format='%(asctime)s,%(msecs)d %(levelname)-8s [%(filename)s:%(lineno)d] %(message)s',
datefmt='%Y-%m-%d:%H:%M:%S', level=logging.INFO)
# make sure this is set to WHAM root directory
WHAM_ROOT = os.getenv("WHAM_ROOT")
NUM_WORKERS = multiprocessing.cpu_count() // 4
OUTPUT_DIR = os.path.expanduser('~/.nussl/recipes/ideal_ratio_mask/')
RESULTS_DIR = os.path.join(OUTPUT_DIR, 'results')
# APPROACH, KWARGS = 'psa', {'range_min': -np.inf, 'range_max':np.inf}
APPROACH, KWARGS = 'msa', {}
shutil.rmtree(os.path.join(RESULTS_DIR), ignore_errors=True)
os.makedirs(RESULTS_DIR, exist_ok=True)
test_dataset = datasets.WHAM(WHAM_ROOT, sample_rate=8000, split='tt')
def separate_and_evaluate(item_):
separator = separation.benchmark.IdealRatioMask(
item_['mix'], item_['sources'], approach=APPROACH,
mask_type='soft', **KWARGS)
estimates = separator()
evaluator = evaluation.BSSEvalScale(
list(item_['sources'].values()), estimates, compute_permutation=True)
scores = evaluator.evaluate()
output_path = os.path.join(RESULTS_DIR, f"{item_['mix'].file_name}.json")
with open(output_path, 'w') as f:
json.dump(scores, f)
400) # get 400 frame excerpts (3.2 seconds)
])
return tfm
def cache_dataset(_dataset):
cache_dataloader = torch.utils.data.DataLoader(_dataset,
num_workers=NUM_WORKERS,
batch_size=BATCH_SIZE)
ml.train.cache_dataset(cache_dataloader)
_dataset.cache_populated = True
tfm = construct_transforms(os.path.join(CACHE_ROOT, 'tr'))
dataset = datasets.WHAM(WHAM_ROOT,
split='tr',
transform=tfm,
cache_populated=CACHE_POPULATED)
tfm = construct_transforms(os.path.join(CACHE_ROOT, 'cv'))
val_dataset = datasets.WHAM(WHAM_ROOT,
split='cv',
transform=tfm,
cache_populated=CACHE_POPULATED)
if not CACHE_POPULATED:
# cache datasets for speed
cache_dataset(dataset)
cache_dataset(val_dataset)
# ----------------------------------------------------
# -------------------- TRAINING ----------------------