1 = nn.BatchNorm2d(num_features=64)
        self.pool_3 = nn.AvgPool2d(kernel_size=3, stride=2, padding=1)

        self.conv_4 = nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=1, padding=1)
        self.bn_5 = nn.BatchNorm2d(num_features=64)
        self.conv_7 = nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=1, padding=1)
        self.bn_8 = nn.BatchNorm2d(num_features=64)

        self.conv_11 = nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=1, padding=1)
        self.bn_12 = nn.BatchNorm2d(num_features=64)
        self.conv_14 = nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=1, padding=1)
        self.bn_15 = nn.BatchNorm2d(num_features=64)

        self.conv_18 = nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=2, padding=1)
        self.bn_19 = nn.BatchNorm2d(num_features=128)
        self.conv_21 = nn.Conv2d(in_channels=128, out_channels=128, kernel_size=3, stride=1, padding=1)
        self.bn_22 = nn.BatchNorm2d(num_features=128)
        self.conv_23 = nn.Conv2d(in_channels=64, out_channels=128, kernel_size=1, stride=2, padding=0)
        self.bn_24 = nn.BatchNorm2d(num_features=128)

        self.conv_27 = nn.Conv2d(in_channels=128, out_channels=128, kernel_size=3, stride=1, padding=1)
        self.bn_28 = nn.BatchNorm2d(num_features=128)
        self.conv_30 = nn.Conv2d(in_channels=128, out_channels=128, kernel_size=3, stride=1, padding=1)
        self.bn_31 = nn.BatchNorm2d(num_features=128)

        self.conv_34 = nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, stride=2, padding=1)
        self.bn_35 = nn.BatchNorm2d(num_features=256)
        self.conv_37 = nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, stride=1, padding=1)
        self.bn_38 = nn.BatchNorm2d(num_features=256)
        self.conv_39 = nn.Conv2d(in_channels=128, out_channels=256, kernel_size=1, stride=2, padding=0)
        self.bn_40 = nn.BatchNorm2d(num_features=256)

        self.conv_43 = nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, stride=1, padding=1)
        self.bn_44 = nn.BatchNorm2d(num_features=256)
        self.conv_46 = nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, stride=1, padding=1)
        self.bn_47 = nn.BatchNorm2d(num_features=256)

        self.conv_50 = nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, stride=2, padding=1)
        self.bn_51 = nn.BatchNorm2d(num_features=512)
        self.conv_53 = nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1)
        self.bn_54 = nn.BatchNorm2d(num_features=512)
        self.conv_55 = nn.Conv2d(in_channels=256, out_channels=512, kernel_size=1, stride=2, padding=0)
        self.bn_56 = nn.BatchNorm2d(num_features=512)

        self.conv_59 = nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1)
        self.bn_60 = nn.BatchNorm2d(num_features=512)
        self.conv_62 = nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1)
        self.bn_63 = nn.BatchNorm2d(num_features=512)

        self.pool_66 = nn.AvgPool2d(kernel_size=7, stride=1, padding=0)

        self.fc_68 = nn.Linear(512, 1000)

    def forward(self, x):

        x = self.conv_0(x)
        x = self.bn_1(x)
        x = torch.square(x)
        x = self.pool_3(x)

        b1 = self.conv_4(x)
        b1 = self.bn_5(b1)
        b1 = torch.square(b1)
        b1 = self.conv_7(b1)
        b1 = self.bn_8(b1)

        x = torch.add(x, b1)
        x = torch.square(x)

        b1 = self.conv_11(x)
        b1 = self.bn_12(b1)
        b1 = torch.square(b1)
        b1 = self.conv_14(b1)
        b1 = self.bn_15(b1)

        x = torch.add(x, b1)
        x = torch.square(x)

        b1 = self.conv_18(x)
        b1 = self.bn_19(b1)
        b1 = torch.square(b1)
        b1 = self.conv_21(b1)
        b1 = self.bn_22(b1)

        b2 = self.conv_23(x)
        b2 = self.bn_24(b2)

        x = torch.add(b1, b2)
        x = torch.square(x)

        b1 = self.conv_27(x)
        b1 = self.bn_28(b1)
        b1 = torch.square(b1)
        b1 = self.conv_30(b1)
        b1 = self.bn_31(b1)

        x = torch.add(x, b1)
        x = torch.square(x)

        b1 = self.conv_34(x)
        b1 = self.bn_35(b1)
        b1 = torch.square(b1)
        b1 = self.conv_37(b1)
        b1 = self.bn_38(b1)

        b2 = self.conv_39(x)
        b2 = self.bn_40(b2)

        x = torch.add(b1, b2)
        x = torch.square(x)

        b1 = self.conv_43(x)
        b1 = self.bn_44(b1)
        b1 = torch.square(b1)
        b1 = self.conv_46(b1)
        b1 = self.bn_47(b1)

        x = torch.add(x, b1)
        x = torch.square(x)

        b1 = self.conv_50(x)
        b1 = self.bn_51(b1)
        b1 = torch.square(b1)
        b1 = self.conv_53(b1)
        b1 = self.bn_54(b1)

        b2 = self.conv_55(x)
        b2 = self.bn_56(b2)

        x = torch.add(b1, b2)
        x = torch.square(x)

        b1 = self.conv_59(x)
        b1 = self.bn_60(b1)
        b1 = torch.square(b1)
        b1 = self.conv_62(b1)
        b1 = self.bn_63(b1)

        x = torch.add(x, b1)
        x = torch.square(x)

        x = self.pool_66(x)
        x = torch.flatten(x, start_dim=1)
        x = self.fc_68(x)

        return x

# define input
input = torch.randn(1, 3, 224, 224) # B,C,X,Y

# run inference
my_nn = MyNeuralNet()
my_nn.eval()

# save model
torch.onnx.export(my_nn, input, "model.onnx", training=torch.onnx.TrainingMode.PRESERVE, export_params=False)
print("Saved model")