def main_train():
# train
args = parse_args()
config_args, device, checkpoint = transform_args(args)
train(config_args, device, checkpoint, dataplace=args.dataplace)
# test at best epoch
best_checkpoint = misc.get_checkpoint(
output_folder=config_args['training']['output_folder'],
epoch="best",
)
evaluate(
config_args=config_args,
device=device,
checkpoint=best_checkpoint,
tempscale=False,
corruptions=False,
dataplace=args.dataplace
)
LOGGER.error(f"Finish: {config_args['training']['config_path']}")
def test_evaluate(deepspeech: DeepSpeech, generator: Iterable,
store_path: str) -> pd.DataFrame:
metrics = evaluate(deepspeech,
generator,
save_activations=True,
store_path=store_path)
with pd.HDFStore(store_path, mode='r') as store:
references = store['references']
assert len(references) == len(metrics) == 12
assert all(references.columns.values == np.array(
['transcript', 'prediction', 'wer', 'cer']))
def main(store_path: str, features_store_path: str, batch_size: int,
save_activations: bool, mask: bool, mask_F: int, mask_mf: int,
mask_T: int, mask_mt: int, mask_ratio_t: float):
""" Evaluate model using prepared features. """
deepspeech = load_extended_model(CONFIG_PATH, ALPHABET_PATH, WEIGHTS_PATH)
generator = DataGenerator.from_prepared_features(
features_store_path,
alphabet=deepspeech.alphabet,
features_extractor=deepspeech.features_extractor,
batch_size=batch_size,
mask=mask,
mask_params=dict(F=mask_F,
mf=mask_mf,
T=mask_T,
mt=mask_mt,
ratio_t=mask_ratio_t))
units = calculate_units(deepspeech.model)
logger.info(f'Model contains: {units//1e6:.0f}M units ({units})')
metrics = evaluate(deepspeech, generator, save_activations, store_path)
logger.info(f'Mean CER: {metrics.cer.mean():.4f}')
logger.info(f'Mean WER: {metrics.wer.mean():.4f}')
for k in to_evaluate.keys():
print(k)
base_data_path = k
for change, movement_data_path in to_evaluate[k]:
real_change = {
"x": 0,
"y": 0,
"z": 0,
"roll": 0,
"pitch": 0,
"yaw": 0,
}
for k in change.keys():
real_change[k] = change[k]
left_side = "x y z: ({} {} {}) φ θ ψ: ({} {} {}),".format(
real_change['x'], real_change['y'], real_change['z'],
real_change['roll'], real_change['pitch'], real_change['yaw'])
results = evaluate(base_data_path,
movement_data_path,
settings,
real_change,
CALCULATE_ERROR=False)
f.write(left_side + ",".join(["{:.3f}".format(x)
for x in results]) + "\n")
f.flush()
end = time()
print("Took:", end - start, "seconds")
print("done")
if __name__ == '__main__':
if args.train_ubm:
data = prepare_data.load_train_file()
ubm = classify.fit_ubm(data, args.save_path)
classify.fit_adap(data, ubm, args.save_path)
args.load_path = args.save_path
if args.train:
data = prepare_data.load_train_file()
classify.fit(data, args.save_path)
args.load_path = args.save_path
if args.validate_ubm:
data = prepare_data.load_test('val')
scores = classify.predict_ubm(data, args.load_path)
Y, Y_pred = evaluate.get_predictions(scores)
evaluate.evaluate(Y, Y_pred)
if args.confusion_matrix:
evaluate.confusion_matrix(Y, Y_pred)
if args.test_ubm:
data = prepare_data.load_test('test')
scores = classify.predict_ubm(data, args.load_path)
Y, Y_pred = evaluate.get_predictions(scores)
evaluate.evaluate(Y, Y_pred)
if args.confusion_matrix:
evaluate.confusion_matrix(Y, Y_pred)
if args.validate:
data = prepare_data.load_test('val')
scores = classify.predict(data, args.load_path)
Y, Y_pred = evaluate.get_predictions(scores)
evaluate.evaluate(Y, Y_pred)
if args.confusion_matrix: