def fit_model_mix_shuffle(
experiment_name
): # mix folds of AudioSet and LiveSet train tensors and fit
path_to_liveset = os.path.join('auxiliary_files', 'dataset',
'live_set_genres.pkl')
# load audioset
sets = dg.get_generated_sample(kd.AUDIOSET, [7, 2, 1])
x_tr_a, y_tr_a = sets['train']
x_val_a, y_val_a = sets['val']
# load liveset
with open(path_to_liveset, "rb") as handle:
dataset_dict = pickle.load(handle)
x_tr_l, y_tr_l, _ = dataset_dict[
'train'] # the last tuple unit is genre labels
x_val_l, y_val_l, _ = dataset_dict['valid']
# concat datasets
folds_count = 10
x_train, y_train = mix_folds(x_tr_a, x_tr_l, y_tr_a, y_tr_l, folds_count)
x_valid, y_valid = mix_folds(x_val_a, x_val_l, y_val_a, y_val_l,
folds_count)
# checkpoints
checkpoints_file = os.path.join('auxiliary_files', 'checkpoints',
'fullfit', experiment_name)
if not os.path.isdir(checkpoints_file):
os.makedirs(checkpoints_file, exist_ok=True)
checkpoints_file = os.path.join(checkpoints_file, 'cpt.h5')
callback_list = mix_model_callbacks(checkpoints_file)
segmentator = sg() # init model
history = segmentator.exec_fit(
x_train,
x_valid,
y_train,
y_valid,
checkpoints_file,
epochs=0,
callback_list=callback_list) # 0 epochs for early stopping
# save history
history_file = os.path.join('auxiliary_files', 'history', 'new',
experiment_name)
if not os.path.isdir(history_file):
os.makedirs(history_file, exist_ok=True)
history_file = os.path.join(history_file, 'history.txt')
with open(history_file, 'w') as f:
print(history.history, file=f)
metrics_dir = os.path.join('auxiliary_files', 'eval', 'new',
experiment_name, 'metrics')
if not os.path.isdir(metrics_dir):
os.makedirs(metrics_dir, exist_ok=True)
segmentator.evaluate(x_val_l, y_val_l, metrics_dir, plot_time_clamp=2000)
def dual_fit(experiment_name
): # fit on AudioSet, don't fix any weights, fit on LiveSet
path_to_liveset = os.path.join('auxiliary_files', 'dataset',
'live_set_genres.pkl')
segmentator = sg() # init model
fit_model_audioset(segmentator, experiment_name)
fit_model_live(segmentator, path_to_liveset, experiment_name)
def full_fit_pipeline(
experiment_name): # fit on AudioSet, fix GRU & Dense, fit on LiveSet
path_to_liveset = os.path.join('auxiliary_files', 'dataset',
'live_set_genres.pkl')
segmentator = sg() # init model
fit_model_audioset(segmentator, experiment_name)
fix_model_params(segmentator)
fit_model_live(segmentator, path_to_liveset, experiment_name)
if __name__ == "__main__":
# preprocesing live samples ang get embeddings
pp.preprocess_train(PATH_TO_META_FOLDER,
PATH_TO_VIDEO_FOLDER,
PATH_TO_WAV_FOLDER,
seq_len=96)
fe.get_audioset_features(PATH_TO_LIVE_DATA,
PATH_TO_LIVE_DATA_WITH_EMBEDDINGS)
# generate audioset and liveset samples
sets = dg.get_generated_sample(kd.AUDIOSET, [7, 3, 1])
x_tr, y_tr = sets['train']
x_val, y_val = sets['val']
x_test, y_test = sets['test']
sets_l = dg.get_generated_sample(
kd.LIVE, [7, 3, 1],
path_to_live_data=PATH_TO_LIVE_DATA_WITH_EMBEDDINGS)
x_tr_l, y_tr_l = sets_l['train']
x_val_l, y_val_l = sets_l['val']
x_test_l, y_test_l = sets_l['test']
# create model
model = sg()
# train model and evaluate
model.exec_fit(x_tr, x_val, y_tr, y_val, PATH_TO_CHECKPOINT_FILE)
model.evaluate(x_test_l, y_test_l, PATH_TO_METRIC_WITHOUT_TR)
# additional train and evaluate
model.exec_fit(x_tr_l, x_val_l, y_tr_l, y_val_l, PATH_TO_CHECKPOINT_FILE)
model.evaluate(x_test_l, y_test_l, PATH_TO_METRIC_WITH_TR)