ziatdinovmax / pyroved Goto Github PK

Currently, the manifold_traversal() in joint iVAE supports only plotting the 2D images. It needs to be expanded to 1D spectral data since there's an increasing interest in applying joint iVAE to this type of data.

I am not sure which is best practice, since the way that you're handling imports does make sense. However, I think for readability, it might make more sense to precisely specify imports.

For example, consider the __init__.py in pyroved.utils,

from .coord import (generate_grid, generate_latent_grid,
                    generate_latent_grid_traversal, transform_coordinates)
from .data import init_dataloader, init_ssvae_dataloaders
from .nn import (get_activation, get_bnorm, get_conv, get_maxpool,
                 set_deterministic_mode, to_onehot)
from .prob import get_sampler
from .viz import plot_grid_traversal, plot_img_grid, plot_spect_grid

__all__ = ['generate_grid', 'transform_coordinates', 'generate_latent_grid',
           'get_sampler', 'init_dataloader', 'init_ssvae_dataloaders',
           'get_activation', 'get_bnorm', 'get_conv', 'get_maxpool',
           'to_onehot', 'set_deterministic_mode', 'get_sampler',
           'plot_img_grid', 'plot_spect_grid', 'plot_grid_traversal',
           'generate_latent_grid_traversal']

It too me a minute to figure out where set_deterministic_mode was when reading trvae.py,

from pyroved.utils import (
    generate_grid, generate_latent_grid, get_sampler,
    plot_img_grid, plot_spect_grid, set_deterministic_mode,
    to_onehot, transform_coordinates
)

In other words, I would modify the imports here to say

from pyroved.utils.nn import set_deterministic_mode

and remove the code in the __init__.py, for example. However, I'm happy to be convinced that your current way is better. What do you think?

Hi @ziatdinovmax ,

Say I defined a vae = pv.models.iVAE(in_dim, latent_dim=2, invariances=['r'], seed=0)

1. vae.manifold2d(d=8,cmap="jet") ---> plots the decoded latent space on grid, how to save it to a png file directly on disc. I know
   we can pass savefig = True in atomai but couldnt find here.

2. Is there a function or modular way to create latent images ?

While it is a huge pain to implement when using custom libraries, e.g. torch, since everything is typed in the code anyway, it might be worthwhile to add a proper mypy type check into the CI pipeline. That said, I really am no expert on using stubs so it could be a challenge/time consuming.

Edit: actually it doesn't look too bad: https://realpython.com/python-type-checking/#adding-stubs

What are these objects?

I'd recommend against module-level warnings suppression in favor of

with warnings.catch_warnings():
    warnings.simplefilter('ignore')

at the precise spot where the warnings are thrown.

Thoughts?

See here. I prefer this style personally and it definitely is just a personal choice, but if there are no objections this is how I'll start documenting!

This should be easy to implement by using von Mises Fisher distribution (or the qualitatively similar Projected Normal distribution) instead of the Normal distribution for sampling a latent angle. One will also need to modify the outputs of the encoder NN such that it outputs a "concentration" parameter in addition to z_loc and z_scale.

It will run Gaussian Process regression to interpolate between existing encodings in order to predict 'labels' on new inputs X.

@utkarshp1161

Originally posted by @ziatdinovmax in #53 (reply in thread)

Currently, all the VED/VAE models work only with a single channel data (e.g. grayscale images but not color images). The extension to multiple channels should be pretty straightforward if there is a use case.

There appears to be quite a bit of repetition in the classes in models.py, especially in e.g. the __init__. I'd suggest refactoring, and moving a lot of the shared code into the base class (and then adding documentation to the base class, of course).

The following modules may benefit from more extensive documentation: conv.py, fc.py, and pretty much all modules in the utils.