File size: 9,212 Bytes
162fa19 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 | import yaml
import random
import torch
import torchaudio
import numpy as np
from pathlib import Path
from collections import Counter
def select_inference_device() -> torch.device:
"""Prefer CUDA, then Apple MPS, else CPU."""
if torch.cuda.is_available():
return torch.device("cuda")
mps_backend = getattr(torch.backends, "mps", None)
if mps_backend is not None and mps_backend.is_available():
return torch.device("mps")
return torch.device("cpu")
def empty_accelerator_cache(device: torch.device) -> None:
"""Release GPU / MPS cache between batches when supported."""
if device.type == "cuda" and torch.cuda.is_available():
torch.cuda.empty_cache()
elif device.type == "mps" and hasattr(torch, "mps") and torch.mps.is_available():
torch.mps.empty_cache()
def _causal_mask(seq_len: int, device=None):
"""Causal (triangular) mask for decoder self-attention.
Returns (seq_len, seq_len) bool tensor where True = masked (cannot attend).
Position i cannot attend to position j when j > i.
"""
return torch.triu(torch.ones(seq_len, seq_len, device=device, dtype=torch.bool), diagonal=1)
def _key_padding_mask_from_lengths(lengths, seq_len, device=None):
"""Key padding mask for PyTorch Transformer (True = ignore/padding).
lengths: (N,) valid length per batch. Returns (N, seq_len) bool."""
if device is None and isinstance(lengths, torch.Tensor):
device = lengths.device
lengths_t = lengths.detach().clone() if isinstance(lengths, torch.Tensor) else torch.tensor(lengths, device=device)
# True where position >= length (i.e. padding)
return torch.arange(seq_len, device=device).unsqueeze(0) >= lengths_t.unsqueeze(1)
def create_mask_plain(tgt_seq_len, tgt_lengths=None, device=None):
"""Create causal mask and optional key padding mask (PyTorch convention).
Returns:
tgt_mask: (tgt_seq_len, tgt_seq_len) bool, True = masked
tgt_padding_mask: (N, tgt_seq_len) bool, True = padding; or None if tgt_lengths is None
"""
tgt_mask = _causal_mask(tgt_seq_len, device)
if tgt_lengths is None:
return tgt_mask, None
tgt_lengths_t = tgt_lengths.detach().clone() if isinstance(tgt_lengths, torch.Tensor) else torch.tensor(tgt_lengths, device=device)
tgt_padding_mask = _key_padding_mask_from_lengths(tgt_lengths_t, tgt_seq_len, device)
return tgt_mask, tgt_padding_mask
def create_bool_matrix(lengths, max_len=None, mask_prob=0.0):
"""
Create a `bool_matrix` based on `lengths`,
and add functionality to optionally mask tokens at positions 5k+3 with probability `mask_prob`.
Arguments:
lengths: List or PyTorch Tensor (N,) indicating the valid length for each sequence.
max_len: Maximum sequence length (default is lengths.max()).
mask_prob: Probability to mask tokens at positions 5k+3 (default: 95%).
Returns:
bool_matrix: PyTorch boolean mask of shape (N, max_len)
"""
# Convert the input list to a tensor.
if isinstance(lengths, list):
lengths = torch.tensor(lengths, dtype=torch.int64)
max_len = max_len or lengths.max()
batch_size = lengths.size(0)
with torch.no_grad(): # Avoid gradient tracking for a cheaper mask computation.
indices = torch.arange(max_len, device=lengths.device)
bool_matrix = (indices.view(1, -1) < lengths.view(-1, 1)) # Standard length mask.
# Identify instrument-token positions (every 5k+3 slot).
mask_positions = (indices % 5 == 3).view(1, -1)
# Randomly mask instrument-token positions with probability mask_prob.
random_mask = torch.bernoulli(torch.full(
(batch_size, max_len), mask_prob, dtype=torch.float32, device=lengths.device
)).bool()
bool_matrix = bool_matrix & ~(mask_positions & random_mask)
return bool_matrix
def create_mask(tgt_seq_len, tgt_lengths=None, device=None, inst_mask_proba=0.0):
"""Create causal mask and padding masks for full sequence (with inst masking).
Returns:
tgt_mask: (tgt_seq_len-1, tgt_seq_len-1) causal bool mask
tgt_input_padding_mask: (N, tgt_seq_len-1) bool, True = padding (PyTorch convention)
tgt_out_padding_mask: (N, tgt_seq_len-1) bool, True = valid (for loss masking)
"""
seq_len = tgt_seq_len - 1
tgt_mask = _causal_mask(seq_len, device)
if tgt_lengths is None:
return tgt_mask, None, None
tgt_bool_matrix = create_bool_matrix(tgt_lengths, mask_prob=inst_mask_proba)
tgt_input_valid = tgt_bool_matrix[:, :-1]
tgt_out_padding_mask = tgt_bool_matrix[:, 1:]
# PyTorch key_padding_mask: True = ignore
tgt_input_padding_mask = ~tgt_input_valid
return tgt_mask, tgt_input_padding_mask, tgt_out_padding_mask
def calc_separate_idx(total_files, separate, sepa_tgt, manual_file_range):
if manual_file_range is not None:
return np.array([manual_file_range]), [0, 1]
if sepa_tgt is None:
sepa_tgt = [0, separate]
files_per_hdffile = total_files // separate
file_range = np.zeros([separate, 2], dtype=int)
for i in range(separate):
file_range[i, 0] = i * files_per_hdffile
if i < separate - 1:
file_range[i, 1] = file_range[i, 0] + files_per_hdffile
else:
file_range[i, 1] = total_files
return file_range, sepa_tgt
def myrelu(x, thr=0, rev=False):
if rev:
return x if x < thr else thr
else:
return x if x > thr else thr
def save_to_hdf(hf, inst_group, wav_seg, tokens):
if inst_group is not None:
if inst_group not in hf:
group = hf.create_group(inst_group)
else:
group = hf[inst_group]
else:
group = hf
seg_num = len(group.keys())
seg_group = group.create_group(f"seg{seg_num}")
seg_group.create_dataset("wav", data=wav_seg)
if tokens is not None:
seg_group.create_dataset("midi", data=tokens)
def save_to_hdf_variations(hf, inst_group, wav_seg, tokens):
if inst_group is not None:
if inst_group not in hf:
group = hf.create_group(inst_group)
else:
group = hf[inst_group]
else:
group = hf
seg_num = len(group.keys())
seg_group = group.create_group(f"seg{seg_num}")
seg_group.create_dataset("wav", data=wav_seg)
if tokens is not None:
for i in range(3):
seg_group.create_dataset(f"midi{i}", data=tokens[i])
def save_tokens_hdf_variations(hf, inst_group, tokens):
if inst_group is not None:
if inst_group not in hf:
group = hf.create_group(inst_group)
else:
group = hf[inst_group]
else:
group = hf
seg_num = len(group.keys())
seg_group = group.create_group(f"seg{seg_num}")
if tokens is not None:
for i in range(3):
seg_group.create_dataset(f"midi{i}", data=tokens[i])
def file_shuffle(files, seed_fix):
files = sorted(files, key=lambda x: str(Path(x)))
if seed_fix:
random.seed(1)
random.shuffle(files)
random.seed(None)
return files
def debug_message(msg):
with open("debug.txt", "w") as file:
file.write(f"{msg}\n")
assert False, "Debug file created !"
def my_vstack(array1, array2):
if len(array1) == 0:
return array2
if len(array2) == 0:
return array1
else:
return np.vstack((array1, array2))
def get_random_mode(lst): # Return a mode, breaking ties uniformly at random.
counter = Counter(lst)
max_count = max(counter.values())
modes = [key for key, count in counter.items() if count == max_count]
return random.choice(modes)
def create_one_hot(length, index):
one_hot = torch.zeros(length, dtype=torch.float32)
one_hot[index] = 1.0
return one_hot
def pad_arrays(array1, array2):
len1, len2 = len(array1), len(array2)
if len1 > len2:
array2 = np.pad(array2, (0, len1 - len2), mode="constant")
elif len2 > len1:
array1 = np.pad(array1, (0, len2 - len1), mode="constant")
return array1, array2
def load_audio(path, target_sample_rate: int):
# load audio and convert to mono
waveform, sample_rate = torchaudio.load(path)
if waveform.shape[0] > 1:
waveform = torch.mean(waveform, dim=0, keepdim=True)
# resample audio according to target sample rate
if sample_rate != target_sample_rate:
resampler = torchaudio.transforms.Resample(sample_rate, target_sample_rate)
waveform = resampler(waveform)
return waveform
def recursive_merge(default_cfg, experiment_cfg):
for key, value in experiment_cfg.items():
if isinstance(value, dict):
default_cfg[key] = recursive_merge(default_cfg[key], value)
else:
default_cfg[key] = value
return default_cfg
def load_yaml(path):
with open(path, "r") as file:
return yaml.safe_load(file)
def draw_from_normal_distribution(std: float, mean: float, high_bound: float, low_bound: float):
return torch.clamp(
torch.clamp(torch.randn(1) * std + mean, -1.0, 1.0).abs() * high_bound, low_bound, high_bound
).item()
|