1. 数据预处理
from google.colab import drive drive.mount('/content/gdrive') import tensorflow as tf from tensorflow import keras from tensorflow.keras import layers data = keras.datasets.imdb #加载数据 (x_train,y_train),(x_test,y_test)= data.load_data(num_words=10000)
我们发现这些评论的序列长短不一，为了便于后面神经网络的训练，我们需要对较长的序列进行截断，对较短的序列进行填充。我们可以使用keras.preprocessing.sequence.pad_sequences方法对其进行填充。
print("截断前前5个样本序列长度:{}".format([len(x) for x in x_train[:5]])) x_train = keras.preprocessing.sequence.pad_sequences(x_train,300) x_test = keras.preprocessing.sequence.pad_sequences(x_test,300) print("截断后前5个样本序列长度:{}".format([len(x) for x in x_test[:5]]))
截断前前5个样本序列长度:[218, 189, 141, 550, 147] 截断后前5个样本序列长度:[300, 300, 300, 300, 300] • 1 • 2
2. 构建模型
model = keras.Sequential() #输出为（None.300,50），意思是将这300个数据转化为300个50维的特征向量 model.add(layers.Embedding(10000,50,input_length=300)) model.add(layers.Flatten()) model.add(layers.Dense(128,activation= "relu")) y = model.add(layers.Dense(1,activation="sigmoid")) model.summary() model.compile(optimizer = keras.optimizers.Adam(lr=0.001), loss = "binary_crossentropy", metrics = ["acc"] ) history = model.fit(x_train,y_train,epochs=10,batch_size=256,validation_data=(x_test,y_test))
Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= embedding (Embedding) (None, 300, 50) 500000 _________________________________________________________________ flatten (Flatten) (None, 15000) 0 _________________________________________________________________ dense (Dense) (None, 128) 1920128 _________________________________________________________________ dense_1 (Dense) (None, 1) 129 ================================================================= Total params: 2,420,257 Trainable params: 2,420,257 Non-trainable params: 0 _________________________________________________________________ Epoch 1/10 98/98 [==============================] - 3s 33ms/step - loss: 0.5086 - acc: 0.7221 - val_loss: 0.2980 - val_acc: 0.8746 Epoch 2/10 98/98 [==============================] - 3s 32ms/step - loss: 0.1643 - acc: 0.9394 - val_loss: 0.3016 - val_acc: 0.8769 Epoch 3/10 98/98 [==============================] - 3s 32ms/step - loss: 0.0391 - acc: 0.9924 - val_loss: 0.3552 - val_acc: 0.8732 Epoch 4/10 98/98 [==============================] - 3s 33ms/step - loss: 0.0079 - acc: 0.9995 - val_loss: 0.3952 - val_acc: 0.8736 Epoch 5/10 98/98 [==============================] - 3s 34ms/step - loss: 0.0027 - acc: 0.9999 - val_loss: 0.4250 - val_acc: 0.8734 Epoch 6/10 98/98 [==============================] - 3s 34ms/step - loss: 0.0014 - acc: 1.0000 - val_loss: 0.4462 - val_acc: 0.8742 Epoch 7/10 98/98 [==============================] - 3s 32ms/step - loss: 8.4250e-04 - acc: 1.0000 - val_loss: 0.4639 - val_acc: 0.8747 Epoch 8/10 98/98 [==============================] - 3s 31ms/step - loss: 5.9368e-04 - acc: 1.0000 - val_loss: 0.4794 - val_acc: 0.8742 Epoch 9/10 98/98 [==============================] - 3s 33ms/step - loss: 4.3011e-04 - acc: 1.0000 - val_loss: 0.4925 - val_acc: 0.8744 Epoch 10/10 98/98 [==============================] - 3s 34ms/step - loss: 3.2738e-04 - acc: 1.0000 - val_loss: 0.5045 - val_acc: 0.8744
3. 模型分析
import matplotlib.pyplot as plt plt.plot(history.epoch, history.history.get("acc"),label="acc") plt.plot(history.epoch, history.history.get("val_acc"),label="val_acc") plt.legend() plt.show()
import matplotlib.pyplot as plt plt.plot(history.epoch, history.history.get("loss"),label="loss") plt.plot(history.epoch, history.history.get("val_loss"),label="val_loss") plt.legend() plt.show()
训练数据的准确率远高于测试数据的准确率，这是过拟合的问题，我们可以添加dropout,l2正则化等方法解决过拟合。
4. 模型的优化
from tensorflow.keras import regularizers model = keras.Sequential() #输出为（None.300,50），意思是将这300个数据转化为300个50维的特征向量 model.add(layers.Embedding(10000,50,input_length=300,embeddings_regularizer=regularizers.l2(0.001))) model.add(layers.GlobalAvgPool1D()) model.add(layers.BatchNormalization()) model.add(layers.Dropout(0.5)) model.add(layers.Dense(256,activation= "relu",kernel_regularizer=regularizers.l2(0.001))) model.add(layers.BatchNormalization()) model.add(layers.Dropout(0.5)) y = model.add(layers.Dense(1,activation="sigmoid",kernel_regularizer=regularizers.l2(0.001))) model.summary() model.compile(optimizer = keras.optimizers.Adam(lr=0.001), loss = "binary_crossentropy", metrics = ["acc"] ) history = model.fit(x_train,y_train,epochs=10,batch_size=256,validation_data=(x_test,y_test))
Model: "sequential_1" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= embedding_1 (Embedding) (None, 300, 50) 500000 _________________________________________________________________ global_average_pooling1d (Gl (None, 50) 0 _________________________________________________________________ batch_normalization (BatchNo (None, 50) 200 _________________________________________________________________ dropout (Dropout) (None, 50) 0 _________________________________________________________________ dense_2 (Dense) (None, 256) 13056 _________________________________________________________________ batch_normalization_1 (Batch (None, 256) 1024 _________________________________________________________________ dropout_1 (Dropout) (None, 256) 0 _________________________________________________________________ dense_3 (Dense) (None, 1) 257 ================================================================= Total params: 514,537 Trainable params: 513,925 Non-trainable params: 612 _________________________________________________________________ 98/98 [==============================] - 3s 26ms/step - loss: 0.9417 - acc: 0.6739 - val_loss: 0.8925 - val_acc: 0.5110 Epoch 2/10 98/98 [==============================] - 2s 24ms/step - loss: 0.5110 - acc: 0.8672 - val_loss: 0.8516 - val_acc: 0.5066 Epoch 3/10 98/98 [==============================] - 2s 25ms/step - loss: 0.4200 - acc: 0.8984 - val_loss: 0.8033 - val_acc: 0.6794 Epoch 4/10 98/98 [==============================] - 3s 26ms/step - loss: 0.3850 - acc: 0.9087 - val_loss: 0.7196 - val_acc: 0.8363 Epoch 5/10 98/98 [==============================] - 2s 24ms/step - loss: 0.3500 - acc: 0.9196 - val_loss: 0.6267 - val_acc: 0.7583 Epoch 6/10 98/98 [==============================] - 2s 25ms/step - loss: 0.3360 - acc: 0.9216 - val_loss: 0.5075 - val_acc: 0.8535 Epoch 7/10 98/98 [==============================] - 2s 25ms/step - loss: 0.3158 - acc: 0.9280 - val_loss: 0.4609 - val_acc: 0.8490 Epoch 8/10 98/98 [==============================] - 2s 24ms/step - loss: 0.3039 - acc: 0.9316 - val_loss: 0.4430 - val_acc: 0.8666 Epoch 9/10 98/98 [==============================] - 2s 24ms/step - loss: 0.2975 - acc: 0.9316 - val_loss: 0.4554 - val_acc: 0.8658 Epoch 10/10 98/98 [==============================] - 2s 24ms/step - loss: 0.2858 - acc: 0.9362 - val_loss: 0.4749 - val_acc: 0.8608
import matplotlib.pyplot as plt plt.plot(history.epoch, history.history.get("acc"),label="acc") plt.plot(history.epoch, history.history.get("val_acc"),label="val_acc") plt.legend() plt.show()
import matplotlib.pyplot as plt plt.plot(history.epoch, history.history.get("loss"),label="loss") plt.plot(history.epoch, history.history.get("val_loss"),label="val_loss") plt.legend() plt.show()
从图中我们可以知道加入dropout,l2正则换，批量归一化具有抑制过拟合的作用。
