Keras:使用Tensorboard和train_on_batch()

Question

对于keras函数fit(),fit_generator()通过将keras.callbacks.TensorBoard对象传递给函数,可以进行张量板可视化.对于该train_on_batch()功能,显然没有可用的回调.在这种情况下,keras中还有其他选项来创建Tensorboard吗？

Answer 1

我认为目前唯一的选择是使用 TensorFlow 代码。在这个堆栈溢出答案中，我找到了一种手动创建 TensorBoard 日志的方法。
因此，使用 Keras 的代码示例train_on_batch()可能如下所示：

# before training init writer (for tensorboard log) / model
writer = tf.summary.FileWriter(...)
model = ...

# train model
loss = model.train_on_batch(...)
summary = tf.Summary(value=[tf.Summary.Value(tag="loss", 
                                             simple_value=value), ])
writer.add_summary(summary)

注意：对于 TensorBoard 中的此示例，您必须选择Horizo​​ntal Axis "RELATIVE"，因为没有步骤传递到摘要。

Answer 2

创建TensorBoard回调并手动驱动的一种可能方法：

# This example shows how to use keras TensorBoard callback
# with model.train_on_batch

import tensorflow.keras as keras

# Setup the model
model = keras.models.Sequential()
model.add(...) # Add your layers
model.compile(...) # Compile as usual

batch_size=256

# Create the TensorBoard callback,
# which we will drive manually
tensorboard = keras.callbacks.TensorBoard(
  log_dir='/tmp/my_tf_logs',
  histogram_freq=0,
  batch_size=batch_size,
  write_graph=True,
  write_grads=True
)
tensorboard.set_model(model)

# Transform train_on_batch return value
# to dict expected by on_batch_end callback
def named_logs(model, logs):
  result = {}
  for l in zip(model.metrics_names, logs):
    result[l[0]] = l[1]
  return result

# Run training batches, notify tensorboard at the end of each epoch
for batch_id in range(1000):
  x_train,y_train = create_training_data(batch_size)
  logs = model.train_on_batch(x_train, y_train)
  tensorboard.on_epoch_end(batch_id, named_logs(model, logs))

tensorboard.on_train_end(None)