LSTM算法为所有输入生成相同的结果

Question

所以，我目前正在研究一个与车速和角度有关的机器学习算法问题，我正在努力改进我的一些工作。我最近完成了一个XGBRegressor，在我交叉验证的数据上产生了88 - 95%的准确率。但是，我正在尝试改进它，所以我一直在研究LSTM算法，因为我的数据是时间序列相关的。本质上，每个环节都包括转向角、前一次转向角(x-1)、前一次转向角(x-2)以及当前值与前一值之间的差(x - (x-1))。目标是预测某个值是否“异常”。例如，如果一个角度从.1跳到.5 (比例为0-1)，这是不正常的。我之前的算法在分类角度是否异常方面做得很好。不幸的是，我的算法是为每个输入值预测相同的值。例如，这就是它给我的。

test_X = array([[[ 5.86925570e-01,  5.86426251e-01,  5.85832947e-01,
          3.19300000e+03, -5.93304274e-04, -1.09262314e-03]],

       [[ 5.86426251e-01,  5.85832947e-01,  5.85263908e-01,
          3.19400000e+03, -5.69038950e-04, -1.16234322e-03]],

       [[ 5.85832947e-01,  5.85263908e-01,  5.84801158e-01,
          3.19500000e+03, -4.62749993e-04, -1.03178894e-03]],

       ...,

       [[ 4.58070203e-01,  4.57902738e-01,  4.64613980e-01,
          6.38100000e+03,  6.71124195e-03,  6.54377704e-03]],

       [[ 4.57902738e-01,  4.64613980e-01,  7.31314846e-01,
          6.38200000e+03,  2.66700866e-01,  2.73412108e-01]],

       [[ 4.64613980e-01,  7.31314846e-01,  4.68819741e-01,
          6.38300000e+03, -2.62495104e-01,  4.20576175e-03]]])

test_y = array([0, 0, 0, ..., 0, 1, 0], dtype=int64)

yhat = array([[-0.00068355],
       [-0.00068355],
       [-0.00068355],
       ...,
       [-0.00068355],
       [-0.00068355],
       [-0.00068355]], dtype=float32)

到目前为止，我已经尝试根据我在网上读到的一些东西来更改纪元和批处理大小。此外，我还尝试绘制出一些特征，看看是否由于某种原因，算法根本不喜欢它们，但我什么也找不到。我不是机器学习的新手，但我是深度学习的新手，所以如果这是一个愚蠢的问题或问题，我很抱歉。下面的代码。

data = pd.read_csv('final_angles.csv') 
data.dropna(axis=0, subset=['steering_angle'], inplace=True)

from sklearn.preprocessing import MinMaxScaler
scaler = MinMaxScaler()
data['steering_angle'] = scaler.fit_transform(data[['steering_angle']])

y = data.flag #Set y to the value we want to predict, the 'flag' value. 
X = data.drop(['flag', 'frame_id'], axis=1) 

X = concat([X.shift(2), X.shift(1), X], axis=1)
X.columns = ['angle-2', 'id2', 'angle-1', 'id1', 'steering_angle', 'id'] 
X = X.drop(['id2', 'id1'], axis=1)  

X['diff'] = 0;
X['diff2'] = 0;
for index, row in X.iterrows():
    if(index <= 1):
        pass;
    else:
        X.loc[index, "diff"] = row['steering_angle'] - X['steering_angle'][index-1] 
        X.loc[index, "diff2"] = row['steering_angle'] - X['steering_angle'][index-2] 

X = X.iloc[2:,]; 
y = y.iloc[2:,];

train_X, test_X, train_y, test_y = train_test_split(X.as_matrix(), y.as_matrix(), test_size=0.5, shuffle=False)
# reshape input to be 3D [samples, timesteps, features]
train_X = train_X.reshape((train_X.shape[0], 1, train_X.shape[1]))
test_X = test_X.reshape((test_X.shape[0], 1, test_X.shape[1]))
print(train_X.shape, train_y.shape, test_X.shape, test_y.shape)

model = Sequential()
model.add(LSTM(50, input_shape=(train_X.shape[1], train_X.shape[2])))
model.add(Dense(1))
model.compile(loss='mae', optimizer='adam')
# fit network
history = model.fit(train_X, train_y, epochs=50, batch_size=150, validation_data=(test_X, test_y), verbose=2, shuffle=False)

yhat = model.predict(test_X)

而不是预测的值是

array([[-0.00068355],
       [-0.00068355],
       [-0.00068355],
       ...,
       [-0.00068355],
       [-0.00068355],
       [-0.00068355]], dtype=float32)

我期待着更多类似的东西

array([-0.00065207, -0.00065207, -0.00065207,  1.0082773 ,  0.01269123,
        0.01873571, -0.00065207, -0.00065207,  0.99916965,  0.002684  ,
       -0.00018287, -0.00065207, -0.00065207, -0.00065207, -0.00065207,
        1.0021645 ,  0.00654274,  0.01044858, -0.0002622 , -0.0002622 ],
      dtype=float32)

来自前面提到的XGBRegressor测试。

任何帮助，请让我知道，如果需要更多的代码/信息。

编辑:打印语句的结果

Train on 3190 samples, validate on 3191 samples
Epoch 1/50
 - 5s - loss: 0.4268 - val_loss: 0.2820
Epoch 2/50
 - 0s - loss: 0.2053 - val_loss: 0.1256
Epoch 3/50
 - 0s - loss: 0.1442 - val_loss: 0.1256
Epoch 4/50
 - 0s - loss: 0.1276 - val_loss: 0.1198
Epoch 5/50
 - 0s - loss: 0.1256 - val_loss: 0.1179
Epoch 6/50
 - 0s - loss: 0.1250 - val_loss: 0.1188
Epoch 7/50
 - 0s - loss: 0.1258 - val_loss: 0.1183
Epoch 8/50
 - 1s - loss: 0.1258 - val_loss: 0.1199
Epoch 9/50
 - 0s - loss: 0.1256 - val_loss: 0.1179
Epoch 10/50
 - 0s - loss: 0.1255 - val_loss: 0.1192
Epoch 11/50
 - 0s - loss: 0.1247 - val_loss: 0.1180
Epoch 12/50
 - 0s - loss: 0.1254 - val_loss: 0.1185
Epoch 13/50
 - 0s - loss: 0.1252 - val_loss: 0.1176
Epoch 14/50
 - 0s - loss: 0.1258 - val_loss: 0.1197
Epoch 15/50
 - 0s - loss: 0.1251 - val_loss: 0.1175
Epoch 16/50
 - 0s - loss: 0.1253 - val_loss: 0.1176
Epoch 17/50
 - 0s - loss: 0.1247 - val_loss: 0.1183
Epoch 18/50
 - 0s - loss: 0.1249 - val_loss: 0.1178
Epoch 19/50
 - 0s - loss: 0.1253 - val_loss: 0.1178
Epoch 20/50
 - 0s - loss: 0.1253 - val_loss: 0.1181
Epoch 21/50
 - 0s - loss: 0.1245 - val_loss: 0.1192
Epoch 22/50
 - 0s - loss: 0.1250 - val_loss: 0.1187
Epoch 23/50
 - 0s - loss: 0.1244 - val_loss: 0.1184
Epoch 24/50
 - 0s - loss: 0.1252 - val_loss: 0.1188
Epoch 25/50
 - 0s - loss: 0.1253 - val_loss: 0.1197
Epoch 26/50
 - 0s - loss: 0.1253 - val_loss: 0.1192
Epoch 27/50
 - 0s - loss: 0.1267 - val_loss: 0.1177
Epoch 28/50
 - 0s - loss: 0.1256 - val_loss: 0.1182
Epoch 29/50
 - 0s - loss: 0.1247 - val_loss: 0.1178
Epoch 30/50
 - 0s - loss: 0.1249 - val_loss: 0.1183
Epoch 31/50
 - 0s - loss: 0.1259 - val_loss: 0.1189
Epoch 32/50
 - 0s - loss: 0.1258 - val_loss: 0.1187
Epoch 33/50
 - 0s - loss: 0.1248 - val_loss: 0.1179
Epoch 34/50
 - 0s - loss: 0.1259 - val_loss: 0.1203
Epoch 35/50
 - 0s - loss: 0.1252 - val_loss: 0.1190
Epoch 36/50
 - 0s - loss: 0.1260 - val_loss: 0.1192
Epoch 37/50
 - 0s - loss: 0.1249 - val_loss: 0.1183
Epoch 38/50
 - 0s - loss: 0.1249 - val_loss: 0.1187
Epoch 39/50
 - 0s - loss: 0.1252 - val_loss: 0.1185
Epoch 40/50
 - 0s - loss: 0.1246 - val_loss: 0.1183
Epoch 41/50
 - 0s - loss: 0.1247 - val_loss: 0.1179
Epoch 42/50
 - 0s - loss: 0.1242 - val_loss: 0.1194
Epoch 43/50
 - 0s - loss: 0.1255 - val_loss: 0.1187
Epoch 44/50
 - 0s - loss: 0.1244 - val_loss: 0.1176
Epoch 45/50
 - 0s - loss: 0.1248 - val_loss: 0.1183
Epoch 46/50
 - 0s - loss: 0.1257 - val_loss: 0.1179
Epoch 47/50
 - 0s - loss: 0.1248 - val_loss: 0.1177
Epoch 48/50
 - 0s - loss: 0.1247 - val_loss: 0.1194
Epoch 49/50
 - 0s - loss: 0.1248 - val_loss: 0.1181
Epoch 50/50
 - 0s - loss: 0.1245 - val_loss: 0.1182

Answer 1

一个可能的问题可能是您的timestamps。您将输入重塑为timestamps = 1的形状。如果我们想利用LSTM的特性，timestamps应该大于1，不是吗？

如果每个数据点的转向角为3个连续的时间步长，那么您可以尝试使timestamps =3。