Keras字符级别LSTM文本分类不训练

如何解决Keras字符级别LSTM文本分类不训练

我正在使用LSTM（我的第一个模型）制作用于字符级文本分类的keras模型。该模型应该对抽搐聊天中的正常，垃圾邮件和粗鲁消息进行分类。但是我得到的结果令人失望和困惑。

无论我做什么，LSTM网络学到的东西很少，而且准确性也很差。

这是我的代码

import tensorflow as tf
''' import tensorflowjs.converters '''
from tensorflow import keras
from tensorflow.keras import layers
import json
import numpy as np
import re
import random
import sys

np.set_printoptions(threshold=sys.maxsize)

vocab = " qwertyuiopasdfghjklñzxcvbnmç1234567890?¿¡!,:-+/@#áéíóú\/"

dropout = 0.2

x_train = []
y_train = []
one_hot_encode = []
sentence = []

# amount of examples in each class
maxofeachtype = 1600

countnormal = 0
countspam = 0
countofensivo = 0

# Load dataset from data.json
with open("./data.json",'r',encoding="utf8") as file:
    data = json.load(file)

# suffle it
random.shuffle(data)

# create the vocabulary map
mapping = {}
for x in range(len(vocab)):
    mapping[vocab[x]] = x

# this is some to balance the dataset adjusting it to "maxofeachtype"
for example in data:
    if(example["y"] == [1,0] and countnormal < maxofeachtype):
        countnormal += 1
    elif(example["y"] == [0,1,0] and countspam < maxofeachtype):
        countspam += 1
    elif(example["y"] == [0,1] and countofensivo < maxofeachtype):
        countofensivo += 1
    elif(countnormal == maxofeachtype or countspam == maxofeachtype or countofensivo == maxofeachtype):
        continue

    # remove unwanted characters to only have the ones in vocab
    cleanexample = re.sub(
        r'[^qwertyuiopasdfghjklñzxcvbnmç1234567890?¿¡!,:@#áéíóú\/]','',str(example["x"]))

    # if the sentence is less than 500 characters long (the max you can type in twitch) add spaces until it gets to 500 chars long
    if len(cleanexample) != 500:
        for a in range(500 - len(cleanexample)):
            cleanexample = cleanexample + " "
    for character in cleanexample:
        sentence.append(mapping[character])

    # print(sentence)
    y_train_ohe = tf.one_hot(sentence,depth=len(vocab)).numpy()
    # print(y_train_ohe)
    x_train.append(y_train_ohe)
    y_train.append(np.array(example["y"]))
    sentence = []

x_train = np.array(x_train)
y_train = np.array(y_train)
""" print(x_train[0][0:5],x_train[0][-5:],y_train[0]) """
print(x_train.shape[1],x_train.shape[2])
print(x_train.shape)
print(y_train.shape)

# Create the model
model = keras.Sequential()

model.add(layers.LSTM(256,activation="tanh",return_sequences=True,dropout=dropout,input_shape=(500,57)))

model.add(layers.LSTM(128,return_sequences=False,dropout=dropout))

model.add(layers.Dense(3,activation="softmax"))

optimizer = keras.optimizers.Adam(lr=0.01)

model.compile(optimizer=optimizer,loss="categorical_crossentropy",metrics=["accuracy"])

model.summary()

model.fit(x=x_train,y=y_train,epochs=15,shuffle=True,batch_size=25,validation_split=0.2)

model.save('model_py.h5')

''' tensorflowjs.converters.save_keras_model(model,"./modelo_js") '''

训练样本在处理之前看起来像这样。 [1，0，0]表示正常，[0，1，0]粗鲁的语言，[0，0，1]表示垃圾邮件：

"x": "sentence","y": [
            1,0
        ]

经过处理后，它们看起来像这样。我将它们编码为矢量，长度为57，即词汇量。以1开头的数组是空格：

[[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1.
  0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0. 0. 0. 0. 0. 0.]]
  ....
 [1 0 0]

在对每个类的1600个示例进行了训练之后，validation_split = 0.2，结果如下：

(4800,500,57)
(4800,3)
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
lstm (LSTM)                  (None,256)          321536
_________________________________________________________________
lstm_1 (LSTM)                (None,128)               197120
_________________________________________________________________
dense (Dense)                (None,3)                 387
=================================================================
Total params: 519,043
Trainable params: 519,043
Non-trainable params: 0
_________________________________________________________________
Epoch 1/15
2020-09-09 12:35:47.606648: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cublas64_10.dll
2020-09-09 12:35:47.872095: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cudnn64_7.dll
154/154 [==============================] - 13s 87ms/step - loss: 1.0811 - accuracy: 0.4120 - val_loss: 2.0132 - val_accuracy: 0.0219
Epoch 2/15
154/154 [==============================] - 12s 78ms/step - loss: 1.0577 - accuracy: 0.4177 - val_loss: 2.0314 - val_accuracy: 0.0000e+00
Epoch 3/15
154/154 [==============================] - 12s 76ms/step - loss: 1.0457 - accuracy: 0.4154 - val_loss: 1.6968 - val_accuracy: 0.0000e+00
Epoch 4/15
154/154 [==============================] - 12s 76ms/step - loss: 1.0506 - accuracy: 0.4161 - val_loss: 1.7731 - val_accuracy: 0.0000e+00
Epoch 5/15
154/154 [==============================] - 11s 73ms/step - loss: 1.0511 - accuracy: 0.4313 - val_loss: 1.9052 - val_accuracy: 0.0000e+00
Epoch 6/15
154/154 [==============================] - 12s 75ms/step - loss: 1.0473 - accuracy: 0.4104 - val_loss: 1.6291 - val_accuracy: 0.0000e+00
Epoch 7/15
154/154 [==============================] - 13s 84ms/step - loss: 1.0464 - accuracy: 0.4135 - val_loss: 1.8916 - val_accuracy: 0.0000e+00
Epoch 8/15
154/154 [==============================] - 12s 76ms/step - loss: 1.0404 - accuracy: 0.4208 - val_loss: 1.8094 - val_accuracy: 0.0000e+00
Epoch 9/15
154/154 [==============================] - 12s 76ms/step - loss: 1.0449 - accuracy: 0.4096 - val_loss: 1.9690 - val_accuracy: 0.0219
Epoch 10/15
154/154 [==============================] - 12s 77ms/step - loss: 1.0489 - accuracy: 0.4104 - val_loss: 1.9596 - val_accuracy: 0.0000e+00
Epoch 11/15
154/154 [==============================] - 13s 83ms/step - loss: 1.0455 - accuracy: 0.4141 - val_loss: 1.8082 - val_accuracy: 0.0000e+00
Epoch 12/15
154/154 [==============================] - 12s 76ms/step - loss: 1.0465 - accuracy: 0.4219 - val_loss: 1.7066 - val_accuracy: 0.0000e+00
Epoch 13/15
154/154 [==============================] - 12s 75ms/step - loss: 1.0424 - accuracy: 0.4161 - val_loss: 1.5192 - val_accuracy: 0.0000e+00
Epoch 14/15
154/154 [==============================] - 12s 75ms/step - loss: 1.0481 - accuracy: 0.4154 - val_loss: 1.5999 - val_accuracy: 0.0000e+00
Epoch 15/15
154/154 [==============================] - 12s 77ms/step - loss: 1.0476 - accuracy: 0.4008 - val_loss: 2.0612 - val_accuracy: 0.0000e+00 

奇怪的是，如果我增加验证间隔，结果会改善。这对我来说毫无意义，因为它的训练数据较少。

这是带有validate_split = 0.6

77/77 [==============================] - 8s 103ms/step - loss: 1.0352 - accuracy: 0.4432 - val_loss: 1.4233 - val_accuracy: 0.2313
Epoch 2/15
77/77 [==============================] - 7s 93ms/step - loss: 0.9906 - accuracy: 0.4443 - val_loss: 1.7316 - val_accuracy: 0.2937
Epoch 3/15
77/77 [==============================] - 7s 92ms/step - loss: 0.9863 - accuracy: 0.4812 - val_loss: 1.5367 - val_accuracy: 0.2313
Epoch 4/15
77/77 [==============================] - 7s 94ms/step - loss: 0.9874 - accuracy: 0.4635 - val_loss: 1.4075 - val_accuracy: 0.2937
Epoch 5/15
77/77 [==============================] - 7s 93ms/step - loss: 0.9905 - accuracy: 0.4594 - val_loss: 1.5759 - val_accuracy: 0.2937
Epoch 6/15
77/77 [==============================] - 7s 93ms/step - loss: 0.9808 - accuracy: 0.4703 - val_loss: 1.3886 - val_accuracy: 0.2937
Epoch 7/15
77/77 [==============================] - 7s 96ms/step - loss: 0.9815 - accuracy: 0.4781 - val_loss: 1.2495 - val_accuracy: 0.2313
Epoch 8/15
77/77 [==============================] - 7s 96ms/step - loss: 0.9824 - accuracy: 0.4698 - val_loss: 1.4516 - val_accuracy: 0.2313
Epoch 9/15
77/77 [==============================] - 7s 92ms/step - loss: 0.9916 - accuracy: 0.4573 - val_loss: 1.4488 - val_accuracy: 0.2313
Epoch 10/15
77/77 [==============================] - 7s 90ms/step - loss: 0.9858 - accuracy: 0.4760 - val_loss: 1.3868 - val_accuracy: 0.2313
Epoch 11/15
77/77 [==============================] - 7s 93ms/step - loss: 0.9861 - accuracy: 0.4734 - val_loss: 1.5702 - val_accuracy: 0.2313
Epoch 12/15
77/77 [==============================] - 7s 91ms/step - loss: 0.9880 - accuracy: 0.4630 - val_loss: 1.4439 - val_accuracy: 0.2313
Epoch 13/15
77/77 [==============================] - 7s 91ms/step - loss: 0.9796 - accuracy: 0.4865 - val_loss: 1.3597 - val_accuracy: 0.2313
Epoch 14/15
77/77 [==============================] - 7s 91ms/step - loss: 0.9832 - accuracy: 0.4745 - val_loss: 1.5791 - val_accuracy: 0.2313
Epoch 15/15
77/77 [==============================] - 7s 90ms/step - loss: 0.9919 - accuracy: 0.4760 - val_loss: 1.6243 - val_accuracy: 0.2313

并且validation_split = 0.8

39/39 [==============================] - 7s 171ms/step - loss: 1.1238 - accuracy: 0.4484 - val_loss: 1.3041 - val_accuracy: 0.3158
Epoch 2/15
39/39 [==============================] - 6s 143ms/step - loss: 0.9795 - accuracy: 0.4692 - val_loss: 1.2562 - val_accuracy: 0.3174
Epoch 3/15
39/39 [==============================] - 6s 146ms/step - loss: 0.9757 - accuracy: 0.4724 - val_loss: 1.3583 - val_accuracy: 0.3437
Epoch 4/15
39/39 [==============================] - 6s 149ms/step - loss: 0.9741 - accuracy: 0.4703 - val_loss: 1.3565 - val_accuracy: 0.2976
Epoch 5/15
39/39 [==============================] - 6s 148ms/step - loss: 0.9748 - accuracy: 0.4578 - val_loss: 1.3904 - val_accuracy: 0.2976
Epoch 6/15
39/39 [==============================] - 5s 137ms/step - loss: 0.9697 - accuracy: 0.4755 - val_loss: 1.3418 - val_accuracy: 0.2976
Epoch 7/15
39/39 [==============================] - 5s 136ms/step - loss: 0.9716 - accuracy: 0.4765 - val_loss: 1.3053 - val_accuracy: 0.3262
Epoch 8/15
39/39 [==============================] - 5s 136ms/step - loss: 0.9748 - accuracy: 0.4557 - val_loss: 1.3529 - val_accuracy: 0.2976
Epoch 9/15
39/39 [==============================] - 5s 140ms/step - loss: 0.9768 - accuracy: 0.4505 - val_loss: 1.3260 - val_accuracy: 0.2976
Epoch 10/15
39/39 [==============================] - 5s 136ms/step - loss: 0.9724 - accuracy: 0.4859 - val_loss: 1.3351 - val_accuracy: 0.3627
Epoch 11/15
39/39 [==============================] - 6s 143ms/step - loss: 0.9748 - accuracy: 0.4588 - val_loss: 1.3203 - val_accuracy: 0.3770
Epoch 12/15
39/39 [==============================] - 6s 144ms/step - loss: 0.9690 - accuracy: 0.4640 - val_loss: 1.3207 - val_accuracy: 0.3517
Epoch 13/15
39/39 [==============================] - 5s 137ms/step - loss: 0.9661 - accuracy: 0.4369 - val_loss: 1.3153 - val_accuracy: 0.3681
Epoch 14/15
39/39 [==============================] - 6s 141ms/step - loss: 0.9628 - accuracy: 0.4661 - val_loss: 1.3405 - val_accuracy: 0.2976
Epoch 15/15
39/39 [==============================] - 5s 137ms/step - loss: 0.9625 - accuracy: 0.4703 - val_loss: 1.3586 - val_accuracy: 0.3457

我尝试仅使用密集层，但结果却好得多，对我来说这是没有意义的，因为它们无法理解序列。但这清除了数据集错误的选项。

使用此配置（validation_split回到0.2）：

model = keras.Sequential()

model.add(layers.Input(shape=(500,57)))

model.add(layers.Flatten())

model.add(layers.Dense(256,activation="relu"))

model.add(layers.Dense(128,activation="relu"))

model.add(layers.Dense(64,activation="relu"))

model.add(layers.Dense(3,metrics=["accuracy"])

我得到这些结果：

154/154 [==============================] - 1s 6ms/step - loss: 0.7377 - accuracy: 0.7844 - val_loss: 1.4061 - val_accuracy: 0.0250
Epoch 2/15
154/154 [==============================] - 1s 4ms/step - loss: 0.3479 - accuracy: 0.8448 - val_loss: 0.8703 - val_accuracy: 0.6927
Epoch 3/15
154/154 [==============================] - 1s 4ms/step - loss: 0.3033 - accuracy: 0.8794 - val_loss: 1.4597 - val_accuracy: 0.6938
Epoch 4/15
154/154 [==============================] - 1s 4ms/step - loss: 0.2899 - accuracy: 0.8966 - val_loss: 1.6684 - val_accuracy: 0.4896
Epoch 5/15
154/154 [==============================] - 1s 4ms/step - loss: 0.2447 - accuracy: 0.9042 - val_loss: 1.6465 - val_accuracy: 0.4812
Epoch 6/15
154/154 [==============================] - 1s 4ms/step - loss: 0.2269 - accuracy: 0.9211 - val_loss: 3.9954 - val_accuracy: 0.7312
Epoch 7/15
154/154 [==============================] - 1s 4ms/step - loss: 0.2071 - accuracy: 0.9201 - val_loss: 2.7729 - val_accuracy: 0.4698
Epoch 8/15
154/154 [==============================] - 1s 4ms/step - loss: 0.2081 - accuracy: 0.9302 - val_loss: 5.1325 - val_accuracy: 0.4229
Epoch 9/15
154/154 [==============================] - 1s 4ms/step - loss: 0.1581 - accuracy: 0.9378 - val_loss: 4.4410 - val_accuracy: 0.3688
Epoch 10/15
154/154 [==============================] - 1s 4ms/step - loss: 0.2184 - accuracy: 0.9333 - val_loss: 2.6669 - val_accuracy: 0.5396
Epoch 11/15
154/154 [==============================] - 1s 4ms/step - loss: 0.1673 - accuracy: 0.9341 - val_loss: 3.6476 - val_accuracy: 0.2750
Epoch 12/15
154/154 [==============================] - 1s 4ms/step - loss: 0.2111 - accuracy: 0.9443 - val_loss: 1.6768 - val_accuracy: 0.6885
Epoch 13/15
154/154 [==============================] - 1s 4ms/step - loss: 0.1210 - accuracy: 0.9547 - val_loss: 2.6785 - val_accuracy: 0.5406
Epoch 14/15
154/154 [==============================] - 1s 4ms/step - loss: 0.1179 - accuracy: 0.9542 - val_loss: 3.4468 - val_accuracy: 0.4385
Epoch 15/15
154/154 [==============================] - 1s 4ms/step - loss: 0.1265 - accuracy: 0.9469 - val_loss: 2.0159 - val_accuracy: 0.7083

该模型相当大地改变了val_accuracy，但是至少我知道它正在学习一些东西。

我已经尝试过（使用LSTM模型）：

-更改层数 -更改每层神经元的数量 -改变学习率 -将优化器更改为SGD -更改损失功能 -更改时期数 -更改训练样本的数量（复制每个样本） -使用亚当衰变

解决方法

我在这里看到两个问题：

• LSTM在单输入法下不能很好地工作。使用padded sequences（例如[5,6,8]）
• 在LSTM层之前，您需要embedding layer

仅更改数据（和损失函数），我举了一个基于您的体系结构（添加了嵌入层）的示例：

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
import numpy as np
import sys

np.set_printoptions(threshold=sys.maxsize)

X_train = ['They like my dog','I hate my cat','We will love my hamster','I dislike your llama']
X_test = ['We love our hamster','They hate our platypus']
y_train = [1,1,0]
y_test = [1,0]

labels = {0: 'negative',1: 'positive'}

encoder = keras.preprocessing.text.Tokenizer()

encoder.fit_on_texts(X_train)

X_train = encoder.texts_to_sequences(X_train)
X_test = encoder.texts_to_sequences(X_test)

max_length = max(map(len,X_train))

x_train = keras.preprocessing.sequence.pad_sequences(X_train,maxlen=max_length)
x_test = keras.preprocessing.sequence.pad_sequences(X_test,maxlen=max_length)

x_train = np.array(x_train)
x_test = np.array(x_test)
y_train = np.array(y_train)
y_test = np.array(y_test)

embedding_dim = 4
# print(x_train.shape[1],x_train.shape[2])
print(x_train.shape)
print(y_train.shape)

# Create the model
model = keras.Sequential()

model.add(layers.Embedding(len(encoder.index_word) + 1,embedding_dim))

model.add(layers.LSTM(8,activation="tanh",return_sequences=True,dropout=.2))

model.add(layers.LSTM(8,return_sequences=False,dropout=.2))

model.add(layers.Dense(2,activation="softmax"))

optimizer = keras.optimizers.Adam(lr=0.01)

model.compile(optimizer=optimizer,loss="sparse_categorical_crossentropy",metrics=["accuracy"])

model.build(input_shape=x_train.shape)
model.summary()

history = model.fit(x=x_train,y=y_train,epochs=25,shuffle=True,batch_size=25,validation_data=(x_test,y_test))

让我知道是否需要澄清。