DACON 음향 데이터 COVID-19 검출 AI 경진대회
참고 | [Private 6위, 0.60553] DNN 코드 공유
사용 라이브러리
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
| import numpy as np
import pandas as pd
import os
import librosa
import matplotlib.pyplot as plt
import seaborn as sns
import koreanize_matplotlib
from tqdm.auto import tqdm
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout
from sklearn.model_selection import train_test_split
|
Data Load
1
2
3
4
5
| base_path = "./open/"
train = pd.read_csv(base_path+"train_data.csv")
test = pd.read_csv(base_path+"test_data.csv")
train.shape, test.shape
|
Pre-Processing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
| features = []
log_specgrams_hp = []
feature_delta = [] # feature 3
melspectrogram_list = [] # feature 1
harmonic_percussive_list = [] # feature 2
for uid in tqdm(train["id"]):
root_path = os.path.join(base_path, "train")
path = os.path.join(root_path, str(uid).zfill(5)+".wav")
# wav 파일 load
y, sr = librosa.load(path, sr=16000)
# melspectrogram 추출
melspectrogram = librosa.feature.melspectrogram(y=y, sr=16000, n_mels=32)
# 로그스케일링 된 melspectrogram의 델타값
delta = []
for e in melspectrogram:
delta.append(np.mean(librosa.feature.delta(e)))
feature_delta.append(delta)
# 로그 스케일링
feature_log_scale = librosa.power_to_db(S=melspectrogram, ref=1.0)
# feature 1: 추출된 melspectrogram드르이 평균
temp = []
for e in feature_log_scale:
temp.append(np.mean(e))
melspectrogram_list.append(temp)
# feature 2: (운율적 소리(harmonic)+두드리는 소리(percussive)의 구성 요소) 평균
y_harmonic, y_percussive = librosa.effects.hpss(y)
melspectrogram_harmonic = librosa.feature.melspectrogram(y_harmonic, n_mels=32)
melspectrogram_percussive = librosa.feature.melspectrogram(y_percussive, n_mels=32)
log_harmonic = librosa.amplitude_to_db(melspectrogram_harmonic)
log_percussive = librosa.amplitude_to_db(melspectrogram_percussive)
log_hp = np.average([log_harmonic, log_percussive], axis=0)
temp = []
for e in log_hp:
temp.append(np.mean(e))
harmonic_percussive_list.append(temp)
|
1
2
3
4
| # 오디오 Feature 추가
f_list1 = pd.DataFrame(melspectrogram_list)
f_list2 = pd.DataFrame(harmonic_percussive_list)
f_list3 = pd.DataFrame(feature_delta)
|
1
2
3
| f_list1.columns = ["melspectrogram_1_"+str(x) for x in range(1, 33)]
f_list2.columns = ["melspectrogram_2_"+str(x) for x in range(1, 33)]
f_list3.columns = ["melspectrogram_3_"+str(x) for x in range(1, 33)]
|
1
2
| melspectrogram_train_df = pd.concat([f_list1,f_list2,f_list3], axis=1)
train_df = pd.concat([train, melspectrogram_train_df], axis=1)
|
| id | age | gender | respiratory_condition | fever_or_muscle_pain | covid19 | melspectrogram_1_1 | melspectrogram_1_2 | melspectrogram_1_3 | melspectrogram_1_4 | ... | melspectrogram_3_23 | melspectrogram_3_24 | melspectrogram_3_25 | melspectrogram_3_26 | melspectrogram_3_27 | melspectrogram_3_28 | melspectrogram_3_29 | melspectrogram_3_30 | melspectrogram_3_31 | melspectrogram_3_32 |
|---|
| 0 | 1 | 24 | female | 0 | 1 | 0 | -31.943413 | -27.335651 | -23.652283 | -20.332117 | ... | 1.551721e-05 | 5.001302e-05 | 7.300529e-05 | 6.368942e-05 | 8.467526e-05 | 2.460767e-05 | 1.935706e-05 | 3.276046e-05 | 1.080332e-05 | 9.988543e-07 |
|---|
| 1 | 2 | 51 | male | 0 | 0 | 0 | -18.772213 | -17.970600 | -19.152706 | -20.261648 | ... | 4.808880e-07 | 3.241044e-07 | 2.775657e-07 | 2.403397e-07 | 1.801275e-07 | 1.468669e-07 | 1.040892e-07 | 3.793937e-08 | 2.792552e-08 | -5.082057e-09 |
|---|
| 2 | 3 | 22 | male | 0 | 0 | 0 | -41.755840 | -37.208641 | -36.034973 | -36.101788 | ... | -2.077132e-08 | -1.249289e-08 | -1.005954e-08 | -6.208817e-09 | -1.373466e-09 | -1.762301e-09 | -2.323603e-09 | -1.048914e-09 | -2.308708e-10 | -4.273706e-11 |
|---|
| 3 | 4 | 29 | female | 1 | 0 | 0 | -36.795872 | -28.668598 | -27.015005 | -27.315031 | ... | -8.004472e-08 | -6.886885e-08 | -4.426233e-08 | -2.452074e-08 | -3.148728e-08 | -4.078213e-08 | -1.627670e-08 | -1.823432e-08 | -7.824335e-09 | -9.221319e-10 |
|---|
| 4 | 5 | 23 | male | 0 | 0 | 0 | -50.650784 | -46.894409 | -47.190170 | -48.408485 | ... | -5.674101e-10 | -8.946563e-10 | 5.834516e-10 | -1.668314e-10 | 1.389479e-09 | -1.114242e-09 | 2.274453e-11 | 3.102881e-10 | 2.110314e-10 | 8.379827e-11 |
|---|
5 rows × 102 columns
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
| # test
features = []
log_specgrams_hp = []
feature_delta = [] # feature 3
melspectrogram_list = [] # feature 1
harmonic_percussive_list = [] # feature 2
for uid in tqdm(test["id"]):
root_path = os.path.join(base_path, "test")
path = os.path.join(root_path, str(uid).zfill(5)+".wav")
# wav 파일 load
y, sr = librosa.load(path, sr=16000)
# melspectrogram 추출
melspectrogram = librosa.feature.melspectrogram(y=y, sr=16000, n_mels=32)
# 로그스케일링 된 melspectrogram의 델타값
delta = []
for e in melspectrogram:
delta.append(np.mean(librosa.feature.delta(e)))
feature_delta.append(delta)
# 로그 스케일링
feature_log_scale = librosa.power_to_db(S=melspectrogram, ref=1.0)
# feature 1: 추출된 melspectrogram드르이 평균
temp = []
for e in feature_log_scale:
temp.append(np.mean(e))
melspectrogram_list.append(temp)
# feature 2: (운율적 소리(harmonic)+두드리는 소리(percussive)의 구성 요소) 평균
y_harmonic, y_percussive = librosa.effects.hpss(y)
melspectrogram_harmonic = librosa.feature.melspectrogram(y_harmonic, n_mels=32)
melspectrogram_percussive = librosa.feature.melspectrogram(y_percussive, n_mels=32)
log_harmonic = librosa.amplitude_to_db(melspectrogram_harmonic)
log_percussive = librosa.amplitude_to_db(melspectrogram_percussive)
log_hp = np.average([log_harmonic, log_percussive], axis=0)
temp = []
for e in log_hp:
temp.append(np.mean(e))
harmonic_percussive_list.append(temp)
|
1
2
3
4
| # 오디오 Feature 추가
f_list1 = pd.DataFrame(melspectrogram_list)
f_list2 = pd.DataFrame(harmonic_percussive_list)
f_list3 = pd.DataFrame(feature_delta)
|
1
2
3
| f_list1.columns = ["melspectrogram_1_"+str(x) for x in range(1, 33)]
f_list2.columns = ["melspectrogram_2_"+str(x) for x in range(1, 33)]
f_list3.columns = ["melspectrogram_3_"+str(x) for x in range(1, 33)]
|
1
2
| melspectrogram_test_df = pd.concat([f_list1,f_list2,f_list3], axis=1)
test_df = pd.concat([test, melspectrogram_test_df], axis=1)
|
| id | age | gender | respiratory_condition | fever_or_muscle_pain | melspectrogram_1_1 | melspectrogram_1_2 | melspectrogram_1_3 | melspectrogram_1_4 | melspectrogram_1_5 | ... | melspectrogram_3_23 | melspectrogram_3_24 | melspectrogram_3_25 | melspectrogram_3_26 | melspectrogram_3_27 | melspectrogram_3_28 | melspectrogram_3_29 | melspectrogram_3_30 | melspectrogram_3_31 | melspectrogram_3_32 |
|---|
| 0 | 3806 | 48 | female | 1 | 0 | -55.853134 | -55.158051 | -54.891762 | -54.874367 | -54.994320 | ... | 0.000000e+00 | 3.043538e-12 | 7.608845e-13 | 0.000000e+00 | 0.000000e+00 | -3.043538e-12 | -7.608845e-13 | -7.608845e-13 | 3.091093e-13 | -5.944410e-15 |
|---|
| 1 | 3807 | 24 | female | 0 | 0 | -46.948128 | -46.494526 | -46.730667 | -46.501385 | -46.066654 | ... | 4.068705e-10 | -3.715323e-10 | -7.672049e-10 | -1.703404e-10 | 1.796148e-09 | 2.624169e-10 | 1.179117e-10 | -4.361051e-11 | 1.998812e-11 | -1.802107e-11 |
|---|
| 2 | 3808 | 29 | male | 0 | 0 | -99.176407 | -93.005913 | -91.820778 | -92.210541 | -93.268394 | ... | -4.520774e-13 | -3.750947e-13 | -7.825289e-13 | -1.284776e-12 | -6.738704e-13 | -9.572461e-13 | -1.115856e-12 | -6.283318e-13 | -1.451943e-12 | -3.377899e-13 |
|---|
| 3 | 3809 | 39 | female | 0 | 0 | -33.959797 | -27.807213 | -26.108114 | -28.486483 | -29.118986 | ... | -9.174222e-10 | 2.168453e-09 | -4.170101e-10 | 5.560135e-11 | -6.950168e-12 | -5.125749e-11 | -4.969370e-10 | 2.293556e-10 | 1.390034e-10 | 4.734802e-11 |
|---|
| 4 | 3810 | 34 | male | 0 | 0 | -33.321808 | -29.476269 | -30.429527 | -30.836193 | -31.185854 | ... | -5.356627e-10 | 7.791457e-10 | 3.895728e-10 | 0.000000e+00 | 1.217415e-10 | 8.521906e-11 | 6.087076e-12 | 0.000000e+00 | -2.434830e-11 | -1.217415e-11 |
|---|
5 rows × 101 columns
Train
1
2
| x_train = train_df.drop(columns=["id", "covid19"])
y_train = train_df["covid19"]
|
1
| x_train = pd.get_dummies(x_train, columns=["gender"], drop_first=False)
|
1
2
3
4
5
6
7
8
9
10
11
| model = Sequential(name="Covid19_DNN")
model.add(Dense(512, activation="relu", input_shape=(x_train.shape[0], x_train.shape[1])))
model.add(Dense(256, activation="relu"))
model.add(Dropout(0.2))
model.add(Dense(128, activation="relu"))
model.add(Dense(64, activation="relu"))
model.add(Dropout(0.2))
model.add(Dense(10, activation="relu"))
model.add(Dense(2, activation="sigmoid"))
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
| Model: "Covid19_DNN"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_13 (Dense) (None, 3805, 512) 52736
_________________________________________________________________
dense_14 (Dense) (None, 3805, 256) 131328
_________________________________________________________________
dropout_4 (Dropout) (None, 3805, 256) 0
_________________________________________________________________
dense_15 (Dense) (None, 3805, 128) 32896
_________________________________________________________________
dense_16 (Dense) (None, 3805, 64) 8256
_________________________________________________________________
dropout_5 (Dropout) (None, 3805, 64) 0
_________________________________________________________________
dense_17 (Dense) (None, 3805, 10) 650
_________________________________________________________________
dense_18 (Dense) (None, 3805, 2) 22
=================================================================
Total params: 225,888
Trainable params: 225,888
Non-trainable params: 0
_________________________________________________________________
|
1
2
3
4
5
| model.compile(
optimizer = "adam",
loss = tf.keras.losses.SparseCategoricalCrossentropy(),
metrics = ["accuracy"]
)
|
1
2
3
| from tensorflow.keras.callbacks import EarlyStopping
early_stop = EarlyStopping(monitor="val_loss", patience=20, restore_best_weights=True)
|
1
2
| with tf.device("/device:GPU:0"):
histroy = model.fit(x_train, y_train, validation_split=0.2, epochs=100, batch_size=64, callbacks=[early_stop])
|
1
2
3
4
5
6
7
8
9
10
11
| Epoch 1/100
48/48 [==============================] - 2s 9ms/step - loss: 0.7508 - accuracy: 0.8811 - val_loss: 0.3203 - val_accuracy: 0.9080
Epoch 2/100
48/48 [==============================] - 0s 6ms/step - loss: 0.3339 - accuracy: 0.9149 - val_loss: 0.3312 - val_accuracy: 0.9080
Epoch 3/100
48/48 [==============================] - 0s 6ms/step - loss: 0.3159 - accuracy: 0.9215 - val_loss: 0.3192 - val_accuracy: 0.9080
Epoch 4/100
48/48 [==============================] - 0s 6ms/step - loss: 0.
...
Epoch 66/100
48/48 [==============================] - 0s 5ms/step - loss: 0.2662 - accuracy: 0.9231 - val_loss: 0.2988 - val_accuracy: 0.9080
|
Evalutate
1
| test_loss, test_acc = model.evaluate(x_train, y_train)
|
1
| 119/119 [==============================] - 0s 3ms/step - loss: 0.2735 - accuracy: 0.9198
|
1
2
3
4
5
6
| hist_df = pd.DataFrame(histroy.history)
fig, ax = plt.subplots(1, 2, figsize=(14, 5))
hist_df[["accuracy", "val_accuracy"]].plot(ax=ax[0])
hist_df[["loss", "val_loss"]].plot(ax=ax[1])
plt.show()
|
Submission
1
2
| x_test = test_df.drop(columns=["id"])
x_test = pd.get_dummies(x_test, columns=["gender"], drop_first=False)
|
1
| pred = model.predict(x_test)
|
1
2
| y_pred = pd.DataFrame(np.where((pred>0.84), 0, 1))
y_pred = pd.DataFrame(y_pred.iloc[:, 0])
|
1
2
3
| sub = pd.read_csv("./open/sample_submission.csv")
sub["covid19"] = y_pred
sub.to_csv("sub_dnn.csv", index=False)
|
마무리
음향 데이터를 사용해 볼 목적으로 비슷한 경진대회를 필사해봤다.
처음 다뤄보는 데이터라 막막했지만, 막상 진행해보니 음향 데이터에서 피처를 뽑는 부분만 새롭고 이후 부분들은 일반적인 방식과 동일했다. 음향 데이터를 위해 사용한 librosa 라이브러리와 DSP 같은 이론을 좀 더 알면 음향 데이터도 쉽게 사용할 수 있을꺼 같다. DACON 제출시 정확도 50%로 170등대에 랭크되었다. 참고한 노트북에서는 softmax를 출력층 활성 함수로 사용했는데, 왜 softmax를 사용한지 모르겠다. 확진 여부를 분리하는 모델이므로 나는 sigmoid를 사용했는데 성능 차이가 이 부분에서 발생한거 같다.
softmax로 활성 함수를 변경하면 58% 정확도로 59등까지 등수가 올라간다. (22-07-29일 기준)
validation에서는 크게 성능 차이가 안나는데 제출시에는 성능 차이가 많이 발생한다.
++ 분류 문제가 아니라 회귀 문제라 끝단에 softmax를 사용한거 같다.
Comments powered by Disqus.