def extract(dataset, recompute=False):
"""Extract feature vectors from the given dataset.
Args:
dataset: Dataset to extract features from.
recompute (bool): Whether to recompute existing features.
"""
import features
# Use a logmel representation for feature extraction
extractor = features.LogmelExtractor(
cfg.sample_rate,
cfg.n_window,
cfg.hop_length,
cfg.n_mels,
)
# Ensure output directory exists and set file path
os.makedirs(cfg.extraction_path, exist_ok=True)
output_path = os.path.join(cfg.extraction_path, dataset.name + '.h5')
# Save free parameters to disk
utils.log_parameters(cfg.logmel,
os.path.join(cfg.extraction_path, 'parameters.json'))
# Extract features for each audio clip in the dataset
df = io.read_metadata(dataset.metadata_path)
features.extract_dataset(
dataset_path=dataset.path,
file_names=df.index.tolist(),
extractor=extractor,
output_path=output_path,
recompute=recompute,
)
def train():
"""Train the neural network model.
See Also:
:func:`training.train`
Note:
For reproducibility, the random seed is set to a fixed value.
"""
import training
# Ensure output directories exist
os.makedirs(os.path.dirname(cfg.scaler_path), exist_ok=True)
os.makedirs(cfg.model_path, exist_ok=True)
os.makedirs(cfg.log_path, exist_ok=True)
# Load (standardized) input data and target values
tr_x, tr_y, _ = _load_data(cfg.training_set, is_training=True)
val_x, val_y, _ = _load_data(cfg.validation_set)
# Try to create reproducible results
np.random.seed(cfg.initial_seed)
# Save free parameters to disk
utils.log_parameters(cfg.training, os.path.join(cfg.model_path,
'parameters.json'))
training.train(tr_x, tr_y, val_x, val_y)
def predict(dataset, fold):
"""Generate predictions for audio tagging.
This function uses an ensemble of trained models to generate the
predictions, with the averaging function being an arithmetic mean.
Computed predictions are then saved to disk.
Args:
dataset: Dataset to generate predictions for.
fold (int): The specific fold to generate predictions for. Only
applicable for the training dataset.
"""
import inference
# Load input data and associated metadata
x, df = _load_data(dataset)
dataset_name = dataset.name
if dataset.name == 'training':
if fold == -1:
raise ValueError('Invalid fold: %d' % fold)
dataset_name += str(fold)
mask = df.fold == fold
tr_x = x[~mask]
x = x[mask]
df = df[mask]
else:
tr_x, tr_df = _load_data(cfg.to_dataset('training'))
if fold >= 0:
dataset_name += str(fold)
tr_x = tr_x[tr_df.fold != fold]
generator = utils.fit_scaler(tr_x)
x = generator.standardize(x)
# Predict class probabilities for each model (epoch)
preds = []
for epoch in _determine_epochs(cfg.prediction_epochs, fold, n=4):
pred = utils.timeit(lambda: _load_model(fold, epoch).predict(x),
'[Epoch %d] Predicted class probabilities' % epoch)
preds.append(inference.merge_predictions(pred, df.index))
pred_mean = pd.concat(preds).groupby(level=0).mean()
# Ensure output directory exists and set file path format
os.makedirs(os.path.dirname(cfg.predictions_path), exist_ok=True)
predictions_path = cfg.predictions_path.format('%s', dataset_name)
# Save free parameters to disk
utils.log_parameters({'prediction_epochs': cfg.prediction_epochs},
os.path.join(os.path.dirname(cfg.predictions_path),
'parameters.json'))
# Write predictions to disk
pred_mean.to_csv(predictions_path % 'predictions')
io.write_predictions(pred_mean, predictions_path % 'submission')
def predict(dataset):
"""Generate predictions for audio tagging and sound event detection.
This function uses an ensemble of trained models to generate the
predictions, with the averaging function being an arithmetic mean.
Computed predictions are then saved to disk.
Args:
dataset: Dataset to generate predictions for.
"""
import capsnet
# Load (standardized) input data and associated file names
test_x, _, names = _load_data(dataset)
# Predict class probabilities for each model (epoch)
at_preds, sed_preds = [], []
for epoch in _determine_epochs(cfg.prediction_epochs):
model = _load_model(epoch)
at_pred, sed_pred = utils.timeit(
lambda: capsnet.gccaps_predict(test_x, model),
'[Epoch %d] Predicted class probabilities' % epoch)
at_preds.append(at_pred)
sed_preds.append(sed_pred)
# Average predictions to give an overall output
total_at_pred = np.mean(at_preds, axis=0)
total_sed_pred = np.mean(sed_preds, axis=0)
# Ensure output directory exists and set file path format
os.makedirs(os.path.dirname(cfg.predictions_path), exist_ok=True)
predictions_path = cfg.predictions_path.format('%s', dataset.name)
# Save free parameters to disk
utils.log_parameters({'prediction_epochs': cfg.prediction_epochs},
os.path.join(os.path.dirname(cfg.predictions_path),
'parameters.json'))
# Write predictions to disk
utils.write_predictions(names, total_at_pred, predictions_path % 'at')
utils.write_predictions(names, total_sed_pred, predictions_path % 'sed')
def extract(dataset):
"""Extract feature vectors from the given dataset.
Args:
dataset: Dataset to extract features from.
"""
import data_augmentation as aug
import features
# Use a logmel representation for feature extraction
extractor = features.LogmelExtractor(sample_rate=cfg.sample_rate,
n_window=cfg.n_window,
hop_length=cfg.hop_length,
n_mels=cfg.n_mels,
)
# Prepare for data augmentation if enabled
file_names, target_values = utils.read_metadata(dataset.metadata_path)
if dataset == cfg.training_set and cfg.enable_augmentation:
n_transforms_iter = aug.transform_counts(target_values)
file_names = aug.expand_metadata((file_names, target_values))[0]
else:
n_transforms_iter = None
# Ensure output directory exists and set file path
os.makedirs(cfg.extraction_path, exist_ok=True)
output_path = os.path.join(cfg.extraction_path, dataset.name + '.h5')
# Save free parameters to disk
utils.log_parameters(cfg.logmel, os.path.join(cfg.extraction_path,
'parameters.json'))
# Generate features for each audio clip in the dataset
features.extract_dataset(dataset.path,
file_names,
extractor,
cfg.clip_duration,
output_path,
n_transforms_iter=n_transforms_iter,
)
def train(model, fold, use_class_weight, noisy_sample_weight):
"""Train the neural network model.
Args:
model (str): The neural network architecture.
fold (int): The fold to use for validation.
use_class_weight (bool): Whether to use class-wise weights.
noisy_sample_weight (float): Examples that are not verified are
weighted according to this value.
Note:
For reproducibility, the random seed is set to a fixed value.
"""
import training
# Try to create reproducible results
np.random.seed(cfg.initial_seed)
# Load training data and associated metadata
x, df = _load_data(cfg.to_dataset('training'))
# Get one-hot representation of target values
y = utils.to_categorical(df.label)
# Split training data into training and validation
if fold >= 0:
mask = df.fold == fold
else:
mask = np.zeros(len(df), dtype=bool)
val_mask = mask & (df.manually_verified == 1)
tr_x = x[~mask]
tr_y = y[~mask]
val_x = x[val_mask]
val_y = y[val_mask]
val_index = df.index[val_mask]
# Compute class weights based on number of class examples
if use_class_weight:
group = utils.group_by_name(df)
n_examples = group.first().groupby('label').size().values
class_weight = len(group) / (len(n_examples) * n_examples)
else:
class_weight = None
# Assign a specific sample weight to unverified examples
if noisy_sample_weight:
sample_weight = df[~mask].manually_verified.values.astype(float)
sample_weight[sample_weight == 0] = noisy_sample_weight
else:
sample_weight = None
# Ensure output directories exist
fold_dir = str(fold) if fold >= 0 else 'all'
os.makedirs(os.path.join(cfg.model_path, fold_dir), exist_ok=True)
os.makedirs(cfg.log_path.format(fold_dir), exist_ok=True)
# Save free parameters to disk
utils.log_parameters(cfg.training,
os.path.join(cfg.model_path, 'parameters.json'))
training.train(tr_x,
tr_y,
val_x,
val_y,
val_index,
model,
fold,
class_weight=class_weight,
sample_weight=sample_weight)
