walloc 0.3.3

Wavelet Learned Lossy Compression

---

datasets:

• danjacobellis/LSDIR_540

---

Wavelet Learned Lossy Compression (WaLLoC)

WaLLoC sandwiches a convolutional autoencoder between time-frequency analysis and synthesis transforms using CDF 9/7 wavelet filters. The time-frequency transform increases the number of signal channels, but reduces the temporal or spatial resolution, resulting in lower GPU memory consumption and higher throughput. WaLLoC's training procedure is highly simplified compared to other $\beta$-VAEs, VQ-VAEs, and neural codecs, but still offers significant dimensionality reduction and compression. This makes it suitable for dataset storage and compressed-domain learning. It currently supports 2D signals (e.g. grayscale, RGB, or hyperspectral images). Support for 1D and 3D signals is in progress.

Installation

1. Follow the installation instructions for torch
2. Install WaLLoC and other dependencies via pip

pip install walloc PyWavelets pytorch-wavelets

Pre-trained checkpoints

Pre-trained checkpoints are available on Hugging Face.

Training

Access to training code is provided by request via email.

Usage example

import os
import torch
import matplotlib.pyplot as plt
import numpy as np
from PIL import Image
from IPython.display import display
from torchvision.transforms import ToPILImage, PILToTensor
from walloc import walloc
class Args: pass

Load the model from a pre-trained checkpoint

wget https://hf.co/danjacobellis/walloc/resolve/main/v0.6.3_ext.pth

device = "cpu"
checkpoint = torch.load("v0.6.3_ext.pth",map_location="cpu")
args = checkpoint['args']
codec = walloc.Walloc(
    channels = args.channels,
    J = args.J,
    N = args.N,
    latent_dim = args.latent_dim,
    latent_bits = 5
)
codec.load_state_dict(checkpoint['model_state_dict'])
codec = codec.to(device)

Load an example image

wget "https://r0k.us/graphics/kodak/kodak/kodim05.png"

img = Image.open("kodim05.png")
img

Full encoding and decoding pipeline with .forward()

• If codec.eval() is called, the latent is rounded to nearest integer.

• If codec.train() is called, uniform noise is added instead of rounding.

with torch.no_grad():
    codec.eval()
    x = PILToTensor()(img).to(torch.float)
    x = (x/255 - 0.5).unsqueeze(0).to(device)
    x_hat, _, _ = codec(x)
ToPILImage()(x_hat[0]+0.5)

Accessing latents

with torch.no_grad():
    codec.eval()
    X = codec.wavelet_analysis(x,J=codec.J)
    Y = codec.encoder(X)
    X_hat = codec.decoder(Y)
    x_hat = codec.wavelet_synthesis(X_hat,J=codec.J)

print(f"dimensionality reduction: {x.numel()/Y.numel()}×")

dimensionality reduction: 12.0×

Y.unique()

tensor([-15., -14., -13., -12., -11., -10.,  -9.,  -8.,  -7.,  -6.,  -5.,  -4.,
         -3.,  -2.,  -1.,  -0.,   1.,   2.,   3.,   4.,   5.,   6.,   7.,   8.,
          9.,  10.,  11.,  12.,  13.,  14.,  15.])

plt.figure(figsize=(5,3),dpi=150)
plt.hist(
    Y.flatten().numpy(),
    range=(-17.5,17.5),
    bins=35,
    density=True,
    width=0.8);
plt.title("Histogram of latents")
plt.xticks(range(-15,16,5));

Lossless compression of latents using PNG

Y_pil = walloc.latent_to_pil(Y,5)
Y_pil[0]

Y_pil[0].save('latent.png')
print("compression_ratio: ", x.numel()/os.path.getsize("latent.png"))

compression_ratio:  20.307596963280485

Y2 = walloc.pil_to_latent(Y_pil, 16, 5)
(Y == Y2).sum()/Y.numel()

tensor(1.)

!jupyter nbconvert --to markdown README.ipynb

[NbConvertApp] Converting notebook README.ipynb to markdown
[NbConvertApp] Support files will be in README_files/
[NbConvertApp] Writing 4133 bytes to README.md

!sed -i 's|!\[png\](README_files/\(README_[0-9]*_[0-9]*.png\))|![png](https://github.com/danjacobellis/walloc/raw/main/README_files/\1)|g' README.md