search

opeltre / revert

Revert
MIT License
2 stars 0 forks source link

Conditional GANs for pulse generation #22

Open opeltre opened 2 years ago

opeltre commented 2 years ago

Implement a section of a BarlowTwins/VICReg pretrained encoder by a conditional GAN. 

x     : pulse ------fibration------------> E(x) : code
G(y') : pulse <-------section--------------- y' : code

In order to beat the discriminator D(x, y), the couple G(y), y must be consistent with the encoder fibration E, i.e. y =~ E(G(y)) while G(y) should also belong to the true data distribution of x.

N.B. when E : x -> y is constant this is like training a simple GAN.

opeltre commented 2 years ago

@guillaumebressan les twins sont réimplémentés ici: 5ba79c6 !

Je vais ajouter quelques tests avant de merge + adapter le script experiments/twins.py

La signature de la classe Twins est devenue [n_batch, 2, d_in] -> [n_batch, 2, d_out] pour garder la batch dimension en premier (et rester cohérent avec la classe Module)

opeltre commented 2 years ago

voir #27