使用 tf.keras.Model.fit 进行训练时如何将自定义摘要添加到 tensorboard

Posted 2023-02-23

【中文标题】使用 tf.keras.Model.fit 进行训练时如何将自定义摘要添加到 tensorboard

【英文标题】：How to add custom summaries to tensorboard when training with tf.keras.Model.fit

【发布时间】：2019-04-13 15:43:21

【问题描述】：

我正在训练一个模型：

with tf.Graph().as_default():
        with tf.Session(config=tf.ConfigProto(allow_soft_placement = True)) as sess:
                K.set_session(sess)
                tf.train.create_global_step()
                #with tf.device('/gpu:0:'):
                m = GAReader.Model(nlayers, data.vocab_size, data.num_chars, W_init,
                        nhidden, embed_dim, dropout, train_emb,
                        char_dim, use_feat, gating_fn, words).build_network()
                m.compile(optimizer=tf.train.AdamOptimizer(0.01),
                          loss=tf.keras.losses.categorical_crossentropy,
                          metrics=[tf.keras.metrics.categorical_accuracy])
                tensorboard = TensorBoardCustom(log_dir="logs", sess=sess)
                m.fit_generator(generator=batch_loader_train, steps_per_epoch=len(batch_loader_train.batch_pool), epochs=100, callbacks=[tensorboard])

我定义了一个自定义回调扩展 keras.callbacks.Tensorboard 为：

class TensorBoardCustom(TensorBoard):

    def __init__(self, log_dir, sess, **kwargs):
        super(TensorBoardCustom, self).__init__(log_dir, **kwargs)
        self.sess = sess

    def on_batch_end(self, batch, logs=):
        summary = tf.summary.merge_all()
        writer = tf.summary.FileWriter(self.log_dir)
        s = self.sess.run(summary)
        writer.add_summary(s, batch)
        writer.close()
        super(TensorBoardCustom, self).on_batch_end(batch, logs)

我正在添加一个新的摘要：

l_docin = tf.keras.layers.Input(shape=(None,))
with tf.name_scope('summaries'):
            table = tf.contrib.lookup.index_to_string_table_from_tensor(
                    self.mapping_string, default_value="UNKNOWN")
            words = table.lookup(tf.cast(l_qin, tf.int64))
            text = tf.reduce_join(words, 1, separator=' ')
            tf.summary.text('text', text)

但是，这不起作用，我收到以下错误：

InvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'input_2' with dtype float and shape [?,?]
     [[node input_2 = Placeholder[dtype=DT_FLOAT, shape=[?,?], _device="/job:localhost/replica:0/task:0/device:CPU:0"]()]]

有人可以解释为什么会发生这种情况以及我该如何纠正它吗？有没有更简单/更好的添加自定义摘要的方法？

【参考方案1】：

TensorFlow 回调TensorBoardWithTime 定义如下，记录累积训练和评估批处理时间。 它依赖于私有属性TensorBoard._train_writer 和TensorBoard._val_writer。 它适用于 TensorFlow 2.4.0rc2。

import tensorflow as tf


class TensorBoardWithTime(tf.keras.callbacks.TensorBoard):
    def __init__(self, **kwargs):
        super().__init__(**kwargs)
        self.timer_test_batch = Timer()
        self.timer_train_batch = Timer()

    def on_epoch_end(self, epoch, logs=None):
        super().on_epoch_end(epoch, logs=logs)
        with self._train_writer.as_default():
            tf.summary.scalar('time_batch_sum', self.timer_train_batch.pop_result(), step=epoch)
        with self._val_writer.as_default():
            tf.summary.scalar('time_batch_sum', self.timer_test_batch.pop_result(), step=epoch)

    def on_test_batch_begin(self, batch, logs=None):
        self.timer_test_batch.begin()
        super().on_test_batch_begin(batch, logs=logs)

    def on_test_batch_end(self, batch, logs=None):
        super().on_test_batch_end(batch, logs=logs)
        self.timer_test_batch.end()

    def on_train_batch_begin(self, batch, logs=None):
        self.timer_train_batch.begin()
        super().on_train_batch_begin(batch, logs=logs)

    def on_train_batch_end(self, batch, logs=None):
        super().on_train_batch_end(batch, logs=logs)
        self.timer_train_batch.end()


class Timer(tf.keras.metrics.Sum):
    def __init__(self):
        super().__init__()
        self.time_begin = None

    def pop_result(self):
        result = self.result()
        self.reset_states()
        return result

    def begin(self):
        self.time_begin = tf.timestamp()

    def end(self):
        self.update_state(tf.timestamp() - self.time_begin)
        self.time_begin = None