Hello,
I encountered some issues while using the DaCapo benchmarks:
Currently, all the benchmarks adopt the SiLU+avgpooling scheme. Although DaCapo provides the model parameters for the ReLU+maxpooling scheme, I still failed to use the ReLU+maxpooling scheme. Could you guide me on how to adjust the code to correctly implement the ReLU+maxpooling scheme?
At present, DaCapo uses FHE parameters with a polynomial degree of 217. If I want to use FHE parameters with a polynomial degree of 216, how should I modify the code? Is it sufficient to only change the value of "nt"?
When modifying AlexNet, I made the following attempts:
model_dict = torch.load(str(source_dir)+"/../data/alexNet_silu_avgpool_model", map_location=torch.device('cpu'))
->
model_dict = torch.load(str(source_dir)+"/../data/alexNet_relu_maxpool_model", map_location=torch.device('cpu'))
return HE_SiLU(x)
->
return HE_ReLU(x)
return HE_Avg(close, x)
->
return HE_MaxPad(close, x)
"nt" : 216
->
"nt" : 215
I performed simulations with plaintext. However, I found that after the first MaxPooling operation, the plaintext data started to deviate from the PyTorch execution results.
Hello,
I encountered some issues while using the DaCapo benchmarks:
Currently, all the benchmarks adopt the SiLU+avgpooling scheme. Although DaCapo provides the model parameters for the ReLU+maxpooling scheme, I still failed to use the ReLU+maxpooling scheme. Could you guide me on how to adjust the code to correctly implement the ReLU+maxpooling scheme?
At present, DaCapo uses FHE parameters with a polynomial degree of 217. If I want to use FHE parameters with a polynomial degree of 216, how should I modify the code? Is it sufficient to only change the value of "nt"?
When modifying AlexNet, I made the following attempts:
model_dict = torch.load(str(source_dir)+"/../data/alexNet_silu_avgpool_model", map_location=torch.device('cpu'))
->
model_dict = torch.load(str(source_dir)+"/../data/alexNet_relu_maxpool_model", map_location=torch.device('cpu'))
return HE_SiLU(x)
->
return HE_ReLU(x)
return HE_Avg(close, x)
->
return HE_MaxPad(close, x)
"nt" : 216
->
"nt" : 215
I performed simulations with plaintext. However, I found that after the first MaxPooling operation, the plaintext data started to deviate from the PyTorch execution results.