第7章代码及详解

Author: sunjiaxin111

chapter7/custom-callback.py

@@ -0,0 +1,24 @@
+import keras
+import numpy as np
+
+
+class ActivationLogger(keras.callbacks.Callback):
+
+    def set_model(self, model):
+        # 这个方法会在父模型训练前被调用
+        self.model = model
+        layer_outputs = [layer.output for layer in model.layers]
+        # 这是一个model实例，返回每层的激活值
+        self.activations_model = keras.models.Model(model.input, layer_outputs)
+
+    def on_epoch_end(self, epoch, logs=None):
+        if self.validation_data is None:
+            raise RuntimeError('Requires validation_data.')
+
+        # 获得验证集的第一个输入样本
+        validation_sample = self.validation_data[0][0:1]
+        activations = self.activations_model.predict(validation_sample)
+        # 保存数组到硬盘
+        f = open('activations_at_epoch_' + str(epoch) + '.npz', 'w')
+        np.savez(f, activations)
+        f.close()

chapter7/functional_api.py

@@ -0,0 +1,50 @@
+from keras import Input, layers
+from keras.models import Sequential, Model
+import numpy as np
+
+# 把层看作一个函数
+# 这是一个张量
+input_tensor = Input(shape=(32,))
+
+# 一层是一个函数
+dense = layers.Dense(32, activation='relu')
+
+# 一层调用一个张量，然后返回一个张量
+output_tensor = dense(input_tensor)
+
+# 与序列模型等价的functional API
+# 序列模型
+seq_model = Sequential()
+seq_model.add(layers.Dense(32, activation='relu', input_shape=(64,)))
+seq_model.add(layers.Dense(32, activation='relu'))
+seq_model.add(layers.Dense(10, activation='softmax'))
+
+# functional API
+input_tensor = Input(shape=(64,))
+x = layers.Dense(32, activation='relu')(input_tensor)
+x = layers.Dense(32, activation='relu')(x)
+output_tensor = layers.Dense(10, activation='softmax')(x)
+
+# 把一个输入张量和一个输出张量变成一个模型
+model = Model(input_tensor, output_tensor)
+
+model.summary()
+
+# 如果输入张量和输出张量没有关联,将会报错
+# unrelated_input = Input(shape=(32,))
+# bad_model = Model(unrelated_input, output_tensor)
+
+# 编译、训练和评估是一样的
+# 编译模型
+model.compile(optimizer='rmsprop',
+              loss='categorical_crossentropy')
+
+# 生成Numpy的数据去训练
+x_train = np.random.random((1000, 64))
+y_train = np.random.random((1000, 10))
+
+# 训练模型
+model.fit(x_train, y_train, epochs=10, batch_size=128)
+
+# 评估模型
+score = model.evaluate(x_train, y_train)

chapter7/inception-module.py

@@ -0,0 +1,20 @@
+from keras import layers
+from keras import Input
+
+# 假定x是一个四维张量
+x = Input(shape=(None,), dtype='int32')
+
+# 每个分支的输出必须是同样大小
+branch_a = layers.Conv2D(128, 1, activation='relu', strides=2)(x)
+
+branch_b = layers.Conv2D(128, 1, activation='relu')(x)
+branch_b = layers.Conv2D(128, 3, activation='relu', strides=2)(branch_b)
+
+branch_c = layers.AveragePooling2D(3, strides=2, activation='relu')(x)
+branch_c = layers.Conv2D(128, 3, activation='relu')(branch_c)
+
+branch_d = layers.Conv2D(128, 1, activation='relu')(x)
+branch_d = layers.Conv2D(128, 3, activation='relu')(branch_d)
+branch_d = layers.Conv2D(128, 3, activation='relu', strides=2)(branch_d)
+
+output = layers.concatenate([branch_a, branch_b, branch_c, branch_d], axis=-1)

chapter7/layer-weight-sharing.py

@@ -0,0 +1,21 @@
+from keras import layers
+from keras import Input
+from keras.models import Model
+
+# 实例化一个LSTM层
+lstm = layers.LSTM(32)
+
+left_input = Input(shape=(None, 128))
+left_output = lstm(left_input)
+
+# 共用lstm层的权重
+right_input = Input(shape=(None, 128))
+right_output = lstm(right_input)
+
+# 构建一个分类器
+merged = layers.concatenate([left_output, right_output], axis=-1)
+predictions = layers.Dense(1, activation='sigmoid')(merged)
+
+# 实例化、训练模型
+model = Model([left_input, right_input], predictions)
+# model.fit([left_data, right_data], targets)

chapter7/logs/checkpoint

@@ -0,0 +1,2 @@
+model_checkpoint_path: "keras_embedding.ckpt-0"
+all_model_checkpoint_paths: "keras_embedding.ckpt-0"

chapter7/logs/events.out.tfevents.1515131424.DESKTOP-JBB0OFG

@@ -0,0 +1,3 @@
+embeddings {
+  tensor_name: "embed/embeddings:0"
+}

chapter7/logs/events.out.tfevents.1515133925.DESKTOP-JBB0OFG

@@ -0,0 +1,50 @@
+from keras.models import Model
+from keras import layers
+from keras import Input
+import numpy as np
+
+text_vocabulary_size = 10000
+question_vocabulary_size = 10000
+answer_vocabulary_size = 500
+
+# 文本输入是变长的序列
+text_input = Input(shape=(None,), dtype='int32', name='text')
+
+# 把文本输入变成大小为64的词向量
+embedded_text = layers.Embedding(64, text_vocabulary_size)(text_input)
+
+# LSTM
+encoded_text = layers.LSTM(32)(embedded_text)
+
+# 对question做相同步骤
+question_input = Input(shape=(None,), dtype='int32', name='question')
+embedded_question = layers.Embedding(32, question_vocabulary_size)(question_input)
+encoded_question = layers.LSTM(16)(embedded_question)
+
+# 连接encoded_text和encoded_question
+concatenated = layers.concatenate([encoded_text, encoded_question], axis=-1)
+
+# 添加一个softmax分类器
+answer = layers.Dense(answer_vocabulary_size, activation='softmax')(concatenated)
+
+# 指定二个输入和一个输出
+model = Model([text_input, question_input], answer)
+model.compile(optimizer='rmsprop',
+              loss='categorical_crossentropy',
+              metrics=['acc'])
+
+# 训练模型
+# 生成一些模拟数据
+num_samples = 500
+max_length = 500
+text = np.random.randint(1, text_vocabulary_size, size=(num_samples, max_length))
+question = np.random.randint(1, question_vocabulary_size, size=(num_samples, max_length))
+
+# one-hot编码
+answers = np.random.randint(0, 1, size=(num_samples, answer_vocabulary_size))
+
+# 输入list
+model.fit([text, question], answers, epochs=10, batch_size=128)
+
+# 字典方式
+model.fit({'text': text, 'question': question}, answers, epochs=10, batch_size=128)

chapter7/logs/keras_embedding.ckpt-0.data-00000-of-00001

@@ -0,0 +1,24 @@
+from keras import layers
+from keras import Input
+
+# 假定x是一个四维张量
+x = Input(shape=(None,), dtype='int32')
+y = layers.Conv2D(128, 3, activation='relu')(x)
+y = layers.Conv2D(128, 3, activation='relu')(y)
+y = layers.Conv2D(128, 3, activation='relu')(y)
+
+# 把x加到y的输出中
+y = layers.add([y, x])
+
+# --------------------------------------------
+# 当特征大小不一样时
+# 假定x是一个四维张量
+x = Input(shape=(None,), dtype='int32')
+y = layers.Conv2D(128, 3, activation='relu')(x)
+y = layers.Conv2D(128, 3, activation='relu')(y)
+y = layers.MaxPooling2D(2, strides=2)(y)
+
+# 使用线性的降低取样
+residual = layers.Conv2D(1, strides=2)(x)
+
+y = layers.add([y, residual])

chapter7/logs/keras_embedding.ckpt-0.index

@@ -0,0 +1,16 @@
+from keras import layers
+from keras import applications
+from keras import Input
+
+# 共用convolutional base
+xception_base = applications.Xception(weights=None, include_top=False)
+
+# 输入为250*250的RGB图像
+left_input = Input(shape=(250, 250, 3))
+right_input = Input(shape=(250, 250, 3))
+
+# 调用xception_base两次
+left_features = xception_base(left_input)
+right_features = xception_base(right_input)
+
+merged_features = layers.concatenate([left_features, right_features], axis=-1)

chapter7/logs/keras_embedding.ckpt-0.meta

@@ -0,0 +1,44 @@
+import keras
+from keras import layers
+from keras.datasets import imdb
+from keras.preprocessing import sequence
+from keras.utils import plot_model
+
+max_features = 2000
+max_len = 500
+
+(x_train, y_train), (x_test, y_test) = imdb.load_data(num_words=max_features)
+x_train = sequence.pad_sequences(x_train, maxlen=max_len)
+x_test = sequence.pad_sequences(x_test, maxlen=max_len)
+
+model = keras.models.Sequential()
+model.add(layers.Embedding(max_features, 128, input_length=max_len, name='embed'))
+model.add(layers.Conv1D(32, 7, activation='relu'))
+model.add(layers.MaxPooling1D(5))
+model.add(layers.Conv1D(32, 7, activation='relu'))
+model.add(layers.GlobalMaxPooling1D())
+model.add(layers.Dense(1))
+model.summary()
+model.compile(optimizer='rmsprop',
+              loss='binary_crossentropy',
+              metrics=['acc'])
+
+# 画图
+plot_model(model, show_shapes=True, to_file='model.png')
+
+callbacks = [
+    keras.callbacks.TensorBoard(
+        # 日志目录
+        log_dir='logs',
+        # 每个epoch记录一次activation直方图
+        histogram_freq=1,
+        # 每个epoch记录一次词向量数据
+        embeddings_freq=1,
+    )
+]
+
+history = model.fit(x_train, y_train,
+                    epochs=20,
+                    batch_size=128,
+                    validation_split=0.2,
+                    callbacks=callbacks)

chapter7/logs/keras_embedding.ckpt-15.data-00000-of-00001

@@ -0,0 +1,60 @@
+from keras import layers
+from keras import Input
+from keras.models import Model
+
+vocabulary_size = 50000
+num_income_groups = 10
+
+posts_input = Input(shape=(None,), dtype='int32', name='posts')
+embedded_posts = layers.Embedding(256, vocabulary_size)(posts_input)
+x = layers.Conv1D(128, 5, activation='relu')(embedded_posts)
+x = layers.MaxPooling1D(5)(x)
+x = layers.Conv1D(256, 5, activation='relu')(x)
+x = layers.Conv1D(256, 5, activation='relu')(x)
+x = layers.MaxPooling1D(5)(x)
+x = layers.Conv1D(256, 5, activation='relu')(x)
+x = layers.Conv1D(256, 5, activation='relu')(x)
+x = layers.GlobalMaxPooling1D()(x)
+x = layers.Dense(128, activation='relu')(x)
+
+# 3个输出
+age_prediction = layers.Dense(1, name='age')(x)
+income_prediction = layers.Dense(num_income_groups, activation='softmax', name='income')(x)
+gender_prediction = layers.Dense(1, activation='sigmoid', name='gender')(x)
+
+model = Model(posts_input, [age_prediction, income_prediction, gender_prediction])
+
+# 多个loss
+model.compile(optimizer='rmsprop',
+              loss=['mse', 'categorical_crossentropy', 'binary_crossentropy'])
+
+# 等价编译
+model.compile(optimizer='rmsprop',
+              loss={'age': 'mse',
+                    'income': 'categorical_crossentropy',
+                    'gender': 'binary_crossentropy'})
+
+# loss weighting
+model.compile(optimizer='rmsprop',
+              loss=['mse', 'categorical_crossentropy', 'binary_crossentropy'],
+              loss_weights=[0.25, 1., 10.])
+
+# 等价编译
+model.compile(optimizer='rmsprop',
+              loss={'age': 'mse',
+                    'income': 'categorical_crossentropy',
+                    'gender': 'binary_crossentropy'},
+              loss_weights={'age': 0.25,
+                            'income': 1.,
+                            'gender': 10.})
+'''
+# 训练模型
+model.fit(posts_input, [age_targets, income_targets, gender_targets],
+          epochs=10, batch_size=64)
+
+# 等价训练
+model.fit(posts_input, {'age': age_targets,
+                        'income': income_targets,
+                        'gender': gender_targets},
+          epochs=10, batch_size=64)
+'''

chapter7/logs/keras_embedding.ckpt-15.index

@@ -0,0 +1,46 @@
+import keras
+from keras import Model
+
+callbacks_list = [
+    # 这个callback会在模型停止优化时打断训练
+    keras.callbacks.EarlyStopping(
+        # 监测验证集的准确度,这边应该是val_acc？？？
+        monitor='acc',
+        # 当准确率已经停止优化超过1个epochs，也就是2个epochs，训练将会被打断
+        patience=1,
+    ),
+    # 这个callback将会在每个epoch后保存当前的权重
+    keras.callbacks.ModelCheckpoint(
+        filepath='my_model.h5',
+        # 下面2个参数意味着除非val_loss有优化，不然model文件不会被重写
+        monitor='val_loss',
+        save_best_only=True,
+    )
+]
+
+model = Model()
+
+# 因为我们监测acc指标，acc应该成为一个度量标准
+model.compile(optimizer='rmsprop',
+              loss='binary_crossentropy',
+              metrics=['acc'])
+
+'''
+model.fit(x, y,
+          epochs=10,
+          batch_size=32,
+          callbacks=callbacks_list,
+          validation_data=(x_val, y_val))
+'''
+
+# 使用ReduceLROnPlateau
+callbacks_list = [
+    keras.callbacks.ReduceLROnPlateau(
+        # 监测验证集的loss
+        monitor='val_loss',
+        # 触发时把学习速率除以10
+        factor=0.1,
+        # 在验证集loss停止优化10个epochs时触发
+        patience=10,
+    )
+]