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()
_img = _img.view(bs, -1)
def __init__(self, sample_rate, min_gain_in_db=-12, max_gain_in_db=12, p=0.5, **kwargs):
store_attr('min_gain_in_db'), store_attr('max_gain_in_db'), store_attr('p')
super().__init__(**kwargs)
self.tfm = partial(aug.Gain(min_gain_in_db=min_gain_in_db,
max_gain_in_db=max_gain_in_db, p=p), sample_rate=sample_rate)
