def test_create_training_dataset(dataset_name, settings):
dir_path = annotator_utils.get_dataset_dir_path(settings.clip_type,
dataset_name)
dataset = dataset_utils.create_training_dataset(dir_path, settings)
show_training_dataset_stats(dataset)
def test_create_waveform_dataset_from_tfrecord_files(dataset_name, settings):
dir_path = annotator_utils.get_dataset_dir_path(settings.clip_type,
dataset_name)
dataset = dataset_utils.create_waveform_dataset_from_tfrecord_files(
dir_path)
show_waveform_dataset_stats(dataset, settings.waveform_sample_rate)
def evaluate_annotator(training_name, epoch_num):
_, settings = annotator_utils.load_model_and_settings(
training_name, epoch_num)
dir_path = annotator_utils.get_dataset_dir_path(settings.clip_type,
'Validation')
dataset = dataset_utils.create_validation_dataset(dir_path, settings)
dataset = dataset.take(settings.validation_step_count)
inferrer = Inferrer((training_name, epoch_num))
bounds = inferrer.get_call_bounds(dataset)
start_diff_counts = defaultdict(int)
end_diff_counts = defaultdict(int)
for (inferred_start_index, inferred_end_index, dataset_start_index,
dataset_end_index) in bounds:
dataset_start_index = dataset_start_index.numpy()
dataset_end_index = dataset_end_index.numpy()
sample_rate = settings.waveform_sample_rate
start_diff = _get_diff(inferred_start_index, dataset_start_index,
sample_rate)
end_diff = _get_diff(inferred_end_index, dataset_end_index,
sample_rate)
if start_diff is not None:
start_diff_counts[start_diff] += 1
end_diff_counts[end_diff] += 1
# print(
# start_diff, end_diff,
# inferred_start_index, inferred_end_index,
# dataset_start_index, dataset_end_index)
_show_diff_counts('Start', start_diff_counts, settings)
_show_diff_counts('End', end_diff_counts, settings)
_plot_diff_counts(training_name, epoch_num, start_diff_counts,
end_diff_counts, settings)
def show_dataset_sizes(settings):
from tensorflow.data import TFRecordDataset
for dataset_name in ('Training', 'Validation'):
total_size = 0
print(f'Sizes of files in dataset "{dataset_name}":')
dir_path = annotator_utils.get_dataset_dir_path(
settings.clip_type, dataset_name)
file_paths = sorted(dir_path.glob('*.tfrecords'))
for file_path in file_paths:
dataset = TFRecordDataset([str(file_path)])
size = 0
for _ in dataset:
size += 1
print(f' {file_path.name}: {size}')
total_size += size
print(f'Total size of dataset "{dataset_name}": {total_size}')
def get_dataset(name, settings):
dir_path = annotator_utils.get_dataset_dir_path(settings.clip_type, name)
return dataset_utils.create_training_dataset(dir_path, settings)