def fasfasa():
import audiomentations
from src import core
crop_size = 98303 # 768 * 128 - 1
transforms = audiomentations.Compose([
core.transforms.ToMono(),
core.transforms.Squeeze(),
core.transforms.ToNumpy(),
audiomentations.TimeStretch(min_rate=0.7, max_rate=1.3, p=0.5),
audiomentations.PitchShift(min_semitones=-4,
max_semitones=4,
p=0.5),
core.transforms.ToTorch()
])
train_dataset = DSD100(root=Path.home() / 'Data' / 'Audio' / 'DSD100',
crop_size=crop_size,
transforms=transforms)
for i, e in zip(range(10), train_dataset):
print(e)
for fold, (t_idx, v_idx) in enumerate(kfold.split(X, y)):
train_gby.loc[v_idx, "kfold"] = fold
train_df = train_df.merge(train_gby[["recording_id", "kfold"]],
on="recording_id",
how="left")
print(train_df.kfold.value_counts())
train_df.to_csv(OUTPUT_DIR / "folds.csv", index=False)
species_fmin_fmax.to_csv(OUTPUT_DIR / "species_fmin_fmax.csv", index=False)
################################################
# audiomentations #
################################################
augmenter = A.Compose([
A.AddGaussianNoise(min_amplitude=0.01, max_amplitude=0.03, p=0.2),
A.PitchShift(min_semitones=-3, max_semitones=3, p=0.2),
A.Gain(p=0.2)
])
################################################
# Dataset #
################################################
def cut_spect(spect: torch.Tensor, fmin_mel: int, fmax_mel: int):
return spect[fmin_mel:fmax_mel]
def do_normalize(img: torch.Tensor):
bs, ch, w, h = img.shape
_img = img.clone()
import audiomentations
import cv2
from model.mixers import UseMixerWithProb, RandomMixer, SigmoidConcatMixer, AddMixer, SigmoidVerticalConcatMixer
from model.random_resized_crop import RandomResizedCrop, RandomResizedCrop2
from model.transforms import Compose, UseWithProb, SpecAugment, SpectreScale, PreprocessMelImage, GaussNoise, OneOf, \
PadToSize, RandomCrop, PreprocessSingleChannelMelImage
wave_augmentations = {
0: None,
1: audiomentations.Compose([
audiomentations.AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.015, p=0.5),
audiomentations.TimeStretch(min_rate=0.8, max_rate=1.25, p=0.5),
audiomentations.PitchShift(min_semitones=-4, max_semitones=4, p=0.5),
audiomentations.Shift(min_fraction=-0.5, max_fraction=0.5, p=0.5),
], p=0.96), # (1-(1-0.5)^4)*0.96==0.9 - In total there will be 90% augmented samples
2: audiomentations.Compose([
audiomentations.AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.010, p=0.95),
audiomentations.Shift(min_fraction=-0.1, max_fraction=0.1, p=0.3),
], p=1),
}
size_4_sec_750_hop = 256
_base_mel_post_process = {
'none': [],
'3ch_1': [
# Use librosa.feature.delta with order 1 and 2 for creating 2 additional channels then divide by 100
PreprocessMelImage(),
],
'1ch_1': [PreprocessSingleChannelMelImage(), ],
def __init__(self, sample_rate, min_semitones=-4, max_semitones=4, p=0.5, **kwargs):
store_attr('min_semitones'), store_attr('max_semitones'), store_attr('p')
super().__init__(**kwargs)
self.tfm = partial(aug.PitchShift(min_semitones=min_semitones,
max_semitones=max_semitones, p=p), sample_rate=sample_rate)