Can not get the same architecture with same random seed and setting

[dzk9528]

Hi. I am run the commands for darts benchamrk search cifar10, but I got different result. The only difference is that because I need to update pytoch version for my gpu is not 1080ti and pytorch updates their methods for computing eigenvalues in later version, I changed

TENAS/lib/procedures/ntk.py

Line 58 in 9df78ff

eigenvalues, _ = torch.symeig(ntk) # ascending

to torch.linalg.eigh, but the result geotype is differnt:

Genotype(normal=[('sep_conv_5x5', 0), ('avg_pool_3x3', 1), ('dil_conv_5x5', 0), ('sep_conv_3x3', 2), ('dil_conv_3x3', 0), ('avg_pool_3x3', 1), ('dil_conv_5x5', 1), ('sep_conv_5x5', 2)], normal_concat=[2, 3, 4, 5], reduce=[('sep_conv_3x3', 0), ('dil_conv_5x5', 1), ('dil_conv_3x3', 0), ('max_pool_3x3', 2), ('dil_conv_3x3', 0), ('sep_conv_5x5', 2), ('sep_conv_3x3', 2), ('dil_conv_5x5', 3)], reduce_concat=[2, 3, 4, 5])

which in DART_evaluation give me 96.79% test accuracy. Do you know the reason behind it?