class config(Config):
groups = ['foo', 'bar']
MY_INT = setting(10, foo=10, bar=20)
MY_FLOAT = setting(10.0, bar=20.0)
MY_STR = setting('foo', foo='asdf')
MY_BOOL = setting(True, bar=False)
class config(Config):
groups = ['train', 'test']
PATH_BASE = 'C:\Proy\sound-source-separation\data\musdb18wav'
PATH_RAW = setting(default=PATH_BASE + 'train/raw_audio/',
train=PATH_BASE + 'train/raw_audio/',
test=PATH_BASE + 'test/raw_audio/')
PATH_SPEC = setting(default=PATH_BASE + 'train/complex/',
train=PATH_BASE + 'train/complex/',
test=PATH_BASE + 'test/complex/')
PATH_INDEXES = setting(default=PATH_BASE + 'train/indexes/',
train=PATH_BASE + 'train/indexes/',
test=PATH_BASE + 'test/indexes/')
LOG_FILENAME = 'computing_spec.log'
FR = 8192
FFT_SIZE = 1024
HOP = 256
MODE = 'conditioned'
INTRUMENTS = ['bass', 'drums', 'other', 'vocals', 'mix']
CONDITIONS = ['bass', 'drums', 'other', 'vocals']
CONDITION_MIX = 1 # complex conditions -> 1 only original instrumets, 2 combination of 2 instruments, etc
ADD_ZERO = True # add the zero condition
ADD_ALL = True # add the all mix condition
ADD_IN_BETWEEN = 1. # in between interval for the combination of several instruments
STEP = 1 # step in frames for creating the input data
CHUNK_SIZE = 4 # chunking the indexes before mixing -> define the number of data points of the same track
SPECTROGRAM_SHAPE = [512, 128, 1] # freq = 512, time = 128
class config(Config):
groups = ['foo', 'bar']
S1 = setting(10, foo=20)
S2 = setting(True, bar=False)
S3 = setting('a', foo='b', bar='c')
S4 = 0.5
class config(Config):
groups = ['large', 'small']
NUM_LAYERS = setting(default=5, large=10, small=3)
NUM_UNITS = setting(default=128, large=512, small=32)
USE_SKIP_CONNECTIONS = setting(default=True, small=False)
LEARNING_RATE = 0.1
OPTIMIZER = 'adam'
class config(Config):
groups = ['foo', 'bar']
S1 = setting(10, foo=20)
S2 = setting(True, bar=False)
S3 = setting('a', foo='b', bar='c')
S4 = 0.5
S5 = setting([1, 2, 3], foo=[4, 5, 6])
S6 = setting({'a': 1, 'b': 2, 'c': 3}, bar={'d': 4, 'e': 5})
class config(Config):
groups = ['foo', 'bar']
MY_INT = setting(10, foo=10, bar=20)
MY_FLOAT = setting(10.0, bar=20.0)
MY_STR = setting('foo', foo='asdf')
@classmethod
def MY_BOOL(cls):
return True
class config(Config):
groups = ["vanilla", "melgan"]
TYPE = setting(default="vanilla", vanilla="vanilla", melgan="melgan")
# MELGAN PARAMETERS
INPUT_SIZE = 128
NGF = 32
N_RES_G = 3
HOP_LENGTH = 256
RATIOS = setting(default=[1, 1, 1, 2, 1, 1, 1],
vanilla=[1, 1, 1, 2, 1, 1, 1],
melgan=[8, 8, 2, 2])
NUM_D = 3
NDF = 16
N_LAYER_D = 4
DOWNSAMP_D = 4
# AUTOENCODER
CHANNELS = [128, 256, 256, 512, 512, 512, 128, 32]
KERNEL = 5
EXTRACT_LOUDNESS = False
AUGMENT = setting(default=5, vanilla=5, melgan=1)
# CLASSIFIER
CLASSIFIER_CHANNELS = [16, 64, 256]
CLASSIFIER_LIN_SIZE = [256, 64, 2]
# TRAIN PARAMETERS
PATH_PREPEND = "./runs/"
SAMPRATE = 24000
N_SIGNAL = setting(default=2**15, vanilla=2**15, melgan=2**14)
EPOCH = 1000
BATCH = 1
LR = 1e-4
NAME = "untitled"
CKPT = None
WAV_LOC = None
FILE_LIST = None
LMDB_LOC = "./preprocessed"
BACKUP = 10000
EVAL = 1000
# INCREMENTAL GENERATION
USE_CACHED_PADDING = False
BUFFER_SIZE = 1024
TARGET_SR = 48000
class config(Config):
MY_INT = setting(10)
MY_FLOAT = setting(10.0)
MY_STR = setting('foo')
MY_BOOL = setting(True)
MY_LIST = setting([1, 2, 3])
MY_DICT = setting({'a': 1, 'b': 2, 'c': 3})
MY_SHORTHAND_INT = 20
class config(Config):
groups = [
'standard', 'simple_dense', 'complex_dense', 'simple_cnn',
'complex_cnn'
]
GROUP = setting(
default='simple_dense',
simple_dense='simple_dense',
complex_dense='complex_dense',
simple_cnn='simple_cnn',
complex_cnn='complex_cnn',
)
PATH_BASE = '/net/guzheng/data2/anasynth_nonbp/meseguerbrocal/source_separation/musdb18/models/'
NAME = 'with_val_all_files'
PATH_MODEL = setting(
os.path.join(PATH_BASE, 'conditioned/simple_dense'),
standard=os.path.join(PATH_BASE, 'standard'),
simple_dense=os.path.join(PATH_BASE, 'conditioned/simple_dense'),
complex_dense=os.path.join(PATH_BASE, 'conditioned/complex_dense'),
simple_cnn=os.path.join(PATH_BASE, 'conditioned/simple_cnn'),
complex_cnn=os.path.join(PATH_BASE, 'conditioned/complex_cnn'))
PATH_AUDIO = '/net/guzheng/data2/anasynth_nonbp/meseguerbrocal/source_separation/musdb18/test/complex'
TARGET = [
'vocals', 'bass', 'drums', 'rest'
] # ['vocals', 'bass', 'bass_vocals'] -> not ready yet for complex conditions
INSTRUMENTS = ['bass', 'drums', 'rest', 'vocals'
] # to check that has the same order than the training
OVERLAP = 0
MODE = setting(default='conditioned', standard='standard')
EMB_TYPE = setting(default='dense',
simple_dense='dense',
complex_dense='dense',
simple_cnn='cnn',
complex_cnn='cnn')
class config(Config):
groups = ['train', 'test']
PATH_BASE = '/data2/anasynth_nonbp/meseguerbrocal/source_separation/musdb18/'
PATH_RAW = setting(default=PATH_BASE + 'train/raw_audio/',
train=PATH_BASE + 'train/raw_audio/',
test=PATH_BASE + 'test/raw_audio/')
PATH_SPEC = setting(default=PATH_BASE + 'train/complex/',
train=PATH_BASE + 'train/complex/',
test=PATH_BASE + 'test/complex/')
PATH_INDEXES = setting(default=PATH_BASE + 'train/indexes/',
train=PATH_BASE + 'train/indexes/',
test=PATH_BASE + 'test/indexes/')
FR = 8192
FFT_SIZE = 1024
HOP = 256
MODE = 'conditioned'
INTRUMENTS = ['bass', 'drums', 'rest', 'vocals', 'mix']
CONDITIONS = ['bass', 'drums', 'rest', 'vocals']
CONDITION_MIX = 1 # complex conditions -> 1 only original instrumets, 2 combination of 2 instruments, etc
ADD_ZERO = True # add the zero condition
ADD_ALL = True # add the all mix condition
ADD_IN_BETWEEN = 1. # in between interval for the combination of several instruments
STEP = 1 # step in frames for creating the input data
CHUNK_SIZE = 4 # chunking the indexes before mixing -> define the number of data points of the same track
class config(Config):
adsf = setting(123)
class config(Config):
groups = ['hello']
FOO = setting(None, hello=123)
class config(Config):
FOO = setting(None)
class config(Config):
FOO = setting(10)
class c(Config):
groups = ['a']
A = setting(1, a=2)
B = 2.0
C = True
D = 'foo'
class config(Config):
MY_LIST = setting(1j)
class config(Config):
groups = ['foo', 'bar']
S1 = setting(10, foo=20)
SECOND_OPTION = setting(True, bar=False)
S3 = setting('a', foo='b', bar='c')
class config(Config):
groups = ['a', 'b']
FOO = setting(10, a=20, b=30)
class config(Config):
groups = ['foo', 'bar']
MY_INT = setting(10, baz=20)
class config(Config):
groups = ['standard', 'simple_dense', 'complex_dense', 'simple_cnn',
'complex_cnn']
# General
MODE = setting(default='conditioned', standard='standard')
NAME = 'with_val_all_files'
ADD_TIME = False # add the time and date in the name
TARGET = 'vocals' # only for standard version
# GENERATOR
PATH_BASE = '/data2/anasynth_nonbp/meseguerbrocal/source_separation/musdb18/'
# default = conditioned
INDEXES_TRAIN = setting(
default=os.path.join(
PATH_BASE, 'train/indexes/indexes_conditioned_1_4_1_True_True_1.0.npz'),
standard=os.path.join(
PATH_BASE, 'train/indexes/indexes_standard_1_4.npz')
)
INDEXES_VAL = setting(
default=os.path.join(
PATH_BASE, 'train/indexes/indexes_conditioned_128_4_1_True_True_1.0.npz'),
standard=os.path.join(
PATH_BASE, 'train/indexes/indexes_standard_128_4.npz')
)
NUM_THREADS = tf.data.experimental.AUTOTUNE # 32
N_PREFETCH = tf.data.experimental.AUTOTUNE # 4096
# checkpoints
EARLY_STOPPING_MIN_DELTA = 1e-6
EARLY_STOPPING_PATIENCE = 15
REDUCE_PLATEAU_PATIENCE = 5
# training
BATCH_SIZE = 64
N_BATCH = 2048
N_EPOCH = 1000
PROGRESSIVE = False
# unet paramters
INPUT_SHAPE = [512, 128, 1] # freq = 512, time = 128
FILTERS_LAYER_1 = 16
N_LAYERS = 6
LR = 1e-3
ACTIVATION_ENCODER = 'leaky_relu'
ACTIVATION_DECODER = 'relu'
ACT_LAST = 'sigmoid'
LOSS = 'mean_absolute_error'
# -------------------------------
# control parameters
CONTROL_TYPE = setting(
'dense', simple_dense='dense', complex_dense='dense',
simple_cnn='cnn', complex_cnn='cnn'
)
FILM_TYPE = setting(
'simple', simple_dense='simple', complex_dense='complex',
simple_cnn='simple', complex_cnn='complex'
)
Z_DIM = 4 # for musdb -> 4 instruments: vocals, drums, bass, rest
ACT_G = 'linear'
ACT_B = 'linear'
N_CONDITIONS = setting(
6, simple_dense=6, complex_dense=1008,
simple_cnn=6, complex_cnn=1008
)
# cnn control
N_FILTERS = setting(
[16, 32, 64], simple_cnn=[16, 32, 64], complex_cnn=[32, 64, 256]
)
PADDING = ['same', 'same', 'valid']
# Dense control
N_NEURONS = setting(
[16, 64, 256], simple_dense=[16, 64, 256],
complex_dense=[16, 256, 1024]
)
