我使用这种架构(一个用于不同轨迹长度的掩蔽层,用0填充到最大长度的轨迹,然后是LSTM,后面有一个密集层,输出2个值(来构建一个基于轨迹预测2个值的回归器。
samples, timesteps, features = x_train.shape[0], x_train.shape[1], x_train.shape[2]
model = Sequential()
model.add(tf.keras.layers.Masking(mask_value=0., input_shape=(timesteps, features), name="mask"))
model.add(LSTM(30, return_sequences=True, name="lstm1"))
model.add(LSTM(30, return_sequences=False, name="lstm2"))
model.add(Dense(20, activation='relu', name="dense1"))
model.add(Dense(20, activation='relu', name="dense2"))
model.add(Dense(2, activation='linear', name="output"))
model.compile(optimizer="adam", loss="mse")
培训:
model.fit(x_train, y_train, epochs = 10, batch_size = 32)
我的输入数据是形状:
x_train (269, 527, 11) (269 trajectories of 527 timesteps of 11 features)
y_train (269, 2) (these 269 trajectories have 2 target values)
x_test (30, 527, 11) (--- same ---)
y_test (30, 2) (--- same ---)
我已经对我的数据进行了预处理,这样我所有的序列都有固定的长度,较小的序列在缺失的时间步长处用0填充。因此,我使用遮罩层来跳过这些时间步长,因为它们没有提供任何信息。
正如预期的那样,输出是稳定的:
(30, 2)
但仔细观察它似乎在回归同样的价值观。
[[37.48257 0.7025466 ]
[37.48258 0.70254654]
[37.48257 0.70254654]
[37.48257 0.7025466 ]
[37.48258 0.70254654]
[37.48258 0.70254654]
[37.48258 0.70254654]
[37.48258 0.7025465 ]
[42.243515 0.6581909 ]
[37.48258 0.70254654]
[37.48257 0.70254654]
[37.48258 0.70254654]
[37.48261 0.7025462 ]
[37.48257 0.7025466 ]
[37.482582 0.70254654]
[37.482567 0.70254654]
[37.48257 0.7025466 ]
[37.48258 0.70254654]
[37.48258 0.70254654]
[37.48257 0.7025466 ]
[37.48258 0.70254654]
[37.48258 0.70254654]
[37.48258 0.70254654]
[37.482567 0.7025465 ]
[37.48261 0.7025462 ]
[37.482574 0.7025466 ]
[37.48261 0.7025462 ]
[37.48261 0.70254624]
[37.48258 0.70254654]
[37.48261 0.7025462 ]]
而我的目标值(y_test(是:
[[70. 0.6]
[40. 0.6]
[ 6. 0.6]
[94. 0.7]
[50. 0.6]
[60. 0.6]
[16. 0.6]
[76. 0.9]
[92. 0.6]
[32. 0.8]
[22. 0.7]
[70. 0.7]
[36. 1. ]
[64. 0.7]
[ 0. 0.9]
[82. 0.9]
[38. 0.6]
[54. 0.8]
[28. 0.8]
[62. 0.7]
[12. 0.6]
[72. 0.8]
[66. 0.8]
[ 2. 1. ]
[98. 1. ]
[20. 0.8]
[82. 1. ]
[38. 1. ]
[68. 0.6]
[62. 1. ]]
这就像将整个数据集作为一个数据点。有经验的人能在这里发现明显的错误吗?
感谢您的帮助!
当权重是随机的时,它们对具体的输入计算的贡献是混乱的,我们总是得到几乎相同的输出。你训练模特了吗?看起来不是,在训练之前考虑简单的MNIST求解器输出:
[-2.39 -2.54 -2.23 -2.24 -2.29 -2.37 -2.39 -2.10 -2.34 -2.20]
[-2.28 -2.43 -2.25 -2.33 -2.28 -2.42 -2.26 -2.19 -2.37 -2.25]
[-2.43 -2.44 -2.25 -2.33 -2.33 -2.37 -2.30 -2.10 -2.37 -2.17]
[-2.33 -2.43 -2.28 -2.27 -2.34 -2.34 -2.28 -2.16 -2.37 -2.26]
以及之后:
[-31.72 -31.65 -25.43 -20.04 -29.68 -0.00 -22.74 -25.88 -16.28 -13.30] (5)
[-12.44 -29.92 -21.19 -25.86 -22.53 -12.01 -0.00 -22.61 -18.88 -23.54] (6)
[-23.86 -25.77 -11.88 -9.18 -19.51 -20.85 -28.71 -0.00 -22.11 -14.57] (7)
[-33.67 -23.45 -17.82 -0.00 -28.89 -14.20 -32.54 -14.45 -11.13 -15.40] (3)
UPD:所以提供了训练,但没有实现其目标。很多事情都可能是原因。除了技术问题之外,对于神经网络来说,任务可能会很复杂,例如,如果目标函数不能随着逐步改进而学习。
检查数据路径,尝试简化任务,找到一些解决封闭问题的示例解决方案,检查并返工。