walloc 0.3.6

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 WEBP

from PIL import Image
import torch
from torchvision.transforms import PILToTensor

def concatenate_channels(x):
    batch_size, N, h, w = x.shape
    if N % 4 != 0 or int((N // 4)**0.5) ** 2 * 4 != N:
        raise ValueError("Number of channels must satisfy N = 4 * (n^2) for some integer n.")
    
    n = int((N // 4)**0.5)
    x = x.view(batch_size, 4, n, n, h, w)
    x = x.permute(0, 2, 4, 3, 5, 1).contiguous()
    x = x.view(batch_size, n*h, n*w, 4)
    return x

def split_channels(x, N):
    batch_size, _, H, W = x.shape
    n = int((N // 4)**0.5)
    h = H // n
    w = W // n
    
    x = x.view(batch_size, 4, n, h, n, w)
    x = x.permute(0, 1, 2, 4, 3, 5).contiguous()
    x = x.view(batch_size, N, h, w)
    return x

def latent_to_pil(latent, n_bits):
    latent_bytes = to_bytes(latent, n_bits)
    concatenated_latent = concatenate_channels(latent_bytes)
    
    pil_images = []
    for i in range(concatenated_latent.shape[0]):
        pil_image = Image.fromarray(concatenated_latent[i].numpy(), mode='RGBA')
        pil_images.append(pil_image)
    
    return pil_images

def pil_to_latent(pil_images, N, n_bits):
    tensor_images = [PILToTensor()(img).unsqueeze(0) for img in pil_images]
    tensor_images = torch.cat(tensor_images, dim=0)
    split_latent = split_channels(tensor_images, N)
    latent = from_bytes(split_latent, n_bits)
    return latent

def to_bytes(x, n_bits):
    max_value = 2**(n_bits - 1) - 1
    min_value = -max_value - 1
    if x.min() < min_value or x.max() > max_value:
        raise ValueError(f"Tensor values should be in the range [{min_value}, {max_value}].")
    return (x + (max_value + 1)).to(torch.uint8)

def from_bytes(x, n_bits):
    max_value = 2**(n_bits - 1) - 1
    return (x.to(torch.float32) - (max_value + 1))

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

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

compression_ratio:  1.789818658509941

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

tensor(1.)

!jupyter nbconvert --to markdown README.ipynb

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