将 LoggingTensorHook 和 StopAtStepHook 迁移到 Keras 回调

在 TensorFlow 1 中，可以使用 tf.estimator.LoggingTensorHook 监视和记录张量，而 tf.estimator.StopAtStepHook 则在使用 tf.estimator.Estimator 进行训练时有助于在指定步骤停止训练。本笔记本演示了如何使用带有 Model.fit 的自定义 Keras 回调 (tf.keras.callbacks.Callback) 从这些 API 迁移到 TensorFlow 2 中的对应项。

Keras 回调是在内置 Keras Model.fit/Model.evaluate/Model.predict API 中的训练/评估/预测期间的不同点调用的对象。可以在 tf.keras.callbacks.Callback API 文档以及编写自己的回调和使用内置方法进行训练和评估（使用回调 部分）指南中详细了解回调。要从 TensorFlow 1 中的 SessionRunHook 迁移到 TensorFlow 2 中的 Keras 回调，请查看迁移使用辅助逻辑的训练指南。

安装

从导入和用于演示目的的简单数据集开始：

import tensorflow as tf
import tensorflow.compat.v1 as tf1

2022-12-14 20:26:46.140102: W tensorflow/compiler/xla/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer.so.7'; dlerror: libnvinfer.so.7: cannot open shared object file: No such file or directory
2022-12-14 20:26:46.140199: W tensorflow/compiler/xla/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer_plugin.so.7'; dlerror: libnvinfer_plugin.so.7: cannot open shared object file: No such file or directory
2022-12-14 20:26:46.140208: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Cannot dlopen some TensorRT libraries. If you would like to use Nvidia GPU with TensorRT, please make sure the missing libraries mentioned above are installed properly.

features = [[1., 1.5], [2., 2.5], [3., 3.5]]
labels = [[0.3], [0.5], [0.7]]

# Define an input function.
def _input_fn():
  return tf1.data.Dataset.from_tensor_slices((features, labels)).batch(1)

TensorFlow 1：使用 tf.estimator API 记录张量和停止训练

在 TensorFlow 1 中，定义各种钩子来控制训练行为。随后，将这些钩子传递给 tf.estimator.EstimatorSpec。

在下面的示例中：

要监视/记录张量（例如模型权重或损失），可以使用 tf.estimator.LoggingTensorHook（tf.train.LoggingTensorHook 是它的别名）。
要在特定步骤停止训练，请使用 tf.estimator.StopAtStepHook（tf.train.StopAtStepHook 是它的别名）。

def _model_fn(features, labels, mode):
  dense = tf1.layers.Dense(1)
  logits = dense(features)
  loss = tf1.losses.mean_squared_error(labels=labels, predictions=logits)
  optimizer = tf1.train.AdagradOptimizer(0.05)
  train_op = optimizer.minimize(loss, global_step=tf1.train.get_global_step())

  # Define the stop hook.
  stop_hook = tf1.train.StopAtStepHook(num_steps=2)

  # Access tensors to be logged by names.
  kernel_name = tf.identity(dense.weights[0])
  bias_name = tf.identity(dense.weights[1])
  logging_weight_hook = tf1.train.LoggingTensorHook(
      tensors=[kernel_name, bias_name],
      every_n_iter=1)
  # Log the training loss by the tensor object.
  logging_loss_hook = tf1.train.LoggingTensorHook(
      {'loss from LoggingTensorHook': loss},
      every_n_secs=3)

  # Pass all hooks to `EstimatorSpec`.
  return tf1.estimator.EstimatorSpec(mode,
                                     loss=loss,
                                     train_op=train_op,
                                     training_hooks=[stop_hook,
                                                     logging_weight_hook,
                                                     logging_loss_hook])

estimator = tf1.estimator.Estimator(model_fn=_model_fn)

# Begin training.
# The training will stop after 2 steps, and the weights/loss will also be logged.
estimator.train(_input_fn)

INFO:tensorflow:Using default config.
WARNING:tensorflow:Using temporary folder as model directory: /tmpfs/tmp/tmp9nj_75jg
INFO:tensorflow:Using config: {'_model_dir': '/tmpfs/tmp/tmp9nj_75jg', '_tf_random_seed': None, '_save_summary_steps': 100, '_save_checkpoints_steps': None, '_save_checkpoints_secs': 600, '_session_config': allow_soft_placement: true
graph_options {
  rewrite_options {
    meta_optimizer_iterations: ONE
  }
}
, '_keep_checkpoint_max': 5, '_keep_checkpoint_every_n_hours': 10000, '_log_step_count_steps': 100, '_train_distribute': None, '_device_fn': None, '_protocol': None, '_eval_distribute': None, '_experimental_distribute': None, '_experimental_max_worker_delay_secs': None, '_session_creation_timeout_secs': 7200, '_checkpoint_save_graph_def': True, '_service': None, '_cluster_spec': ClusterSpec({}), '_task_type': 'worker', '_task_id': 0, '_global_id_in_cluster': 0, '_master': '', '_evaluation_master': '', '_is_chief': True, '_num_ps_replicas': 0, '_num_worker_replicas': 1}
WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.9/site-packages/tensorflow/python/training/training_util.py:396: Variable.initialized_value (from tensorflow.python.ops.variables) is deprecated and will be removed in a future version.
Instructions for updating:
Use Variable.read_value. Variables in 2.X are initialized automatically both in eager and graph (inside tf.defun) contexts.
INFO:tensorflow:Calling model_fn.
WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.9/site-packages/tensorflow/python/training/adagrad.py:138: calling Constant.__init__ (from tensorflow.python.ops.init_ops) with dtype is deprecated and will be removed in a future version.
Instructions for updating:
Call initializer instance with the dtype argument instead of passing it to the constructor
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Create CheckpointSaverHook.
INFO:tensorflow:Graph was finalized.
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
INFO:tensorflow:Calling checkpoint listeners before saving checkpoint 0...
INFO:tensorflow:Saving checkpoints for 0 into /tmpfs/tmp/tmp9nj_75jg/model.ckpt.
INFO:tensorflow:Calling checkpoint listeners after saving checkpoint 0...
INFO:tensorflow:loss = 12.718544, step = 0
INFO:tensorflow:Tensor("Identity:0", shape=(2, 1), dtype=float32) = [[-1.1990657]
 [-1.3781608]], Tensor("Identity_1:0", shape=(1,), dtype=float32) = [0.]
INFO:tensorflow:loss from LoggingTensorHook = 12.718544
INFO:tensorflow:Tensor("Identity:0", shape=(2, 1), dtype=float32) = [[-1.1491147]
 [-1.3281827]], Tensor("Identity_1:0", shape=(1,), dtype=float32) = [0.04995093] (0.027 sec)
INFO:tensorflow:Calling checkpoint listeners before saving checkpoint 2...
INFO:tensorflow:Saving checkpoints for 2 into /tmpfs/tmp/tmp9nj_75jg/model.ckpt.
INFO:tensorflow:Calling checkpoint listeners after saving checkpoint 2...
INFO:tensorflow:Loss for final step: 36.829548.
<tensorflow_estimator.python.estimator.estimator.Estimator at 0x7f226c14eeb0>

TensorFlow 2：使用自定义回调和 Model.fit 记录张量和停止训练

在 TensorFlow 2 中，当您使用内置 Keras Model.fit（或 Model.evaluate）进行训练/评估时，可以通过定义自定义 Keras tf.keras.callbacks.Callback 来配置张量监视和训练停止。随后，将它们传递给 Model.fit（或 Model.evaluate）的 callbacks 参数。（在编写自己的回调指南中了解详情。）

在下面的示例中：

要重新创建 StopAtStepHook 的功能，请定义一个自定义回调（下称 StopAtStepCallback），可以在其中重写 on_batch_end 方法以在一定数量的步骤后停止训练。
要重新创建 LoggingTensorHook 行为，请定义一个自定义回调 (LoggingTensorCallback)，可以在其中手动记录和输出记录的张量，因为不支持按名称访问张量。此外，您还可以在自定义回调中实现记录频率。下面的示例将每两步打印一次权重。每 N 秒记录一次之类的其他策略也是可行的。

class StopAtStepCallback(tf.keras.callbacks.Callback):
  def __init__(self, stop_step=None):
    super().__init__()
    self._stop_step = stop_step

  def on_batch_end(self, batch, logs=None):
    if self.model.optimizer.iterations >= self._stop_step:
      self.model.stop_training = True
      print('\nstop training now')

class LoggingTensorCallback(tf.keras.callbacks.Callback):
  def __init__(self, every_n_iter):
      super().__init__()
      self._every_n_iter = every_n_iter
      self._log_count = every_n_iter

  def on_batch_end(self, batch, logs=None):
    if self._log_count > 0:
      self._log_count -= 1
      print("Logging Tensor Callback: dense/kernel:",
            model.layers[0].weights[0])
      print("Logging Tensor Callback: dense/bias:",
            model.layers[0].weights[1])
      print("Logging Tensor Callback loss:", logs["loss"])
    else:
      self._log_count -= self._every_n_iter

完成后，将新回调（StopAtStepCallback 和 LoggingTensorCallback）传递给 Model.fit 的 callbacks 参数：

dataset = tf.data.Dataset.from_tensor_slices((features, labels)).batch(1)
model = tf.keras.models.Sequential([tf.keras.layers.Dense(1)])
optimizer = tf.keras.optimizers.Adagrad(learning_rate=0.05)
model.compile(optimizer, "mse")

# Begin training.
# The training will stop after 2 steps, and the weights/loss will also be logged.
model.fit(dataset, callbacks=[StopAtStepCallback(stop_step=2),
                              LoggingTensorCallback(every_n_iter=2)])

Logging Tensor Callback: dense/kernel: <tf.Variable 'dense/kernel:0' shape=(2, 1) dtype=float32, numpy=
array([[1.3215477 ],
       [0.41005942]], dtype=float32)>
Logging Tensor Callback: dense/bias: <tf.Variable 'dense/bias:0' shape=(1,) dtype=float32, numpy=array([-0.04979974], dtype=float32)>
Logging Tensor Callback loss: 3.102187156677246
1/3 [=========>....................] - ETA: 0s - loss: 3.1022
stop training now
Logging Tensor Callback: dense/kernel: <tf.Variable 'dense/kernel:0' shape=(2, 1) dtype=float32, numpy=
array([[1.2734439],
       [0.3627134]], dtype=float32)>
Logging Tensor Callback: dense/bias: <tf.Variable 'dense/bias:0' shape=(1,) dtype=float32, numpy=array([-0.09329326], dtype=float32)>
Logging Tensor Callback loss: 6.4134416580200195
3/3 [==============================] - 0s 4ms/step - loss: 6.4134
<keras.callbacks.History at 0x7f2150509220>

后续步骤

通过以下方式详细了解回调：

API 文档：tf.keras.callbacks.Callback
指南：编写自己的回调
指南：使用内置方法进行训练和评估（使用回调部分）

此外，您可能还会发现下列与迁移相关的资源十分有用：

提前停止迁移指南：tf.keras.callbacks.EarlyStopping 是一个内置的提前停止回调
TensorBoard 迁移指南：TensorBoard 支持跟踪和显示指标
使用辅助逻辑进行训练迁移指南：从 TensorFlow 1 中的 SessionRunHook 到 TensorFlow 2 中的 Keras 回调