def prepare_eval_data(audio_id):
file_path = os.path.join(ONBOARDING_DATA_PATH, "dnn2016_%d.pkl" % audio_id)
with open(file_path, 'rb') as file_handler:
data = pickle.load(file_handler)
compact_window = np.array((
data['zxx_log'],
np.roll(data['zxx_log'], -1, axis=1),
np.roll(data['zxx_log'], -2, axis=1),
np.roll(data['zxx_log'], -3, axis=1),
np.roll(data['zxx_log'], -4, axis=1),
np.roll(data['zxx_log'], -5, axis=1),
np.roll(data['zxx_log'], -6, axis=1),
np.roll(data['zxx_log'], -7, axis=1),
np.roll(data['zxx_log'], -8, axis=1),
np.roll(data['zxx_log'], -9, axis=1),
np.roll(data['zxx_log'], -10, axis=1),
np.roll(data['zxx_log'], -11, axis=1),
np.roll(data['zxx_log'], -12, axis=1),
np.roll(data['zxx_log'], -13, axis=1),
np.roll(data['zxx_log'], -14, axis=1),
np.roll(data['zxx_log'], -15, axis=1),
))
compact_window = np.swapaxes(
np.swapaxes(np.swapaxes(compact_window, 0, 1), 0, 2), 1,
2).reshape((-1, 64, 16))
return th.from_numpy(compact_window).double(), rms_to_db(data['rms'])
def label_segmentation(label_list, decibel_lower_bound=0.0):
df_list = []
table_fine = pd.read_csv(os.path.join(BASE_PATH,
'./fine_grained_annotation.csv'),
index_col=0)
for index, row in tqdm(label_list.iterrows(), total=label_list.shape[0]):
record_label = row['label']
audio_id = index
filepath = os.path.join(DATA_PATH, "dnn2016_%d.pkl" % audio_id)
try:
with open(filepath, 'rb') as handler:
data = pickle.load(handler)
audio_length = data['zxx_log'].shape[1]
decibel = rms_to_db(data['rms'])
# COUGH
if record_label:
# Load Fine Grained Data
record_list = table_fine.loc[
table_fine.coarse_grained_annotation_id == audio_id,
['label', 'label_end', 'label_start']]
cough_list = record_list[record_list.label == 'Cough']
# Encode One-hot Array for the cough event
time_seq = np.repeat(False, audio_length)
for index, cough in cough_list.iterrows():
label_start = round(
cough['label_start'] * SR) // WINDOW_SIZE
label_end = round(
cough['label_end'] * SR) // WINDOW_SIZE
time_seq[label_start:label_end - 1] = True
# Sliding Window
padded = np.pad(time_seq, (0, 16), constant_values=(0, 0))
rolling_list = np.array((
padded,
np.roll(padded, -1),
np.roll(padded, -2),
np.roll(padded, -3),
np.roll(padded, -4),
np.roll(padded, -5),
np.roll(padded, -6),
np.roll(padded, -7),
np.roll(padded, -8),
np.roll(padded, -9),
np.roll(padded, -10),
np.roll(padded, -11),
np.roll(padded, -12),
np.roll(padded, -13),
np.roll(padded, -14),
np.roll(padded, -15),
))
labels = np.sum(rolling_list, axis=0) / 16.0
df = pd.DataFrame({
"audio":
audio_id,
"window_index":
np.arange(0, labels[:-16 - 15].shape[0]),
"label":
labels[:-16 - 15],
"db":
decibel[:-15]
})
df = df[df.db >= decibel_lower_bound].drop(columns=['db'])
df_list.append(df)
# NON-COUGH
else:
df = pd.DataFrame({
"audio":
audio_id,
"window_index":
np.arange(0, audio_length - 15),
"label":
0,
"db":
decibel[:-15]
})
df = df[df.db >= decibel_lower_bound].drop(columns=['db'])
df_list.append(df)
pass
except Exception as e:
pass
return pd.concat(df_list, ignore_index=True)
def label_segmentation(label_list, decibel_lower_bound=0.0):
df_list = []
table_fine = pd.read_csv(os.path.join(BASE_PATH,
'./fine_grained_annotation.csv'),
index_col=0)
for index, row in tqdm(label_list.iterrows(), total=label_list.shape[0]):
record_label = row['label']
audio_id = index
filepath = os.path.join(DATA_PATH, "dnn2016_%d.pkl" % audio_id)
try:
with open(filepath, 'rb') as handler:
data = pickle.load(handler)
audio_length = data['zxx_log'].shape[1]
decibel = rms_to_db(data['rms'])
# COUGH
if record_label:
# Load Fine Grained Data
record_list = table_fine.loc[
table_fine.coarse_grained_annotation_id == audio_id,
['label', 'label_end', 'label_start']]
cough_list = record_list[record_list.label == 'Cough']
# Encode One-hot Array for the cough event
#time_seq = np.repeat(False, audio_length)
for index, cough in cough_list.iterrows():
label_start = round(
cough['label_start'] * SR) // WINDOW_SIZE
label_end = round(
cough['label_end'] * SR) // WINDOW_SIZE
#time_seq[label_start: label_end - 1] = True
label_end = audio_length if label_end > audio_length else label_end
df = pd.DataFrame({
"audio":
audio_id,
"window_index":
np.arange(label_start, label_end - 15,
16), # no Sliding
"label":
1,
"db":
decibel[label_start:label_end - 15:16]
})
df = df[df.db >= decibel_lower_bound].drop(
columns=['db'])
df_list.append(df)
# NON-COUGH
else:
df = pd.DataFrame({
"audio":
audio_id,
"window_index":
np.arange(0, audio_length - 15,
16 * DROP_NON_COUGH_RATIO), # no Sliding
"label":
0,
"db":
decibel[:-15:16 * DROP_NON_COUGH_RATIO]
})
df = df[df.db >= decibel_lower_bound].drop(columns=['db'])
df_list.append(df)
pass
except Exception as e:
pass
return pd.concat(df_list, ignore_index=True)