Simple torch.nn.module implementation of Alias-Free-GAN style filter and resample

Last update: Dec 22, 2022

Overview

Alias-Free-Torch

Simple torch module implementation of Alias-Free GAN.

This repository including

Alias-Free GAN style lowpass sinc filter @filter.py
Alias-Free GAN style up/downsample @resample.py
Alias-Free activation @act.py
and test codes @./test

Note: Since this repository is unofficial, filter and upsample could be different with official implementation.

Note: 2d lowpass filter is applying sinc instead of jinc (first order Bessel function of the first kind) in paper

Requirements

Due to torch.kaiser_window and torch.i0 are implemeted after 1.7.0, our repository need torch>=1.7.0.

Pytorch>=1.7.0

TODO

2d sinc filter
2d resample
devide 1d and 2d modules
pip packaging

Test results 1d

Filter sine	Filter noise

upsample	downsample

Test results 2d

Filter L1 norm sine	Filter noise

upsample	downsample

Activation

References

Alias-Free GAN
adefossez/julius
A. V. Oppenheim and R. W. Schafer. Discrete-Time Signal Processing. Pearson, International Edition, 3rd edition, 2010

Acknowledgement

This work is done at MINDsLab Inc.

Thanks to teammates at MINDsLab Inc.

Comments

Batched resampling for the new implementation

Hi, thank you very much for the contribution.

I think the new implementation of resample.Upsample1d and resample.Downsample1d breaks batched resampling when using groups=C without expanding the filter to match the shape. Perhaps the implementation should be like the below (maybe similar goes to 2d):

Upsample1d.forward()

    # x: [B,C,T]
    def forward(self, x):
        B, C, T = x.shape
        x = F.pad(x, (self.pad, self.pad), mode='reflect')
        # TConv with filter expanded to C with C groups for depthwise op
        x = self.ratio * F.conv_transpose1d(
            x, self.filter.expand(C, -1, -1), stride=self.stride, groups=C)
        pad_left = self.pad * self.stride + (self.kernel_size -
                                             self.stride) // 2
        pad_right = self.pad * self.stride + (self.kernel_size - self.stride +
                                              1) // 2
        x = x[..., pad_left:-pad_right]

LowPassFilter1d.forward()

    #input [B,C,T]
    def forward(self, x):
        B, C, T = x.shape
        if self.padding:
            x = F.pad(x, (self.left_pad, self.right_pad),
                      mode=self.padding_mode)
        # Conv with filter expanded to C with C groups for depthwise op
        out = F.conv1d(x, self.filter.expand(C, -1, -1), stride=self.stride, groups=C) # typo 'groupds' btw
        return out

Could you check the correctness? Thanks again for the implementation!

opened by L0SG 2

torch.speical.i1 typo

https://github.com/junjun3518/alias-free-torch/blob/f1fddd52fdd068ee475e82ae60c92e1bc24ffe02/src/alias_free_torch/filter.py#L22

At this line I believe you wanted torch.special.i1.

opened by torridgristle 2
"if self.pad / self.padding" in LowPassFilter2d

https://github.com/junjun3518/alias-free-torch/blob/258551410ff7bf02e06ece7c597466dc970fe5c7/src/alias_free_torch/filter.py#L165 https://github.com/junjun3518/alias-free-torch/blob/258551410ff7bf02e06ece7c597466dc970fe5c7/src/alias_free_torch/filter.py#L173

In LowPassFilter2d it looks like if self.pad: should change to if self.padding:, or self.padding = padding should change to self.pad = padding to match LowPassFilter1d.

opened by torridgristle 1
Padding Bool typo

https://github.com/junjun3518/alias-free-torch/blob/258551410ff7bf02e06ece7c597466dc970fe5c7/src/alias_free_torch/filter.py#L73

padding: bool: True, should be padding: bool = True,

I'm not sure if this causes an error with every version of PyTorch, but it does with PyTorch 1.12.0+cu113 on Python 3.7.13

opened by torridgristle 1
2D Filter Jinc appears to be wrong

Here is a plot of the generated 1D sinc filter kernel.

Here is a plot of the generated 2D jinc filter kernel.

I'd expect it to look more like a series of rings or ripples, rather than a donut or torus.

The FFT output for randn noise put through the 2D filter doesn't look right either.

Changing filter_ = 2 * cutoff * window * jinc(2 * cutoff * time) to filter_ = 2 * cutoff * window * sinc(2 * cutoff * time) in kaiser_jinc_filter2d makes a more familiar kernel.

And the FFT output for randn noise put through this 2D filter looks about how I'd expect.

opened by torridgristle 3

Releases(v0.0.6)

v0.0.6(Jul 26, 2022)

https://pypi.org/project/alias-free-torch/0.0.6/

Tested version
Source code(tar.gz)
Source code(zip)
v0.0.3(Jul 18, 2022)

https://pypi.org/project/alias-free-torch/0.0.3/

Bug fix for torch.special / remove print / split pad from conv_transpose
Source code(tar.gz)
Source code(zip)
v0.0.2(Jun 22, 2022)

https://pypi.org/project/alias-free-torch/0.0.2/

Rewrite upsample, jinc applied
Source code(tar.gz)
Source code(zip)
v0.0.1(Nov 2, 2021)

v0.0.1 released https://pypi.org/project/alias-free-torch/
Source code(tar.gz)
Source code(zip)

Owner

이준혁(Junhyeok Lee)

Audio/Speech Deep Learning Researcher @mindslab-ai

GitHub Repository

Simple torch.nn.module implementation of Alias-Free-GAN style filter and resample

Related tags

Overview

Alias-Free-Torch

Requirements

TODO

Test results 1d

Test results 2d

References

Acknowledgement

Comments

Batched resampling for the new implementation

torch.speical.i1 typo

"if self.pad / self.padding" in LowPassFilter2d

Padding Bool typo

2D Filter Jinc appears to be wrong

Releases(v0.0.6)

v0.0.6(Jul 26, 2022)

v0.0.3(Jul 18, 2022)

v0.0.2(Jun 22, 2022)

v0.0.1(Nov 2, 2021)

Owner

이준혁(Junhyeok Lee)

Code for Environment Dynamics Decomposition (ED2).

RoIAlign & crop_and_resize for PyTorch

Voice control for Garry's Mod

CLNTM - Contrastive Learning for Neural Topic Model

Semi-Supervised Semantic Segmentation with Pixel-Level Contrastive Learning from a Class-wise Memory Bank

The Official PyTorch Implementation of "LSGM: Score-based Generative Modeling in Latent Space" (NeurIPS 2021)

Get 2D point positions (e.g., facial landmarks) projected on 3D mesh

TransPrompt - Towards an Automatic Transferable Prompting Framework for Few-shot Text Classification

Implementation of Memory-Efficient Neural Networks with Multi-Level Generation, ICCV 2021

Pytorch code for "State-only Imitation with Transition Dynamics Mismatch" (ICLR 2020)

[ICCV21] Self-Calibrating Neural Radiance Fields

InDuDoNet+: A Model-Driven Interpretable Dual Domain Network for Metal Artifact Reduction in CT Images

EPSANet：An Efficient Pyramid Split Attention Block on Convolutional Neural Network

Hierarchical Memory Matching Network for Video Object Segmentation (ICCV 2021)

An implementation of shampoo

Self-Supervised Document-to-Document Similarity Ranking via Contextualized Language Models and Hierarchical Inference

On the Adversarial Robustness of Visual Transformer

Facilitating Database Tuning with Hyper-ParameterOptimization: A Comprehensive Experimental Evaluation

NumPy로 구현한 딥러닝 라이브러리입니다. (자동 미분 지원)

HugsVision is a easy to use huggingface wrapper for state-of-the-art computer vision