def load_audios(file_list, opts):
sr = opts.get("sr", 16000)
preloaded_file = Path(
opts.get("src_dir")) / "preloaded_audios_{}.pkl".format(sr)
if preloaded_file.exists():
print("Loading preloaded file {}".format(preloaded_file))
return pickle.load(open(preloaded_file, "rb"))
res = [librosa.load(str(file_path), sr=sr)[0] for file_path in file_list]
if opts.get("load_preloaded", True) and not preloaded_file.exists():
with open(ensure_path_exists(preloaded_file, is_file=True), "wb") as f:
pickle.dump(res, f, -1)
print("Saved file: ", preloaded_file)
return res
file_list = src_raw_root.glob("*.WAV")
opts = {"n_mels": 64, "type": "mel", "n_fft": 2048, "hop_length": 512}
#%%
final_path = dest_root / (plot + ".pkl")
if not final_path.exists():
print(final_path)
tmp = {"specs": [], "infos": []}
for file_path in file_list:
print("Loading file: ", file_path)
spec, info = AudioDataHandler.load_raw_data(file_path, opts)
tmp["specs"].append(spec)
tmp["infos"].append(info)
with open(ensure_path_exists(final_path, is_file=True), "wb") as f:
pickle.dump(tmp, f, -1)
print("Saved file: ", final_path)
#%%
model_opts = ModelOptions({
"model_dir": "/mnt/win/UMoncton/Doctorat/dev/dlbd/results/models",
"name": "CityNetTF2_Dropout_resized_mel64_hop_512",
"class": CityNetTF2Dropout,
"version": 19,
"id": "{model--from_epoch}",
"id_prefixes": {
"model--from_epoch": "_fe-"
},
"model": {
#%%
for plot_id in file_list.plot_id.unique():
final_path = dest_root / (plot_id + ".pkl")
if not final_path.exists():
print(final_path)
tmp = {"specs": [], "infos": []}
tmp_df = file_list.loc[file_list.plot_id == plot_id]
for row in tmp_df.itertuples():
path = src_raw_root / row.path
print("Loading file: ", path)
spec, info = AudioDataHandler.load_raw_data(path, opts)
tmp["specs"].append(spec)
tmp["infos"].append(info)
with open(ensure_path_exists(final_path, is_file=True), "wb") as f:
pickle.dump(tmp, f, -1)
print("Saved file: ", final_path)
#%%
model_opts = ModelOptions({
"model_dir": "results/models",
"name": "CityNetTF2_Dropout_resized",
"class": CityNetTF2Dropout,
"version": 1,
"id": "{model--from_epoch}",
"id_prefixes": {
"model--from_epoch": "_fe-"
},
"model": {
def generate_mix(audios, file_df, opts):
# Shuffle all files
tags = np.random.permutation(file_df.id)
# Index of used file
idx = 0
file_count = 1
sr = opts.get("sr", 16000)
duration = opts.get("duration", 30)
nframes = sr * duration
print(len(audios))
print(len(tags))
print(max(tags))
print(max(file_df.id))
# While we still have unused files, create new mixes
for i in range(opts.get("n_iter", 1)):
print("Performing iteration no: ", i + 1)
while idx < tags.shape[0]:
tic = time.time()
tag_infos = []
# Create empty audio file
generated = np.zeros(sr * duration)
# Create empty labels
labels = np.zeros(sr * duration)
noisy = False
start_pos = np.random.choice(range(0, generated.shape[0]), 1000)
max_filled = random.uniform(0, opts.get("max_filled", 0.3))
for start in start_pos:
if idx > tags.shape[0]:
break
audio = audios[tags[idx]]
if not noisy:
snr = random.uniform(20, 50)
noise = get_white_noise(audio, snr, len(generated))
generated += noise
noisy = True
end = min(start + len(audio), len(generated) - 1)
if not opts.get("allow_overlap", False) and (
labels[start] == 1 or labels[end] == 1):
continue
dur = min(end - start, len(audio))
generated[start:end] += audio[0:dur]
labels[start:end] = 1
tag_infos.append({
"tag": file_df.tag.iloc[tags[idx]],
"start": start / sr,
"end": end / sr,
})
prop_filled = sum(labels) / nframes
# print(prop_filled)
idx += 1
if idx >= tags.shape[0]:
break
if prop_filled > max_filled:
break
file_name = Path(opts.get("dest_dir",
".")) / "mix_{}.wav".format(file_count)
librosa.output.write_wav(
file_name,
generated,
opts.get("sr", 16000),
)
pd.DataFrame(tag_infos).to_csv(
Path(opts.get("dest_dir", ".")) /
"mix_{}_tags.csv".format(file_count))
labels_file = Path(opts.get(
"dest_dir", ".")) / "mix_{}_labels.pkl".format(file_count)
with open(ensure_path_exists(labels_file, is_file=True),
"wb") as f:
pickle.dump(labels, f, -1)
print("Generated file {} in {}s.".format(file_name,
time.time() - tic))
file_count += 1
res = []
i = 0
for plot_id in file_list.plot_id.unique():
tic = time()
plot_tmp = []
pkl_path = src_root / (plot_id + ".pkl")
data = pickle.load(open(pkl_path, "rb"))
for i, spec in enumerate(data["specs"]):
tmp = SongDetectorEvaluator.classify_element(model, spec,
data["infos"][i], sampler)
plot_tmp.append(tmp)
tmp_df = pd.concat(plot_tmp).reset_index()
res.append(tmp_df)
tmp_df.to_feather(
str(
ensure_path_exists(dest_root / "intermediate" /
(plot_id + ".feather"))))
print("Classified {0} in {1}".format(plot_id, time() - tic))
res_df = pd.concat(res).reset_index()
res_df.to_feather(
str(ensure_path_exists(dest_root / "predictions_all.feather",
is_file=True)))
# #%%
# import pandas as pd
# a = pd.read_feather("../results/predictions/tommy_preds.feather")
# print(a)