def parse_args(argv):
parser = common.common_arguments_parser()
parser.add_argument("--spectrogram", type=str, help="Postprocessing layer")
parser.add_argument("--capacity_multiplier",
default=8,
type=int,
help="Capacity")
parser.add_argument("--voicing",
action='store_true',
help="Add voicing model.")
args = parser.parse_args(argv)
hop_length = 512
defaults = {
"samplerate": 44100,
"context_width": 14 * hop_length,
"annotations_per_window": 20,
"hop_size": 1,
"frame_width": hop_length,
"note_range": 72,
"min_note": 24,
"batch_size": 8,
"evaluate_every": 5000,
"evaluate_small_every": 1000,
"annotation_smoothing": 0.177,
"spectrogram": "hcqt",
"capacity_multiplier": 8,
"voicing": False
}
specified_args = common.argument_defaults(args, defaults)
common.name(args, specified_args, "bittner")
# common.name(args, "cqt_voicing_residual_batchnorm")
return args
def parse_args(argv):
parser = common.common_arguments_parser()
parser.add_argument("--capacity_multiplier", type=int, help="Capacity multiplier")
parser.add_argument("--spectrogram", type=str, help="Spectrogram method")
parser.add_argument("--spectrogram_top_db", type=float, help="Spectrogram top_db")
parser.add_argument("--spectrogram_filter_scale", type=float, help="Spectrogram filter_scale")
parser.add_argument("--spectrogram_undertone_stacking", type=int, help="spectrogram undertone stacking")
parser.add_argument("--spectrogram_overtone_stacking", type=int, help="spectrogram overtone stacking")
parser.add_argument("--undertone_stacking", type=int, help="Undertone stacking in the model")
parser.add_argument("--overtone_stacking", type=int, help="Overtone stacking in the model")
args = parser.parse_args(argv)
defaults = {
"samplerate": 44100, "context_width": 0, "annotations_per_window": 50, "hop_size": 1,
"frame_width": HOP_LENGTH,
"note_range": 72, "min_note": 24,
"evaluate_every": 5000,
"evaluate_small_every": 1000,
"spectrogram": "cqt",
"learning_rate": 0.001,
"learning_rate_decay": 0.85,
"learning_rate_decay_steps": 5000,
"undertone_stacking": 0,
"overtone_stacking": 1,
"spectrogram_undertone_stacking": 1,
"spectrogram_overtone_stacking": 5,
"spectrogram_top_db": 80,
"spectrogram_filter_scale": 1.0,
"capacity_multiplier": 64,
}
specified_args = common.argument_defaults(args, defaults)
common.name(args, specified_args, "bittner")
return args
def parse_args(argv):
parser = common.common_arguments_parser()
# Model specific arguments
parser.add_argument("--input_normalization",
type=int,
help="Enable normalizing each input example")
parser.add_argument("--capacity_multiplier",
type=int,
help="Capacity multiplier of the model")
parser.add_argument(
"--multiresolution_convolution",
type=int,
help="Number of different resolution of the first convolution layer")
parser.add_argument("--variable_stride",
action='store_true',
help="Variable stride")
parser.add_argument("--first_layer_capacity",
type=int,
help="Capacity multiplier")
args = parser.parse_args(argv)
defaults = {
# Change some of the common defaults
"context_width": 978,
"input_normalization": 1,
"capacity_multiplier": 16,
"multiresolution_convolution": 0,
"variable_stride": False,
"first_layer_capacity": 1,
}
specified_args = common.argument_defaults(args, defaults)
common.name(args, specified_args, "crepe")
return args
def parse_args(argv):
parser = common.common_arguments_parser()
# Model specific arguments
# input
parser.add_argument("--spectrogram", type=str, help="Spectrogram method")
parser.add_argument("--spectrogram_top_db",
type=float,
help="Spectrogram top_db")
parser.add_argument("--spectrogram_filter_scale",
type=float,
help="Spectrogram filter_scale")
parser.add_argument("--spectrogram_undertone_stacking",
type=int,
help="spectrogram undertone stacking")
parser.add_argument("--spectrogram_overtone_stacking",
type=int,
help="spectrogram overtone stacking")
parser.add_argument(
"--cut_context",
type=int,
help="Cut unnecessary context, doesn't work with dilations!")
# model
parser.add_argument("--architecture", type=str, help="Model architecture")
parser.add_argument("--filters", type=int, help="Filters in convolutions")
parser.add_argument("--stacks", type=int, help="Stacks")
parser.add_argument("--conv_range",
type=int,
help="Stack kernel size in frequency axis")
parser.add_argument("--undertone_stacking",
type=int,
help="Undertone stacking in the model")
parser.add_argument("--overtone_stacking",
type=int,
help="Overtone stacking in the model")
parser.add_argument("--activation",
type=str,
help="Activation function for the convolution stack")
# context
parser.add_argument("--conv_ctx",
nargs="+",
type=int,
help="Stack kernel sizes in time axis")
parser.add_argument("--dilations",
nargs="+",
type=int,
help="Dilation rate for the convolutions")
parser.add_argument("--last_conv_kernel", nargs=2, type=int)
# residual
parser.add_argument(
"--residual_hop",
type=int,
help="Size of one block around which there is a residual connection")
parser.add_argument("--residual_end",
type=int,
help="No residual connection in last N layers")
parser.add_argument(
"--residual_op",
type=str,
help=
"Residual connection operation (add for ResNet, concat for DenseNet)")
# regularization
parser.add_argument("--batchnorm", type=int)
parser.add_argument("--dropout", type=float)
parser.add_argument("--specaugment_prob", type=float)
parser.add_argument("--specaugment_freq_mask_num", type=int)
parser.add_argument("--specaugment_freq_mask_max", type=int)
parser.add_argument("--specaugment_time_mask_num", type=int)
parser.add_argument("--specaugment_time_mask_max", type=int)
# voicing module
parser.add_argument("--voicing", type=int)
parser.add_argument("--voicing_input", type=str)
args = parser.parse_args(argv)
hop_length = 512
defaults = {
# Change some of the common defaults
"samplerate": 44100,
"context_width": 10 * hop_length,
"annotations_per_window": 5,
"hop_size": 1,
"frame_width": hop_length,
"note_range": 72,
"min_note": 24,
"evaluate_every": 5000,
"evaluate_small_every": 1000,
"annotation_smoothing": 0.18,
"batch_size": 8,
# Model specific defaults
"learning_rate_decay_steps": 10000,
"learning_rate_decay": 0.8,
"spectrogram": "cqt",
"spectrogram_top_db": 80,
"spectrogram_filter_scale": 1.0,
"spectrogram_undertone_stacking": 1,
"spectrogram_overtone_stacking": 5,
"cut_context": 1,
"architecture": "deep_hcnn",
"filters": 16,
"stacks": 10,
"conv_range": 3,
"undertone_stacking": 0,
"overtone_stacking": 1,
"activation": "relu",
"conv_ctx": [1],
"dilations": [1],
"last_conv_kernel": [1, 1],
"residual_hop": 1,
"residual_end": 0,
"residual_op": "add",
"batchnorm": 0,
"dropout": 0.3,
"specaugment_prob": 0.0,
"specaugment_freq_mask_num": 2,
"specaugment_freq_mask_max": 27,
"specaugment_time_mask_num": 1,
"specaugment_time_mask_max": 5,
"voicing": 0,
"voicing_input": "spectrogram_salience",
}
specified_args = common.argument_defaults(args, defaults)
common.name(args, specified_args, "spctrgrm")
return args
def parse_args(argv):
parser = common.common_arguments_parser()
# Model specific arguments
# input
parser.add_argument("--spectrogram", type=str, help="Spectrogram method")
parser.add_argument("--spectrogram_top_db",
type=float,
help="Spectrogram top_db")
parser.add_argument("--spectrogram_filter_scale",
type=float,
help="Spectrogram filter_scale")
parser.add_argument("--spectrogram_undertone_stacking",
type=int,
help="spectrogram undertone stacking")
parser.add_argument("--spectrogram_overtone_stacking",
type=int,
help="spectrogram overtone stacking")
parser.add_argument(
"--cut_context",
type=int,
help="Cut unnecessary context, doesn't work with dilations!")
# model
parser.add_argument("--architecture", type=str, help="Model architecture")
parser.add_argument("--faster_hcnn", type=int, help="HCNN implementation")
parser.add_argument("--use_bias", type=int, help="use bias in conv2d")
parser.add_argument("--class_weighting",
type=int,
help="use class weighting")
parser.add_argument("--filters", type=int, help="Filters in convolutions")
parser.add_argument("--stacks", type=int, help="Stacks")
parser.add_argument("--conv_range",
type=int,
help="Stack kernel size in frequency axis")
parser.add_argument("--undertone_stacking",
type=int,
help="Undertone stacking in the model")
parser.add_argument("--overtone_stacking",
type=int,
help="Overtone stacking in the model")
parser.add_argument("--stacking_until",
type=int,
help="Harmonic stacking in the model until Nth layer")
parser.add_argument("--activation",
type=str,
help="Activation function for the convolution stack")
# context
parser.add_argument("--conv_ctx",
nargs="+",
type=int,
help="Stack kernel sizes in time axis")
parser.add_argument("--dilations",
nargs="+",
type=int,
help="Dilation rate for the convolutions")
parser.add_argument("--last_conv_kernel", nargs=2, type=int)
parser.add_argument("--last_pooling", type=str)
# residual
parser.add_argument(
"--residual_hop",
type=int,
help="Size of one block around which there is a residual connection")
parser.add_argument("--residual_end",
type=int,
help="No residual connection in last N layers")
parser.add_argument(
"--residual_op",
type=str,
help=
"Residual connection operation (add for ResNet, concat for DenseNet)")
# regularization
parser.add_argument("--batchnorm", type=int)
parser.add_argument("--dropout", type=float)
args = parser.parse_args(argv)
# hop_length = 256
hop_length = 512 # FRAME-LEVEL INSTRUMENT RECOGNITION BY TIMBRE AND PITCH
defaults = {
# Change some of the common defaults
"samplerate": 44100,
"context_width": 4 * hop_length,
"annotations_per_window": 1,
"hop_size": 1,
"frame_width": hop_length,
"note_range": 11,
"min_note": 0,
"evaluate_every": 30000,
"evaluate_small_every": 1000,
"annotation_smoothing": 0.0,
"batch_size": 32,
"bins_per_semitone": 1,
"unvoiced_loss_weight": 1.0,
"datasets": ["musicnet_mir"],
# Model specific defaults
"learning_rate_decay_steps": 10000,
"learning_rate_decay": 0.8,
"spectrogram": "YunNingHung_cqt",
"spectrogram_top_db": 110,
"spectrogram_filter_scale": 1.0,
"spectrogram_undertone_stacking": 1,
"spectrogram_overtone_stacking": 5,
"cut_context": 1,
"architecture": "baseline",
"faster_hcnn": 0,
"use_bias": 1,
"class_weighting": 1,
"filters": 12,
"stacks": 6,
"conv_range": 3,
"undertone_stacking": 1,
"overtone_stacking": 3,
"stacking_until": 999,
"activation": "relu",
"conv_ctx": [1],
"dilations": [1],
"last_conv_kernel": [1, 72],
"last_pooling": "avg",
"residual_hop": 1,
"residual_end": 0,
"residual_op": "add",
"batchnorm": 0,
"dropout": 0.3,
}
specified_args = common.argument_defaults(args, defaults)
common.name(args, specified_args, "mir")
return args
def parse_args(argv):
parser = common.common_arguments_parser()
# Model specific arguments
parser.add_argument("--use_biases",
action='store_true',
default=False,
help="Use biases in the convolutions")
parser.add_argument("--input_normalization",
type=int,
help="Enable normalizing each input example")
parser.add_argument("--initial_filter_width",
type=int,
help="First conv layer filter width")
parser.add_argument("--initial_filter_padding",
type=str,
help="First conv layer padding")
parser.add_argument("--filter_width",
type=int,
help="Dilation stack filter width (2 or 3)")
parser.add_argument("--skip_channels", type=int, help="Skip channels")
parser.add_argument("--residual_channels",
type=int,
help="Residual channels")
parser.add_argument("--stack_number",
type=int,
help="Number of dilated stacks")
parser.add_argument("--max_dilation",
type=int,
help="Maximum dilation rate")
parser.add_argument("--dilation_layer_dropout",
type=float,
help="Dropout in dilation layer")
parser.add_argument("--skip_layer_dropout",
type=float,
help="Dropout in skip connections")
parser.add_argument("--skip", type=str, help="Skip add or concat")
parser.add_argument("--postprocessing",
type=str,
help="Postprocessing layer")
args = parser.parse_args(argv)
defaults = {
"note_range":
72,
"min_note":
24,
"evaluate_every":
5000,
"evaluate_small_every":
5000,
"batch_size":
8,
"annotations_per_window":
10,
"context_width":
94,
"annotation_smoothing":
0.18,
"input_normalization":
1,
"initial_filter_width":
32,
"initial_filter_padding":
"same",
"filter_width":
3,
"skip_channels":
64,
"residual_channels":
32,
"stack_number":
1,
"max_dilation":
512,
"dilation_layer_dropout":
0.0,
"skip_layer_dropout":
0.0,
"skip":
"add",
"postprocessing":
"avgpool_p93_s93_Psame--conv_f256_k16_s8_Psame_arelu--conv_f256_k16_s8_Psame_arelu",
}
specified_args = common.argument_defaults(args, defaults)
common.name(args, specified_args, "wavenet")
return args
def parse_args(argv):
parser = common.common_arguments_parser()
# Model specific arguments
# input
parser.add_argument("--spectrogram", type=str, help="Spectrogram method")
# model
parser.add_argument("--architecture", type=str, help="Model architecture")
parser.add_argument("--filters", type=int, help="Filters in convolutions")
parser.add_argument("--spectrogram_undertone_stacking",
type=int,
help="Undertone stacking in the spectrogram")
parser.add_argument("--spectrogram_overtone_stacking",
type=int,
help="Overtone stacking in the spectrogram")
parser.add_argument("--undertone_stacking",
type=int,
help="Undertone stacking in the model")
parser.add_argument("--overtone_stacking",
type=int,
help="Overtone stacking in the model")
parser.add_argument("--activation",
type=str,
help="Activation function for the convolution stack")
# regularization
parser.add_argument("--dropout", type=float)
args = parser.parse_args(argv)
hop_length = 441 * 4 # 25 fps
# context_width: 10*hop_length
defaults = {
# Change some of the common defaults
"samplerate": 44100,
"context_width": 2 * hop_length,
"annotations_per_window": 1,
"hop_size": 1,
"frame_width": hop_length,
"note_range": 88,
"min_note": 21,
"evaluate_every": 5000,
"evaluate_small_every": 1000,
"annotation_smoothing": 0.0,
"batch_size": 128,
"batch_size_evaluation": 1024,
"bins_per_semitone": 1,
"datasets": ["maps"],
# Model specific defaults
"learning_rate_decay_steps": 10000,
"learning_rate_decay": 0.8,
"spectrogram": "kelz",
"architecture": "allconv",
"filters": 16,
"undertone_stacking": 0,
"overtone_stacking": 1,
"spectrogram_undertone_stacking": 0,
"spectrogram_overtone_stacking": 1,
"activation": "relu",
"dropout": 0.25,
}
specified_args = common.argument_defaults(args, defaults)
common.name(args, specified_args, "kelz")
return args