Skip to content

Instantly share code, notes, and snippets.

View HudsonHuang's full-sized avatar
🌙
DayDreaming

雲夢 HudsonHuang

🌙
DayDreaming
40 followers · 202 following
  • Hongkong
View GitHub Profile
@HudsonHuang
HudsonHuang / TransLayer.py
Created October 18, 2020 09:50
Transformer for convlution, replacing SE module
import torch
import math
from torch.nn import functional as F
from torch import nn
class PositionalEncoding(nn.Module):
def __init__(self, d_model, height, width, dropout=0.1):
super().__init__()
self.dropout = nn.Dropout(p=dropout)
@HudsonHuang
HudsonHuang / spectral_feature_compare.md
Last active December 31, 2019 03:43
spectral_feature_compare

| Feature | nnAudio | torch.stft | kapre | torchaudio | tf.signal(or else tf. stuff) | torch-stft | librosa | | ---|---|---|---|---|---|---|---|--- | | Trainable | 1 | 0 | 1 | 0 | 0 | 1 | 0 | | ModelConvert* | 1 | 0 | 1 | 0 | 0 | 1 | 0 | | Speed(Need test**) | 0 | 0 | 0 | 0 | 0 | 0 | 0 | | Differentiable(Not sure**) | 1 | 1 | 1 | 1 | 1 | 1 | 0 | | Mel | 1 | 0 | 1 | 1 | 1 | 0 | 1 | | MFCC | 0 | 0 | 0 | 1 | 1 | 0 | 1 | | CQT | 1 | 0 | 0 | 0 | 0 | 0 | 1 | | GPU support | 1 | 1 | 1 | 1 | 1 | 1 | 0 |

@HudsonHuang
HudsonHuang / model_activation_visualization.py
Last active December 20, 2019 11:24
View activation of any pytorch model
import PIL
import torch
import torchvision
import numpy as np
import matplotlib.pyplot as plt
def get_image(path, imsize=-1):
"""Load an image and resize to a cpecific size.
@HudsonHuang
HudsonHuang / benchmark.py
Last active May 11, 2022 19:02
wave file IO benchmark(soundfile and librosa)
## linux + HDD
# soundfile r: 111.3170223236084
# soundfile w: 21.47102665901184
# librosa r: 27.82967972755432
# librosa w: 24.776712656021118
## windows + SSD
# soundfile r: 12.918063640594482
# soundfile w: 11.726674318313599
# librosa r: 16.433390378952026
@HudsonHuang
HudsonHuang / audio_format.py
Created October 11, 2019 15:21
Convert audio data of PCM16/float32 to byte, and vice versa.
"""Helper functions for working with audio files in NumPy."""
"""some code borrowed from https://github.com/mgeier/python-audio/blob/master/audio-files/utility.py"""
import numpy as np
import contextlib
import librosa
import struct
import soundfile
def float_to_byte(sig):
@HudsonHuang
HudsonHuang / do_ifft_using_fft.py
Created June 26, 2019 09:41
do ifft using fft
from scipy.fftpack import fft
import numpy as np
# random complex array of N = 4
a = np.array([11.+11.4j, 2.47+2.3j, 30.89+73.1j, 44.+16.3j])
print(a)
# do FFT to a to transform to frequency domain
da = fft(a)
@HudsonHuang
HudsonHuang / grad.py
Created June 26, 2019 07:56
DFT的梯度
# FFT 的 一阶，二阶和n阶导数：https://math.mit.edu/~stevenj/fft-deriv.pdf
# DFT的导数：https://math.stackexchange.com/a/1658364/684858
import tensorflow as tf
import numpy as np
import torch
from torch.autograd import gradcheck, Variable
# mag loss
def mag(x):
@HudsonHuang
HudsonHuang / mixup.py
Last active December 2, 2022 00:29
mixup in numpy, tensorflow(keras), and pytorch
import numpy as np
import tensorflow as tf
def mixup_np(features, labels, alpha=0.1):
# numpy version
num_examples = features.shape[0]
num_class = labels.shape[-1]
mix = np.random.beta(alpha, alpha, size=[num_examples])
features = np.swapaxes(features, 0, -1)
features = features * mix + features[::-1] * (np.ones_like(mix) - mix)
@HudsonHuang
HudsonHuang / kaldi_install.sh
Last active June 15, 2019 15:03
kaldi_install.sh
git clone https://github.com/kaldi-asr/kaldi.git kaldi --origin upstream
cd kaldi/tools
make -j 4
# NOTE: If error happend, run this command and apt-get install necessary packages：
# sh ./extras/check_dependencies.sh
cd ../src/
./configure --shared
make depend -j 4
make -j 4
./run.sh
@HudsonHuang
HudsonHuang / Pytorch performance guide.md
Last active March 25, 2020 02:43
Pytorch performance guide
  1. Using CUDA in correct way：

  • 确定性卷积：（把所有操作的seed=0，以便重现，会变慢） torch.backends.cudnn.deterministic https://oldpan.me/archives/pytorch-conmon-problem-in-training

    添加torch.cuda.get_device_name和torch.cuda.get_device_capability实现如下功能。例：

    torch.cuda.get_device_name(0) 'Quadro GP100' torch.cuda.get_device_capability(0) (6, 0)