我正在测试哪种架构最适合开始对模型进行微调。我有三个不同的测试用例：一个CuDNNLSTM层，一个CuDNNGRU层以及一个将输入平坦化的密集层。全部都有64个节点。
我的输入数据是汽车速度的50个长度序列，我的输出是预测油门和制动百分比。所有数据均已标准化为0到1。
但是，出乎意料的是，如果我整理数据以便可以使用密集层，那么似乎可以在5个时期内将其表现得更好。这是a幸，还是在某些情况下致密层胜过专用的循环层？
密度：
Epoch 1/5
462280/462280 [==============================] - 4s 8us/step - loss: 0.0090 - mae: 0.0716 - val_loss: 0.0072 - val_mae: 0.0586
Epoch 2/5
462280/462280 [==============================] - 3s 7us/step - loss: 0.0062 - mae: 0.0610 - val_loss: 0.0053 - val_mae: 0.0581
Epoch 3/5
462280/462280 [==============================] - 3s 6us/step - loss: 0.0056 - mae: 0.0575 - val_loss: 0.0056 - val_mae: 0.0624
Epoch 4/5
462280/462280 [==============================] - 3s 8us/step - loss: 0.0052 - mae: 0.0553 - val_loss: 0.0047 - val_mae: 0.0547
Epoch 5/5
462280/462280 [==============================] - 3s 6us/step - loss: 0.0050 - mae: 0.0543 - val_loss: 0.0049 - val_mae: 0.0578
LSTM：
Epoch 1/5:
462280/462280 [==============================] - 10s 22us/step - loss: 0.0105 - mae: 0.0793 - val_loss: 0.0083 - val_mae: 0.0709
Epoch 2/5
462280/462280 [==============================] - 9s 19us/step - loss: 0.0083 - mae: 0.0703 - val_loss: 0.0083 - val_mae: 0.0717
Epoch 3/5
462280/462280 [==============================] - 8s 18us/step - loss: 0.0083 - mae: 0.0697 - val_loss: 0.0082 - val_mae: 0.0704
Epoch 4/5
462280/462280 [==============================] - 9s 18us/step - loss: 0.0082 - mae: 0.0697 - val_loss: 0.0081 - val_mae: 0.0666
Epoch 5/5
462280/462280 [==============================] - 8s 18us/step - loss: 0.0076 - mae: 0.0672 - val_loss: 0.0065 - val_mae: 0.0647
GRU：
Epoch 1/5
462280/462280 [==============================] - 11s 25us/step - loss: 0.0095 - mae: 0.0746 - val_loss: 0.0084 - val_mae: 0.0731
Epoch 2/5
462280/462280 [==============================] - 10s 22us/step - loss: 0.0084 - mae: 0.0701 - val_loss: 0.0082 - val_mae: 0.0692
Epoch 3/5
462280/462280 [==============================] - 10s 21us/step - loss: 0.0083 - mae: 0.0696 - val_loss: 0.0081 - val_mae: 0.0692
Epoch 4/5
462280/462280 [==============================] - 10s 21us/step - loss: 0.0082 - mae: 0.0692 - val_loss: 0.0081 - val_mae: 0.0696
Epoch 5/5
462280/462280 [==============================] - 10s 21us/step - loss: 0.0083 - mae: 0.0692 - val_loss: 0.0083 - val_mae: 0.0717