tiny-diff 0.1.0

Hackable implementation of diffusion models

Tiny Diff (Beta)

This project attempts to provide a minimal implementation of diffusion models written in pure python, without huggingface dependencies.

It aims to:

• Be object oriented to enable end users to extend the SDK for their custom workflows
• Be, as it name states, easy to use (hence focused in simple usecases)
• Have as little duplicated code as possible.
• Be modular so users can reuse layers, blocks if they

You can check more info at the documentation site. Conditional diffusion is still to be included.

At the moment, all the necessary modules to build the original latent diffusion model are included

Installation

Simply run

pip install tiny-diff

Quickstart

Check the examples folder for a general guideline on how to train Variational Autoencoders + GAN and using them as embedders for Latent Diffusion models

Esentially, the repo provides three groups of models right now

UNets

Used for diffusion models, they follow the architecture of Huggingface Diffusers implementation

from tiny_diff.models import UNet

model = UNet(
    base_channels=128,
    channel_mult=[1, 1, 2, 2, 4, 4],
    attn_at_layer=[False, False, False, False, True, False],
    n_res_blocks=3,
    num_groups=-1,
    input_channels=3,
    kernel_size=3,
    factor=2,
    dropout=0.0,
    attention_factory=attention_factory,
)

Variational Autoencoders

Used for latent diffusion models as a way to compress the information into an embedded space.

from tiny_diff.models.vae import ConvVAE
from tiny_diff.losses import VAELoss, BetaPolicy
from torch.nn import L1Loss, LeakyReLU
from tiny_diff.modules.layer_factory import LayerFactory

nonlinearity_factory = LayerFactory(cls=LeakyReLU, kwargs={"negative_slope": 0.03})

vae_loss = VAELoss(recon_l=L1Loss(), kl_w_policy=BetaPolicy(a=1e-6, b=6e-2, offset=40, slope=2))
conv_vae_model = ConvVAE(
    transform=...,  # A callable to apply on the input
    vae_loss=vae_loss,
    component_kwargs={
        "base_channels": 32,
        "channel_mult": [1, 2, 4],
        "n_res_blocks": 2,
        "num_groups": -1,
        "input_channels": 3,
        "kernel_size": 4,
        "factor": 2,
        "dropout": 0.0,
        "attn_channels": [],
        "z_channels": 4,
        "interpolation_mode": "bicubic",
        "nonlinearity_factory": nonlinearity_factory,
    },
)

print(conv_vae_model)

Schedulers

These guide the diffusion process to add and remove noise.

from tiny_diff.schedulers import LatentDDPMScheduler, DDPMScheduler

scheduler = LatentDDPMScheduler(beta_start=0.00085, beta_end=0.012)

Submodules

The framework is modular so you can reuse whatever component you want. Here are some examples:

from tiny_diff.conv import ConvBlock, PreNormActConvBlock, UpsampleConvBlock
from tiny_diff.layer_factory import LayerFactory
from tiny_diff.nonlinearity import Swish
from tiny_diff.resampling import Downsample, Upsample
from tiny_diff.residual import (
    DownsampleResidualLayer,
    ERABlock,
    ERARBlock,
    ERBlock,
    EResidualLayer,
    InterpolationResidualLayerABC,
    RBlock,
    UpsampleResidualLayer,
)

Layer Factory

One simple yet unique element of this framework is the layer factory, which is a simple implementation of the Factory pattern. This allows to change the attention or non linearity layers from a single init param while avoiding aliasing.

from tiny_diff.modules.layer_factory import LayerFactory

nonlinearity_factory = LayerFactory(cls=LeakyReLU, kwargs={"negative_slope": 0.03})

If you look at the VAE or UNet model, it's trivial to see how to change the attention or non_linearity throughout the whole network.