Implementation of DIRT-T

[Solacex]

salad/salad/models/digits/dirtt.py

Line 38 in 5f55d6f

class SVHN_MNIST_Model(nn.Module):

I noticed that the architecture of DIRT-T in salad is different from the setting in the original paper:

including the number of channels in hidden layers and the activation function(the original paper is LeakyReLU)

[stes]

Hey @Solacex , yes you're right; actually all implementations currently using the same model, which is the one used in the Self-Ensembling paper by French et al.
This gives the SelfEnsemblingSolver potentially an unfair advantage when this architecture is used.

Would you be interested in implementing the DIRT-T model and submitting a PR? It would be interesting to see how results change when this model architecture is used.

[stes]

In case you're interested in implementing this, please directly change the code in https://github.com/domainadaptation/salad/blob/master/salad/models/digits/dirtt.py and call the resulting implementation SVHN_MNIST_DIRTTBaseline or similar.

Thanks a lot!

[Solacex]

Hello, @stes
I am trying implementing DIRT-T, I will submit my pytorch implementation as you indicated once finished.
And I am so appreciated you that make and public such a good library, it helps me a lot!