Regarding patch decoding using rVAE

[utkarshp1161]

Hi Max,

Richard(@RichardLiuCoding) and I tried below using regular vae and works great:

1. m, s = vae.encode(patches)

2. Choose a location in the latent space and then filter out patches
   with corresponding latent representation

   z = [-0.6,-0.4]

   Adjust this thres value to control how close these patches are w.r.t. the selected latent values
   thres = 0.25
   index = np.where(np.sqrt((m[:,0]-z[0])**2 + (m[:,1]-z[1])**2) < thres)

What I am trying now:

1. trained a rvae:
   rvae1 = pv.models.iVAE(in_dim, latent_dim=2, # Number of latent dimensions other than the invariancies
   hidden_dim_e = [256, 256], # corresponds to the number of neurons in the hidden layers of the encoder
   hidden_dim_d = [256, 256], # corresponds to the number of neurons in the hidden layers of the decoder
   invariances= ['r'], seed=0)

2. Now I want to choose a z (randomly) and decode it through network

Question. since I have a rvae with 3 latent dim[z1 is angle and z2,z3 are other latent variables] shall I give a z with 3 enteries(I tried and it throws an error)? or two enteries(if two enteries what does these enteries correspond to?) Essentially I want to draw samples from the latent space or rvae and decode it and see.

[ziatdinovmax]

@utkarshp1161 you should be able to pass "angle" as a keyword argument:

pyroVED/pyroved/models/base.py

Line 159 in 7807ffb

a = kwargs.get("angle", tt(0.)).to(device)

[utkarshp1161]

@ziatdinovmax thanks a lot. We were able to use it.