恢复培训tf.keras张量板

Question

当我继续训练模型并在张量板上可视化进度时遇到了一些问题。

我的问题是，如何在不手动指定任何纪元的情况下从同一步骤恢复训练？如果可能，只需加载已保存的模型，它就可以以某种方式从已保存的优化器中读取global_step，然后从那里继续进行训练。

我在下面提供了一些代码来重现类似的错误。

import tensorflow as tf
from tensorflow.keras.callbacks import TensorBoard
from tensorflow.keras.models import load_model

mnist = tf.keras.datasets.mnist

(x_train, y_train),(x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0

model = tf.keras.models.Sequential([
  tf.keras.layers.Flatten(input_shape=(28, 28)),
  tf.keras.layers.Dense(512, activation=tf.nn.relu),
  tf.keras.layers.Dropout(0.2),
  tf.keras.layers.Dense(10, activation=tf.nn.softmax)
])
model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

model.fit(x_train, y_train, epochs=10, callbacks=[Tensorboard()])
model.save('./final_model.h5', include_optimizer=True)

del model

model = load_model('./final_model.h5')
model.fit(x_train, y_train, epochs=10, callbacks=[Tensorboard()])

您可以使用以下命令运行tensorboard：

tensorboard --logdir ./logs

Answer 1

您可以将功能initial_epoch中的参数model.fit()设置为您要开始训练的时期的编号。考虑到模型会一直训练到到达索引epochs的时期（而不是epochs给出的多次迭代）。 在您的示例中，如果您想再训练10个纪元，应该是：

model.fit(x_train, y_train, initial_epoch=9, epochs=19, callbacks=[Tensorboard()])

它将允许您以正确的方式在Tensorboard上可视化绘图。 在docs中可以找到有关这些参数的更多信息。

Answer 2

这很简单。在训练模型时创建检查点，然后使用这些检查点从您离开的地方继续训练。

import tensorflow as tf
from tensorflow.keras.callbacks import TensorBoard
from tensorflow.keras.callbacks import ModelCheckpoint
from tensorflow.keras.models import load_model

mnist = tf.keras.datasets.mnist

(x_train, y_train),(x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0

model = tf.keras.models.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28)),
    tf.keras.layers.Dense(512, activation=tf.nn.relu),
    tf.keras.layers.Dropout(0.2),
    tf.keras.layers.Dense(10, activation=tf.nn.softmax)
])
model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

model.fit(x_train, y_train, epochs=10, callbacks=[Tensorboard()])
model.save('./final_model.h5', include_optimizer=True)

model = load_model('./final_model.h5')

callbacks = list()

tensorboard = Tensorboard()
callbacks.append(tensorboard)

file_path = "model-{epoch:02d}-{loss:.4f}.hdf5"

# now here you can create checkpoints and save according to your need
# here period is the no of epochs after which to save the model every time during training
# another option is save_weights_only, for your case it should be false
checkpoints = ModelCheckpoint(file_path, monitor='loss', verbose=1, period=1, save_weights_only=False)
callbacks.append(checkpoints)

model.fit(x_train, y_train, epochs=10, callbacks=callbacks)

此后，只需从要重新开始训练的位置加载检查点

model = load_model(checkpoint_of_choice)
model.fit(x_train, y_train, epochs=10, callbacks=callbacks)

您完成了。

如果您还有其他疑问，请告诉我。

Answer 3

这里是示例代码，以防有人需要。它实现了Abhinav Anand提出的想法：

mca = ModelCheckpoint(join(dir, 'model_{epoch:03d}.h5'),
                      monitor = 'loss',
                      save_best_only = False)
tb = TensorBoard(log_dir = join(dir, 'logs'),
                 write_graph = True,
                 write_images = True)
files = sorted(glob(join(fold_dir, 'model_???.h5')))
if files:
    model_file = files[-1]
    initial_epoch = int(model_file[-6:-3])
    print('Resuming using saved model %s.' % model_file)
    model = load_model(model_file)
else:
    model = nn.model()
    initial_epoch = 0
model.fit(x_train,
          y_train,
          epochs = 100,
          initial_epoch = initial_epoch,
          callbacks = [mca, tb])

用您自己的用于定义模型的函数替换nn.model()。