Add tensorboard support in Speedometer. (#5345)

* Add tensorboard support in Speedometer.

* fix pylint.

* Add tensorboard_callback.

* Refactor.

* fix lint.

python/mxnet/contrib/__init__.py

@@ -6,3 +6,5 @@
 
 from . import symbol as sym
 from . import ndarray as nd
+
+from . import tensorboard

python/mxnet/contrib/tensorboard.py

@@ -0,0 +1,56 @@
+# coding: utf-8
+"""TensorBoard functions that can be used to log various status during epoch."""
+from __future__ import absolute_import
+
+import logging
+
+
+class LogMetricsCallback(object):
+    """Log metrics periodically in TensorBoard.
+    This callback works almost same as `callback.Speedometer`, but write TensorBoard event file
+    for visualization. For more usage, please refer https://github.com/dmlc/tensorboard
+
+    Parameters
+    ----------
+    logging_dir : str
+        TensorBoard event file directory.
+        After that, use `tensorboard --logdir=path/to/logs` to launch TensorBoard visualization.
+    prefix : str
+        Prefix for a metric name of `scalar` value.
+        You might want to use this param to leverage TensorBoard plot feature,
+        where TensorBoard plots different curves in one graph when they have same `name`.
+        The follow example shows the usage(how to compare a train and eval metric in a same graph).
+
+    Examples
+    --------
+    >>> # log train and eval metrics under different directories.
+    >>> training_log = 'logs/train'
+    >>> evaluation_log = 'logs/eval'
+    >>> # in this case, each training and evaluation metric pairs has same name,
+    >>> # you can add a prefix to make it separate.
+    >>> batch_end_callbacks = [mx.tensorboard.LogMetricsCallback(training_log)]
+    >>> eval_end_callbacks = [mx.tensorboard.LogMetricsCallback(evaluation_log)]
+    >>> # run
+    >>> model.fit(train,
+    >>>     ...
+    >>>     batch_end_callback = batch_end_callbacks,
+    >>>     eval_end_callback  = eval_end_callbacks)
+    >>> # Then use `tensorboard --logdir=logs/` to launch TensorBoard visualization.
+    """
+    def __init__(self, logging_dir, prefix=None):
+        self.prefix = prefix
+        try:
+            from tensorboard import SummaryWriter
+            self.summary_writer = SummaryWriter(logging_dir)
+        except ImportError:
+            logging.error('You can install tensorboard via `pip install tensorboard`.')
+
+    def __call__(self, param):
+        """Callback to log training speed and metrics in TensorBoard."""
+        if param.eval_metric is None:
+            return
+        name_value = param.eval_metric.get_name_value()
+        for name, value in name_value:
+            if self.prefix is not None:
+                name = '%s-%s' % (self.prefix, name)
+            self.summary_writer.add_scalar(name, value)