curriculum_preset.py

from note_attribute_repr import Sampler
import numpy as np


"""
default pad lengths
"""
pad_length100 = 100

"""
default R_base to R_fac conversion
"""
eo_dist0 = Sampler(0, 4, np.array([0.8, 0.08, 0.07, 0.05]))

ep_dist0 = None  # i.e., always estimate ep

w_dist0 = Sampler(-3, 4,
                  np.array([0.05, 0.05, 0.15, 0.5, 0.15, 0.05, 0.05]))

default_autoenc_dict = {
    'nmat_pad_length': pad_length100,
    'atr_mat_pad_length': pad_length100,
    'estimate_ep': True,
    'eo_dist': eo_dist0,
    'ep_dist': ep_dist0,
    'w_dist': w_dist0
}


"""
Preset corrupters
"""

# For light corruption where o_bt and p_hig are not corrupted.
default_corrupter_dict = {
    'corrupt_col_ids': (1, 3, 4, 5, 6),
    'pad_length': pad_length100,
    'mask_ratio': 0.15,
    'unchange_ratio': 0.1,
    'unknown_ratio': 0.8,
    'relmat_cpt_ratio': 0.3
}

# For pre-trained settings where all attributes are masked.
all_corrupter_dict = {
    'corrupt_col_ids': (0, 1, 2, 3, 4, 5, 6),
    'pad_length': pad_length100,
    'mask_ratio': 0.15,
    'unchange_ratio': 0.1,
    'unknown_ratio': 0.8,
    'relmat_cpt_ratio': 0.3
}


# Onset attributes masked: for fine-tuning onset predictor.
onset_corrupter_dict = {
    'corrupt_col_ids': (0, 1),
    'pad_length': pad_length100,
    'mask_ratio': 0.15,
    'unchange_ratio': 0.1,
    'unknown_ratio': 0.8,
    'relmat_cpt_ratio': 0.5
}


# Pitch attributes masked: for fine-tuning pitch predictor.
pitch_corrupter_dict = {
    'corrupt_col_ids': (2, 3, 4),
    'pad_length': pad_length100,
    'mask_ratio': 0.15,
    'unchange_ratio': 0.1,
    'unknown_ratio': 0.8,
    'relmat_cpt_ratio': 0.5
}


# Duration attributes masked: for fine-tuning duration predictor.
duration_corrupter_dict = {
    'corrupt_col_ids': (5, 6),
    'pad_length': pad_length100,
    'mask_ratio': 0.15,
    'unchange_ratio': 0.1,
    'unknown_ratio': 0.8,
    'relmat_cpt_ratio': 0.5
}


"""
Preset MuseBERT parameters
"""

# used with default corrupter
default_model_dict = {
    'loss_inds': (1, 3, 4, 5, 6)
}

# used with pre-training
all_model_dict = {
    'loss_inds': (0, 1, 2, 3, 4, 5, 6),
}

# used with onset, pitch and duration fine-tuning.
onset_model_dict = {'loss_inds': (0, 1)}
pitch_model_dict = {'loss_inds': (2, 3, 4)}
duration_model_dict = {'loss_inds': (5, 6)}


"""
Preset learning rate parameters
"""

# pre-training parameters
default_lr_dict = {
    'lr': 5e-4,
    'final_lr_factor': 1e-2,
    'warmup': 15000,
    'n_epoch': 100,
    'final_epoch': 60
}


# fine-tuning parameters
ft_lr_dict = {
    'lr': 3e-4,
    'final_lr_factor': 1e-2,
    'warmup': 10000,
    'n_epoch': 100,
    'final_epoch': 60
}


"""
Preset training parameters
"""


# used with default corrupter
default_train_dict = {
    'batch_size': 128,
    'beta': (0, 1, 0, 1, 1, 0.1, 0.1),
    'lr_dict': default_lr_dict
}


# used at pre-traning
all_train_dict = {
    'batch_size': 128,
    'beta': (1, 1, 1, 1, 1, 0.1, 0.1),
    'lr_dict': default_lr_dict
}


# used at general fine-tuning
all_train_ft_dict = {
    'batch_size': 128,
    'beta': (1, 1, 1, 1, 1, 0.1, 0.1),
    'lr_dict': ft_lr_dict
}


# used for fine-tuning specific attributes
onset_train_ft_dict = {
    'batch_size': 128,
    'beta': (1, 1, 0, 0, 0, 0, 0),
    'lr_dict': ft_lr_dict
}


pitch_train_ft_dict = {
    'batch_size': 128,
    'beta': (0, 0, 1, 1, 1, 0, 0),
    'lr_dict': ft_lr_dict
}


duration_train_ft_dict = {
    'batch_size': 128,
    'beta': (0, 0, 0, 0, 0, 1, 1),
    'lr_dict': ft_lr_dict
}