def fit_model(io: IO, model: Sequential, preprocessed: List[Preprocessed]):
epochs = io.get("epochs")
model.reset_states()
logline("splitting into training set and testing set ({}%)".format(
io.get("split")))
split = gen_split(preprocessed, io)
log_dir = "logs/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
for i in range(epochs):
logline("generating input and expected data for epoch {}/{}".format(
i + 1, epochs))
train_x, train_y = trim_params(gen_fit_params(split), io)
logline("training epoch {}/{}".format(i + 1, epochs))
callbacks = []
if io.get("profile"):
debug("profiling")
callbacks.append(
tf.keras.callbacks.TensorBoard(log_dir=log_dir,
histogram_freq=1))
model.fit(train_x,
train_y,
batch_size=io.get("batch_size"),
epochs=1,
shuffle=False,
callbacks=callbacks)
model.reset_states()
def output_split(all: List[Preprocessed], train: List[Preprocessed], io: IO):
obj = {
"training_set": list(map(lambda x: x.file_name, train)),
"test_set": list(map(lambda x: x.file_name, filter(lambda x: x not in train, all))),
}
with open(io.get("output_train"), "w+") as out_file:
json.dump(obj, out_file)
logline("wrote training/testing config to {}".format(io.get("output_train")))
def read_test_files(io: IO) -> List[Preprocessed]:
with open(io.get("input_preprocessed"), "rb") as preprocessed_file:
file_configs = pickle.load(preprocessed_file)
with open(io.get("input_train"), "rb") as train_config_file:
train_config = json.load(train_config_file)
test_files_names = train_config["test_set"]
preprocessed = map(lambda x: Preprocessed(x), file_configs)
test_files = list(
filter(lambda x: x.file_name in test_files_names,
preprocessed))
return test_files
def fit_model(io: IO, model: Sequential, preprocessed: List[Preprocessed]):
epochs = io.get("epochs")
model.reset_states()
logline("splitting into training set and testing set ({}%)".format(io.get("split")))
split = gen_split(preprocessed, io)
for i in range(epochs):
logline("generating input and expected data for epoch {}/{}".format(i + 1, epochs))
train_x, train_y = trim_params(gen_fit_params(split), io)
logline("training epoch {}/{}".format(i + 1, epochs))
model.fit(train_x, train_y, batch_size=io.get("batch_size"), epochs=1, shuffle=False)
model.reset_states()
def output_split(all: List[Preprocessed], train: List[Preprocessed], io: IO):
obj = {
"training_set":
list(map(lambda x: x.file_name, train)),
"test_set":
list(map(lambda x: x.file_name, filter(lambda x: x not in train,
all))),
}
pathlib.Path(os.path.dirname(io.get("output_train"))).mkdir(parents=True,
exist_ok=True)
with open(io.get("output_train"), "w+") as out_file:
json.dump(obj, out_file)
logline("wrote training/testing config to {}".format(
io.get("output_train")))
def gen_split(preprocessed: List[Preprocessed], io: IO) -> List[Preprocessed]:
split = io.get("split")
if split == 100:
output_split(preprocessed, preprocessed, io)
return preprocessed
shuffled = random.sample(preprocessed, len(preprocessed))
total_len = sum(map(lambda x: len(x.features), preprocessed))
train_len = (total_len / 100.0) * split
train_items = list()
current_len = 0
for i in range(len(preprocessed) - 1):
new_len = current_len + len(shuffled[i].features)
if new_len >= train_len:
output_split(preprocessed, train_items, io)
return train_items
current_len = new_len
train_items.append(shuffled[i])
output_split(preprocessed, train_items, io)
return train_items
def collect_input_paths(io: IO) -> List[str]:
"""Turn the input glob into file paths"""
all_files = list(set(io.get("input_files")))
wav_files = list(
filter(lambda in_file: in_file.split(".")[-1] == "wav", all_files))
return wav_files
def predictions_to_out_file(predictions: np.array, io: IO):
obj = {"items": [], "genre": {"hard": 0.5, "uptempo": 0.5}}
interval = io.get("interval")
melodies = list()
cur_time = 0
for i in range(len(predictions)):
prediction = predictions[i]
beat, melody = prediction
if is_positive_beat(beat):
cur_obj = {}
cur_obj["type"] = "beat"
cur_obj["time"] = cur_time
obj["items"].append(cur_obj)
if is_positive_melody(melody):
cur_obj = {}
cur_obj["type"] = "melody"
cur_obj["time"] = cur_time
cur_obj["duration"] = interval
melodies.append(cur_obj)
cur_time += interval
obj["items"] = obj["items"] + stitch_melodies(melodies, io)
return obj
def gen_outputs(file: MarkedAudioFile, io: IO) -> List[ExpectedOutput]:
"""Gen a list of marked outputs for given file"""
out_len = len(file.bins_file.bins)
outputs = [ExpectedOutput(False, False) for x in range(out_len)]
interval = io.get("interval")
for timestamp in file.json_file.timestamps:
# Round it to the range
timestamp_time = timestamp.timestamp * 1000
closest = get_closest(timestamp_time, io)
timestamp_index = int(closest / interval)
if timestamp_index >= out_len:
continue
if timestamp.beat_type == "beat":
output_mark = outputs[timestamp_index]
output_mark.is_beat = True
elif timestamp.beat_type == "melody":
closest_end = get_closest(timestamp_time +
(timestamp.length * 1000))
for i in range(int((closest_end - closest) / interval)):
outputs[timestamp_index + i].is_melody = True
return outputs
def run_tests(io: IO, model: Sequential, test_files: List[Preprocessed]):
model.reset_states()
for file in test_files:
logline("creating test params for {}".format(file.file_name))
test_x, test_y = get_test_params(file)
logline("making predictions")
predictions: List[List[float]] = model.predict(test_x,
batch_size=1,
verbose=1)
model.reset_states()
mse_total: List[float] = list()
correct = 0
diff_score = 0
for i in range(len(predictions)):
prediction = predictions[i]
actual: List[float] = test_y[i]
diff = abs(actual[0] - prediction[0])
diff_score += diff
if is_in_range(diff):
correct += 1
mse_total.append(mean_squared_error(actual, prediction))
logline(
"predicted {}/{} within range ({}%) correct, score was {}/{}, mse was {}"
.format(
correct,
len(predictions),
round(correct / len(predictions) * 100, 2),
diff_score,
len(predictions),
round(sum(mse_total) / len(predictions), 4),
))
out_obj = predictions_to_out_file(predictions, io)
pathlib.Path(io.get("output_annotated")).mkdir(parents=True,
exist_ok=True)
out_path = os.path.join(io.get("output_annotated"),
"{}.json".format(file.file_name))
with open(out_path, "w+") as out_file:
json.dump(out_obj, out_file)
logline("wrote object to {}".format(out_path))
def start_server(io: IO):
global interval
interval = io.get("interval")
port = io.get("port")
httpd = HTTPServer(("", port),
partial(WebServer,
directory=os.path.join(CUR_DIR, "public")))
logline("listening at port", port)
enter_group()
try:
httpd.serve_forever()
except KeyboardInterrupt:
pass
httpd.server_close()
exit_group()
logline("stopped listening")
def get_closest(timestamp_time: float, io: IO) -> int:
"""Get the closest multiple of INTERVAL to the timestamp"""
interval = io.get("interval")
lowerbound = (timestamp_time // interval) * interval
upperbound = lowerbound + interval
lowerbound_diff = timestamp_time - lowerbound
upperbound_diff = upperbound - timestamp_time
return lowerbound if lowerbound_diff <= upperbound_diff else upperbound
def trim_params(params: Tuple[np.ndarray, np.ndarray], io: IO) -> Tuple[np.ndarray, np.ndarray]:
batch_size = io.get("batch_size")
x_param, y_param = params
length = x_param.shape[0]
remainder = length % batch_size
if remainder == 0:
return params
return x_param[:-remainder], y_param[:-remainder]
def collect_input_paths(io: IO) -> Union[None, List[str]]:
"""Turn the input glob into file paths"""
all_files = io.get("input_files")
wav_files = list(
filter(lambda in_file: in_file.split(".")[-1] == "wav", all_files))
if len(wav_files) == 0:
return None
annotated_files = list(filter(has_json_file, wav_files))
return annotated_files
def stitch_melodies(obj: List[Dict[str, Union[str, float]]],
io: IO) -> List[Dict[str, Union[str, float]]]:
new_melodies = list()
interval = io.get("interval")
i = 0
while i < len(obj):
if len(new_melodies) > 0:
if new_melodies[-1]["time"] == obj[i]["time"] - interval:
new_melodies[-1]["time"] += interval
i += 1
continue
new_melodies.append(obj[i])
i += 1
return new_melodies
def predictions_to_out_file(predictions: np.array, io: IO):
obj: Dict[str, Any] = {"items": [], "genre": {"hard": 0.5, "uptempo": 0.5}}
interval = io.get("interval")
cur_time = 0
for i in range(len(predictions)):
prediction: Tuple[float] = predictions[i]
confidence: float = prediction[0]
if is_in_range(confidence):
cur_obj = {}
cur_obj["type"] = "beat"
cur_obj["time"] = cur_time
obj["items"].append(cur_obj)
cur_time += interval
return obj
def match_files(io: IO, input_paths: List[str]):
"""Match found files to analysis file contents"""
analysis_file = io.get("analysis")
logline(analysis_file)
analysis = AnalysisFile(analysis_file)
mapped: Dict[str, str] = {}
reverse_map: Dict[str, str] = {}
for in_path in input_paths:
file_name = in_path.split("/")[-1].split(".")[0]
for track_analysis in analysis.tracks:
if track_analysis.name.lower() in file_name.lower():
mapped[in_path] = track_analysis.name
reverse_map[track_analysis.name] = file_name
break
logline("came up with the following mapping:")
logline("")
for file_name in mapped:
logline('"{}" -> "{}"'.format(file_name, mapped[file_name]))
unmapped_amount: int = 0
for in_path in input_paths:
if in_path not in mapped:
warn('input file "{}" not mapped'.format(in_path))
unmapped_amount += 1
for track_analysis in analysis.tracks:
if track_analysis.name not in reverse_map:
warn('analysed file "{}" not mapped'.format(track_analysis.name))
unmapped_amount += 1
logline("")
if unmapped_amount > 0:
try:
correct = input("is this correct? Y/n")
if correct.lower() == "n":
return None
except KeyboardInterrupt:
return None
return analysis, mapped
def gen_outputs(file: MarkedAudioFile, io: IO) -> List[ExpectedOutput]:
"""Gen a list of marked outputs for given file"""
out_len = len(file.bins_file.bins)
# TODO: change
outputs = [ExpectedOutput(0) for _ in range(out_len)]
interval = io.get("interval")
for timestamp in file.timestamps:
# Round it to the range
timestamp_time = timestamp.timestamp * 1000
closest = get_closest(timestamp_time, io)
timestamp_index = int(closest / interval)
if timestamp_index >= out_len:
continue
output_mark = outputs[timestamp_index]
output_mark.beat_confidence = timestamp.confidence
return outputs
def run_tests(io: IO, model: Sequential, test_files: List[Preprocessed]):
model.reset_states()
for file in test_files:
logline("creating test params for {}".format(file.file_name))
test_x, test_y = get_test_params(file)
logline("making predictions")
predictions = model.predict(test_x, batch_size=1, verbose=1)
model.reset_states()
mse_total = list()
correct = 0
for i in range(len(predictions)):
prediction = predictions[i]
actual = test_y[i]
if actual[0] == is_positive_beat(
prediction[0]) and actual[1] == is_positive_melody(
prediction[1]):
correct += 1
mse_total.append(mean_squared_error(actual, prediction))
logline("predicted {}/{} ({}%) correct, mse was {}".format(
correct,
len(predictions),
round(correct / len(predictions) * 100, 2),
round(sum(mse_total) / len(predictions), 4),
))
out_obj = predictions_to_out_file(predictions, io)
out_path = os.path.join(io.get("output_annotated"),
"{}.json".format(file.file_name))
with open(out_path, "w+") as out_file:
json.dump(out_obj, out_file)
logline("wrote object to {}".format(out_path))
def export_model(model: Sequential, io: IO):
logline('wrote weights to file "{}"'.format(io.get("output_weights")))
model.save_weights(io.get("output_weights"))
def apply_weights(model: Sequential, io: IO) -> Sequential:
model.load_weights(io.get("input_weights"))
return model
def load_preprocessed(io: IO) -> List[Preprocessed]:
with open(io.get("input_file"), "rb") as in_file:
return list(map(lambda x: Preprocessed(x), pickle.load(in_file)))
