使用 Tensorboard 实时监控训练并可视化模型架构

Posted 2023-02-16

【中文标题】使用 Tensorboard 实时监控训练并可视化模型架构

【英文标题】：Using Tensorboard to monitor training real time and visualize the model architecture

【发布时间】：2020-01-26 15:19:49

【问题描述】：

我正在学习使用 Tensorboard -- Tensorflow 2.0。

特别是，我想实时监控学习曲线，并直观地检查和交流我的模型的架构。

下面我将提供一个可重现示例的代码。

我有三个问题：

虽然训练结束后我得到了学习曲线，但我不知道应该怎么做才能实时监控它们

我从 Tensorboard 得到的学习曲线与 history.history 的情节不符。事实上，它的反转很奇怪且难以解释。

我无法理解图表。我已经训练了一个顺序模型，其中包含 5 个密集层和中间的 dropout 层。 Tensorboard 向我展示的是其中包含更多元素的东西。

我的代码如下：

from keras.datasets import boston_housing

(train_data, train_targets), (test_data, test_targets) = boston_housing.load_data()

inputs = Input(shape = (train_data.shape[1], ))
x1 = Dense(100, kernel_initializer = 'he_normal', activation = 'elu')(inputs)
x1a = Dropout(0.5)(x1)
x2 = Dense(100, kernel_initializer = 'he_normal', activation = 'elu')(x1a)
x2a = Dropout(0.5)(x2)
x3 = Dense(100, kernel_initializer = 'he_normal', activation = 'elu')(x2a)
x3a = Dropout(0.5)(x3)
x4 = Dense(100, kernel_initializer = 'he_normal', activation = 'elu')(x3a)
x4a = Dropout(0.5)(x4)
x5 = Dense(100, kernel_initializer = 'he_normal', activation = 'elu')(x4a)
predictions = Dense(1)(x5)
model = Model(inputs = inputs, outputs = predictions)

model.compile(optimizer = 'Adam', loss = 'mse')

logdir="logs\\fit\\" + datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard_callback = keras.callbacks.TensorBoard(log_dir=logdir)

history = model.fit(train_data, train_targets,
          batch_size= 32,
          epochs= 20,
          validation_data=(test_data, test_targets),
          shuffle=True,
          callbacks=[tensorboard_callback ])

plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])

plt.plot(history.history['val_loss'])

【参考方案1】：

我认为您可以做的是在您的模型上调用 .fit() 之前启动 TensorBoard。如果您使用的是 IPython（Jupyter 或 Colab），并且已经安装了 TensorBoard，那么您可以通过以下方式修改代码；

from keras.datasets import boston_housing

(train_data, train_targets), (test_data, test_targets) = boston_housing.load_data()

inputs = Input(shape = (train_data.shape[1], ))
x1 = Dense(100, kernel_initializer = 'he_normal', activation = 'relu')(inputs)
x1a = Dropout(0.5)(x1)
x2 = Dense(100, kernel_initializer = 'he_normal', activation = 'relu')(x1a)
x2a = Dropout(0.5)(x2)
x3 = Dense(100, kernel_initializer = 'he_normal', activation = 'relu')(x2a)
x3a = Dropout(0.5)(x3)
x4 = Dense(100, kernel_initializer = 'he_normal', activation = 'relu')(x3a)
x4a = Dropout(0.5)(x4)
x5 = Dense(100, kernel_initializer = 'he_normal', activation = 'relu')(x4a)
predictions = Dense(1)(x5)
model = Model(inputs = inputs, outputs = predictions)

model.compile(optimizer = 'Adam', loss = 'mse')

logdir="logs\\fit\\" + datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard_callback = keras.callbacks.TensorBoard(log_dir=logdir)

在另一个单元格中，你可以运行；

# Magic func to use TensorBoard directly in IPython
%load_ext tensorboard

通过在另一个单元格中运行来启动 TensorBoard；

# Launch TensorBoard with objects in the log directory
# This should launch tensorboard in your browser, but you may not see your metadata.
%tensorboard --logdir=logdir 

您终于可以在另一个单元格中的模型上调用.fit()；

history = model.fit(train_data, train_targets,
          batch_size= 32,
          epochs= 20,
          validation_data=(test_data, test_targets),
          shuffle=True,
          callbacks=[tensorboard_callback ])

plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])

如果您不使用 IPython，您可能只需要在训练模型期间或之前启动 TensorBoard 以实时监控它。