diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/models/resnet/TrainImageNet.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/models/resnet/TrainImageNet.scala
index b854e3ac729..53ac3765505 100644
--- a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/models/resnet/TrainImageNet.scala
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/models/resnet/TrainImageNet.scala
@@ -26,7 +26,7 @@ import com.intel.analytics.bigdl.dllib.optim._
 import com.intel.analytics.bigdl.dllib.utils._
 import com.intel.analytics.bigdl.dllib.tensor.TensorNumericMath.TensorNumeric._
 import com.intel.analytics.bigdl.utils._
-import com.intel.analytics.bigdl.dllib.utils.visualization.{TrainSummary, ValidationSummary}
+import com.intel.analytics.bigdl.dllib.visualization.{TrainSummary, ValidationSummary}
 import org.apache.log4j.{Level, Logger}
 import org.apache.spark.SparkContext
 
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/models/vgg/Train.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/models/vgg/Train.scala
index 7818e60ed4f..cfddf5879bd 100644
--- a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/models/vgg/Train.scala
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/models/vgg/Train.scala
@@ -27,7 +27,7 @@ import com.intel.analytics.bigdl.dllib.tensor.TensorNumericMath.TensorNumeric._
 import com.intel.analytics.bigdl.dllib.utils.{T, Table}
 import com.intel.analytics.bigdl.utils.{Engine, OptimizerV1, OptimizerV2}
 import com.intel.analytics.bigdl.utils.LoggerFilter
-import com.intel.analytics.bigdl.dllib.utils.visualization.{TrainSummary, ValidationSummary}
+import com.intel.analytics.bigdl.dllib.visualization.{TrainSummary, ValidationSummary}
 import org.apache.log4j.{Level, Logger}
 import org.apache.spark.SparkContext
 
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/models/vgg/TrainImageNet.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/models/vgg/TrainImageNet.scala
index 11162007d22..ec96f72e56a 100644
--- a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/models/vgg/TrainImageNet.scala
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/models/vgg/TrainImageNet.scala
@@ -23,7 +23,7 @@ import com.intel.analytics.bigdl.dllib.optim._
 import com.intel.analytics.bigdl.utils.{Engine, MklBlas, MklDnn, OptimizerV1, OptimizerV2}
 import com.intel.analytics.bigdl.dllib.utils._
 import com.intel.analytics.bigdl.utils.LoggerFilter
-import com.intel.analytics.bigdl.dllib.utils.visualization.TrainSummary
+import com.intel.analytics.bigdl.dllib.visualization.TrainSummary
 import org.apache.log4j.{Level, Logger}
 import org.apache.spark.SparkContext
 
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/nn/Graph.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/nn/Graph.scala
index 7403cd4bcf4..a56b597a447 100644
--- a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/nn/Graph.scala
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/nn/Graph.scala
@@ -28,7 +28,7 @@ import com.intel.analytics.bigdl.dllib.utils._
 import com.intel.analytics.bigdl.dllib.utils.serializer._
 import com.intel.analytics.bigdl.dllib.utils.serializer.converters.DataConverter
 import com.intel.analytics.bigdl.dllib.utils.tf.Tensorflow
-import com.intel.analytics.bigdl.dllib.utils.visualization.tensorboard.{FileWriter => TFFileWriter}
+import com.intel.analytics.bigdl.dllib.visualization.tensorboard.{FileWriter => TFFileWriter}
 
 import scala.collection.mutable
 import scala.collection.mutable.ArrayBuffer
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/AbstractOptimizer.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/AbstractOptimizer.scala
index eb3708b328a..5a48d5dbd92 100644
--- a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/AbstractOptimizer.scala
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/AbstractOptimizer.scala
@@ -26,7 +26,7 @@ import com.intel.analytics.bigdl.dllib.tensor.TensorNumericMath.TensorNumeric
 import com.intel.analytics.bigdl.dllib.utils.intermediate.IRGraph
 import com.intel.analytics.bigdl.utils.{Engine, MklBlas, MklDnn}
 import com.intel.analytics.bigdl.dllib.utils.{Table}
-import com.intel.analytics.bigdl.dllib.utils.visualization.{TrainSummary, ValidationSummary}
+import com.intel.analytics.bigdl.dllib.visualization.{TrainSummary, ValidationSummary}
 import org.apache.spark.rdd.{RDD, ZippedPartitionsWithLocalityRDD}
 
 import scala.reflect.ClassTag
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/DistriOptimizer.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/DistriOptimizer.scala
index 27b1ae79d2f..e18ffde0838 100644
--- a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/DistriOptimizer.scala
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/DistriOptimizer.scala
@@ -26,7 +26,7 @@ import com.intel.analytics.bigdl.dllib.tensor.Tensor
 import com.intel.analytics.bigdl.dllib.tensor.TensorNumericMath.TensorNumeric
 import com.intel.analytics.bigdl.dllib.utils._
 import com.intel.analytics.bigdl.dllib.utils.intermediate.{ConversionUtils, IRGraph}
-import com.intel.analytics.bigdl.dllib.utils.visualization.{TrainSummary, ValidationSummary}
+import com.intel.analytics.bigdl.dllib.visualization.{TrainSummary, ValidationSummary}
 import com.intel.analytics.bigdl.utils._
 import com.intel.analytics.bigdl.{Module, _}
 import java.io.{File, FilenameFilter}
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/DistriOptimizerV2.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/DistriOptimizerV2.scala
index e9a85144af8..428a323f71b 100644
--- a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/DistriOptimizerV2.scala
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/DistriOptimizerV2.scala
@@ -29,7 +29,7 @@ import com.intel.analytics.bigdl.dllib.tensor.TensorNumericMath.TensorNumeric
 import com.intel.analytics.bigdl.utils._
 import com.intel.analytics.bigdl.dllib.utils._
 import com.intel.analytics.bigdl.dllib.utils.intermediate.ConversionUtils
-import com.intel.analytics.bigdl.dllib.utils.visualization.{TrainSummary, ValidationSummary}
+import com.intel.analytics.bigdl.dllib.visualization.{TrainSummary, ValidationSummary}
 import com.intel.analytics.bigdl.{Module, _}
 import org.apache.log4j.Logger
 import org.apache.spark.rdd.RDD
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/Optimizer.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/Optimizer.scala
index 402253963a2..5d61c0fe528 100644
--- a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/Optimizer.scala
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/Optimizer.scala
@@ -26,7 +26,7 @@ import com.intel.analytics.bigdl.dllib.optim.parameters.{ConstantClippingProcess
   L2NormClippingProcessor, ParameterProcessor}
 import com.intel.analytics.bigdl.dllib.tensor.TensorNumericMath.TensorNumeric
 import com.intel.analytics.bigdl.dllib.utils._
-import com.intel.analytics.bigdl.dllib.utils.visualization.{TrainSummary, ValidationSummary}
+import com.intel.analytics.bigdl.dllib.visualization.{TrainSummary, ValidationSummary}
 import com.intel.analytics.bigdl.utils._
 import org.apache.log4j.Logger
 import org.apache.spark.rdd.RDD
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/ParallelOptimizer.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/ParallelOptimizer.scala
index 3695e17de98..4bae6b6b9ed 100644
--- a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/ParallelOptimizer.scala
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/ParallelOptimizer.scala
@@ -27,7 +27,7 @@ import com.intel.analytics.bigdl.dllib.tensor.Tensor
 import com.intel.analytics.bigdl.dllib.tensor.TensorNumericMath.TensorNumeric
 import com.intel.analytics.bigdl.dllib.utils._
 import com.intel.analytics.bigdl.utils._
-import com.intel.analytics.bigdl.dllib.utils.visualization.{TrainSummary, ValidationSummary}
+import com.intel.analytics.bigdl.dllib.visualization.{TrainSummary, ValidationSummary}
 import com.intel.analytics.bigdl.{Module, _}
 import java.io.{File, FilenameFilter}
 import java.text.SimpleDateFormat
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/utils/python/api/PythonBigDL.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/utils/python/api/PythonBigDL.scala
index 9b51ae5276d..05dd3b6a04a 100644
--- a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/utils/python/api/PythonBigDL.scala
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/utils/python/api/PythonBigDL.scala
@@ -28,7 +28,7 @@ import com.intel.analytics.bigdl.dllib.optim.{Optimizer, _}
 import com.intel.analytics.bigdl.dllib.tensor.{DenseType, SparseType, Storage, Tensor}
 import com.intel.analytics.bigdl.dllib.tensor.TensorNumericMath.TensorNumeric
 import com.intel.analytics.bigdl.dllib.utils.{Table, _}
-import com.intel.analytics.bigdl.dllib.utils.visualization.{Summary, TrainSummary, ValidationSummary}
+import com.intel.analytics.bigdl.dllib.visualization.{Summary, TrainSummary, ValidationSummary}
 import com.intel.analytics.bigdl.utils._
 import org.apache.spark.api.java.{JavaRDD, JavaSparkContext}
 import org.apache.spark.rdd.RDD
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/README.md b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/README.md
new file mode 100644
index 00000000000..fbbbd2041e7
--- /dev/null
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/README.md
@@ -0,0 +1,40 @@
+# Visualization via Tensorboard
+
+> This wheel distribution is provided by https://github.com/dmlc/tensorboard
+
+> You can find the wheel repository at https://pypi.python.org/pypi/tensorboard
+
+Please follow the instructions below to install TensorBoard; it has been tested on both Ubuntu and Mac OS. Please refer to the [Know Issues](https://github.com/122689305/BigDL/tree/readme/spark/dl/src/main/scala/com/intel/analytics/bigdl/visualization#known-issues)  section for possible errors.
+
+## Requirement
+
+Python verison: 2.7, 3.4, 3.5, 3.6
+
+Pip version >= 9.0.1
+
+## Installation
+
+### Python 2
+```pip install tensorboard==1.0.0a4```
+### Python 3
+```pip3 install tensorboard==1.0.0a4```
+
+## Known Issues
+
+> #### 1. Issue: No compatible version of tensorboard
+
+Solutions
+*  [Update](https://pip.pypa.io/en/stable/installing/) your pip version to the latest: https://pip.pypa.io/en/stable/installing/
+*  Check whether your python support wide unicode if you use python 2.7 
+```
+python -c 'import sys;print(sys.maxunicode)'
+```
+　　It should output `1114111`
+
+> #### 2. RuntimeError: module compiled against API version 0xa but this version of numpy is 0x9
+
+　　Check your python library path (sys.path) to see whether it includes numpy module
+
+> #### 3. RuntimeError: Cannot load some specific libraries, like '_pywrap_tensorflow.so'. 
+
+　　Set your 'PATH' environment variable so that `$ which python` outputs the path of your python that has installed tensorboard.
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/Summary.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/Summary.scala
new file mode 100644
index 00000000000..81b87a58626
--- /dev/null
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/Summary.scala
@@ -0,0 +1,186 @@
+/*
+ * Copyright 2016 The BigDL Authors.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package com.intel.analytics.bigdl.dllib.visualization
+
+import com.intel.analytics.bigdl.dllib.tensor.Tensor
+import com.intel.analytics.bigdl.dllib.tensor.TensorNumericMath.TensorNumeric
+import com.intel.analytics.bigdl.dllib.visualization.tensorboard.{FileWriter}
+import org.tensorflow
+
+import scala.reflect.ClassTag
+
+/**
+ * Logger for tensorboard.
+ * Support scalar and histogram now.
+ * @param logDir
+ * @param appName
+ */
+abstract class Summary(
+                        logDir: String,
+                        appName: String) {
+  protected val writer: FileWriter
+
+  /**
+   * Add a scalar summary.
+   * @param tag tag name.
+   * @param value tag value.
+   * @param step current step.
+   * @return this
+   */
+  def addScalar(
+                 tag: String,
+                 value: Float,
+                 step: Long): this.type = {
+    writer.addSummary(
+      Summary.scalar(tag, value), step
+    )
+    this
+  }
+
+  /**
+   * Add a histogram summary.
+   * @param tag tag name.
+   * @param value a tensor.
+   * @param step current step.
+   * @return this
+   */
+  def addHistogram[T: ClassTag](
+                                 tag: String,
+                                 value: Tensor[T],
+                                 step: Long)(implicit ev: TensorNumeric[T]): this.type = {
+    writer.addSummary(
+      Summary.histogram[T](tag, value), step
+    )
+    this
+  }
+
+  /**
+   * Read scalar values to an array of triple by tag name.
+   * First element of the triple is step, second is value, third is wallclocktime.
+   * @param tag tag name.
+   * @return an array of triple.
+   */
+  def readScalar(tag: String): Array[(Long, Float, Double)]
+
+  /**
+   * Close this logger.
+   */
+  def close(): Unit = {
+    writer.close()
+  }
+}
+
+object Summary {
+
+  /**
+   * Create a scalar summary.
+   * @param tag tag name
+   * @param scalar scalar value
+   * @return
+   */
+  def scalar(tag: String, scalar : Float): tensorflow.framework.Summary = {
+    val v = tensorflow.framework.Summary.Value.newBuilder().setTag(tag).setSimpleValue(scalar)
+    tensorflow.framework.Summary.newBuilder().addValue(v).build()
+  }
+
+  private val limits = makeHistogramBuckets()
+
+  /**
+   * Create a histogram summary.
+   * @param tag tag name.
+   * @param values values.
+   * @return
+   */
+  def histogram[T: ClassTag](
+      tag: String,
+      values: Tensor[T])(implicit ev: TensorNumeric[T]): tensorflow.framework.Summary = {
+    val counts = new Array[Int](limits.length)
+
+    var squares = 0.0
+    values.apply1{value =>
+      val v = ev.toType[Double](value)
+      squares += v * v
+      val index = bisectLeft(limits, v)
+      counts(index) += 1
+      value
+    }
+
+    val histogram = tensorflow.framework.HistogramProto.newBuilder()
+      .setMin(ev.toType[Double](values.min()))
+      .setMax(ev.toType[Double](values.max()))
+      .setNum(values.nElement())
+      .setSum(ev.toType[Double](values.sum()))
+      .setSumSquares(squares)
+
+    var i = 0
+    while (i < counts.length) {
+      if (counts(i) != 0) {
+        histogram.addBucket(counts(i))
+        histogram.addBucketLimit(limits(i))
+      }
+      i += 1
+    }
+    val v = tensorflow.framework.Summary.Value.newBuilder().setTag(tag).setHisto(histogram)
+    tensorflow.framework.Summary.newBuilder().addValue(v).build()
+  }
+
+  /**
+   * Find a bucket for x.
+   */
+  private def bisectLeft(
+                          a: Array[Double],
+                          x: Double,
+                          lo: Int = 0,
+                          hi: Int = -1): Int = {
+    require(lo >= 0)
+    var high = if (hi == -1) {
+      a.length
+    } else {
+      hi
+    }
+    var low = lo
+
+    while (low < high) {
+      val mid = (low + high) / 2
+      if (a(mid) < x) {
+        low = mid + 1
+      } else {
+        high = mid
+      }
+    }
+    low
+  }
+
+  /**
+   * Create a histogram buckets.
+   * @return
+   */
+  private def makeHistogramBuckets(): Array[Double] = {
+    var v = 1e-12
+    val buckets = new Array[Double](1549)
+    var i = 1
+    buckets(774) = 0.0
+    while (i <= 774) {
+      buckets(774 + i) = v
+      buckets(774 - i) = -v
+      v *= 1.1
+      i += 1
+    }
+    buckets
+  }
+
+}
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/TrainSummary.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/TrainSummary.scala
new file mode 100644
index 00000000000..98ee138d343
--- /dev/null
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/TrainSummary.scala
@@ -0,0 +1,95 @@
+/*
+ * Copyright 2016 The BigDL Authors.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package com.intel.analytics.bigdl.dllib.visualization
+
+import com.intel.analytics.bigdl.dllib.optim.Trigger
+import com.intel.analytics.bigdl.dllib.visualization.tensorboard.{FileReader, FileWriter}
+
+import scala.collection.mutable
+
+/**
+ * Train logger for tensorboard.
+ * Use optimize.setTrainSummary to enable train logger. Then the log will be saved to
+ * logDir/appName/train.
+ *
+ * @param logDir log dir.
+ * @param appName application Name.
+ */
+class TrainSummary(
+                    logDir: String,
+                    appName: String) extends Summary(logDir, appName) {
+  protected val folder = s"$logDir/$appName/train"
+  protected override val writer = new FileWriter(folder)
+  private val triggers: mutable.HashMap[String, Trigger] = mutable.HashMap(
+    "Loss" -> Trigger.severalIteration(1),
+    "Throughput" -> Trigger.severalIteration(1))
+
+  /**
+   * Read scalar values to an array of triple by tag name.
+   * First element of the triple is step, second is value, third is wallClockTime.
+   * @param tag tag name. Supported tag names is "LearningRate", "Loss", "Throughput"
+   * @return an array of triple.
+   */
+  override def readScalar(tag: String): Array[(Long, Float, Double)] = {
+    FileReader.readScalar(folder, tag)
+  }
+
+  /**
+   * Supported tag name are LearningRate, Loss, Throughput, Parameters.
+   * Parameters contains weight, bias, gradWeight, gradBias, and some running status(eg.
+   * runningMean and runningVar in BatchNormalization).
+   *
+   * Notice: By default, we record LearningRate, Loss and Throughput each iteration, while
+   * recording parameters is disabled. The reason is getting parameters from workers is a
+   * heavy operation when the model is very big.
+   *
+   * @param tag tag name
+   * @param trigger trigger
+   * @return
+   */
+  def setSummaryTrigger(tag: String, trigger: Trigger): this.type = {
+    require(tag.equals("LearningRate") || tag.equals("Loss") ||
+      tag.equals("Throughput") | tag.equals("Parameters"),
+      s"TrainSummary: only support LearningRate, Loss, Parameters and Throughput")
+    triggers(tag) = trigger
+    this
+  }
+
+  /**
+   * Get a trigger by tag name.
+   * @param tag
+   * @return
+   */
+  def getSummaryTrigger(tag: String): Option[Trigger] = {
+    if (triggers.contains(tag)) {
+      Some(triggers(tag))
+    } else {
+      None
+    }
+  }
+
+  private[bigdl] def getScalarTriggers(): Iterator[(String, Trigger)] = {
+    triggers.filter(!_._1.equals("Parameters")).toIterator
+  }
+}
+
+object TrainSummary{
+  def apply(logDir: String,
+            appName: String): TrainSummary = {
+    new TrainSummary(logDir, appName)
+  }
+}
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/ValidationSummary.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/ValidationSummary.scala
new file mode 100644
index 00000000000..5a66599abfc
--- /dev/null
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/ValidationSummary.scala
@@ -0,0 +1,51 @@
+/*
+ * Copyright 2016 The BigDL Authors.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package com.intel.analytics.bigdl.dllib.visualization
+
+import com.intel.analytics.bigdl.dllib.visualization.tensorboard.{FileReader, FileWriter}
+
+/**
+ * Validation logger for tensorboard.
+ * Use optimize.setValidation to enable validation logger. Then the log will be saved to
+ * logDir/appName/Validation.
+ *
+ * @param logDir
+ * @param appName
+ */
+class ValidationSummary(
+                         logDir: String,
+                         appName: String) extends Summary(logDir, appName) {
+  protected val folder = s"$logDir/$appName/validation"
+  protected override val writer = new FileWriter(folder)
+
+  /**
+   * ReadScalar by tag name. Optional tag name is based on ValidationMethod, "Loss",
+   * "Top1Accuracy" or "Top5Accuracy".
+   * @param tag tag name.
+   * @return an array of triple.
+   */
+  override def readScalar(tag: String): Array[(Long, Float, Double)] = {
+    FileReader.readScalar(folder, tag)
+  }
+}
+
+object ValidationSummary{
+  def apply(logDir: String,
+            appName: String): ValidationSummary = {
+    new ValidationSummary(logDir, appName)
+  }
+}
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/tensorboard/EventWriter.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/tensorboard/EventWriter.scala
new file mode 100644
index 00000000000..29dd80a0b4b
--- /dev/null
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/tensorboard/EventWriter.scala
@@ -0,0 +1,73 @@
+/*
+ * Copyright 2016 The BigDL Authors.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package com.intel.analytics.bigdl.dllib.visualization.tensorboard
+
+import java.net.InetAddress
+import java.util.concurrent.{LinkedBlockingDeque, TimeUnit}
+
+import org.apache.hadoop.fs.{FileSystem, Path}
+import org.tensorflow.util.Event
+
+/**
+ * Event writer, write event protocol buffers to file.
+ *
+ * @param logDir Support local directory and HDFS directory
+ * @param flushMillis
+ */
+private[bigdl] class EventWriter(logDir: String,
+                                 flushMillis: Int = 1000,
+                                 fs: FileSystem) extends Runnable {
+  private val eventQueue = new LinkedBlockingDeque[Event]()
+  private val outputFile = new Path(logDir +
+    s"/bigdl.tfevents.${(System.currentTimeMillis() / 1e3).toInt}" +
+    s".${InetAddress.getLocalHost().getHostName()}")
+  private val recordWriter = new RecordWriter(outputFile, fs)
+  // Add an empty Event to the queue.
+  eventQueue.add(Event.newBuilder().setWallTime(System.currentTimeMillis() / 1e3).build())
+  @volatile private var running: Boolean = true
+
+  def addEvent(event: Event): this.type = {
+    eventQueue.add(event)
+    this
+  }
+
+  private def flush(): this.type = {
+    while (!eventQueue.isEmpty) {
+      recordWriter.write(eventQueue.pop())
+    }
+    this
+  }
+
+  private def writeEvent(): this.type = {
+    val e = eventQueue.poll(flushMillis, TimeUnit.MILLISECONDS)
+    if (null != e) recordWriter.write(e)
+    this
+  }
+
+  def close(): this.type = {
+    running = false
+    this
+  }
+
+  override def run(): Unit = {
+    while (running) {
+      writeEvent()
+    }
+    flush()
+    recordWriter.close()
+  }
+}
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/tensorboard/FileReader.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/tensorboard/FileReader.scala
new file mode 100644
index 00000000000..744abf37579
--- /dev/null
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/tensorboard/FileReader.scala
@@ -0,0 +1,122 @@
+/*
+ * Copyright 2016 The BigDL Authors.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package com.intel.analytics.bigdl.dllib.visualization.tensorboard
+
+import java.io.{BufferedInputStream}
+import java.nio.ByteBuffer
+
+import org.apache.hadoop.conf.Configuration
+import org.apache.hadoop.fs.{FileSystem, Path}
+import org.tensorflow.util.Event
+
+import scala.collection.mutable.ArrayBuffer
+import scala.util.matching.Regex
+
+private[bigdl] object FileReader {
+  val fileNameRegex = """bigdl.tfevents.*""".r
+
+  /**
+   * Search file with regex.
+   * @param f
+   * @param r
+   * @return
+   */
+  private def recursiveListFiles(f: Path, r: Regex, fs: FileSystem): Array[Path] = {
+    val buffer = new ArrayBuffer[Path]()
+    val files = fs.listFiles(f, true)
+    while (files.hasNext) {
+      val file = files.next().getPath
+      if (r.findFirstIn(file.getName).isDefined) {
+        buffer.append(file)
+      }
+    }
+    buffer.toArray
+  }
+
+  /**
+   * List all events file in path.
+   * @param path should be a local/HDFS folder.
+   * @return
+   */
+  def listFiles(path: String): Array[Path] = {
+    val logPath = new Path(path)
+    val fs = logPath.getFileSystem(new Configuration(false))
+    require(fs.isDirectory(logPath), s"FileReader: $path should be a directory")
+    FileReader.recursiveListFiles(logPath, fileNameRegex, fs)
+  }
+
+  /**
+   * List all folders contains event files in path.
+   * @param path should be a local/HDFS folder.
+   * @return
+   */
+  def list(path: String): Array[String] = {
+    val logPath = new Path(path)
+    val fs = logPath.getFileSystem(new Configuration(false))
+    require(fs.isDirectory(logPath), s"FileReader: $path should be a directory")
+    FileReader.recursiveListFiles(logPath, fileNameRegex, fs).map(_.getParent.toString).distinct
+  }
+
+  /**
+   * Read all scalar events named tag from a path.
+   * @param path should be a local/HDFS folder.
+   * @param tag tag name.
+   * @return
+   */
+  def readScalar(path: String, tag: String): Array[(Long, Float, Double)] = {
+    val logPath = new Path(path)
+    val fs = logPath.getFileSystem(new Configuration(false))
+    require(fs.isDirectory(logPath), s"FileReader: $path should be a directory")
+    val files = FileReader.recursiveListFiles(logPath, fileNameRegex, fs)
+    files.map{file =>
+      readScalar(file, tag, fs)
+    }.flatMap(_.toIterator).sortWith(_._1 < _._1)
+  }
+
+  /**
+   * Read all scalar events named tag from a file.
+   * @param file The path of file. Support local/HDFS path.
+   * @param tag tag name.
+   * @return
+   */
+  def readScalar(file: Path, tag: String, fs: FileSystem): Array[(Long, Float, Double)] = {
+    require(fs.isFile(file), s"FileReader: ${file} should be a file")
+    val bis = new BufferedInputStream(fs.open(file))
+    val longBuffer = new Array[Byte](8)
+    val crcBuffer = new Array[Byte](4)
+    val bf = new ArrayBuffer[(Long, Float, Double)]
+    while (bis.read(longBuffer) > 0) {
+      val l = ByteBuffer.wrap(longBuffer.reverse).getLong()
+      bis.read(crcBuffer)
+      // TODO: checksum
+      //      val crc1 = ByteBuffer.wrap(crcBuffer.reverse).getInt()
+      val eventBuffer = new Array[Byte](l.toInt)
+      bis.read(eventBuffer)
+      val e = Event.parseFrom(eventBuffer)
+      if (e.getSummary.getValueCount == 1 &&
+        tag.equals(e.getSummary.getValue(0).getTag())) {
+        bf.append((e.getStep, e.getSummary.getValue(0).getSimpleValue,
+          e.getWallTime))
+      }
+      bis.read(crcBuffer)
+      //      val crc2 = ByteBuffer.wrap(crcBuffer.reverse).getInt()
+    }
+    bis.close()
+    bf.toArray.sortWith(_._1 < _._1)
+  }
+}
+
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/tensorboard/FileWriter.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/tensorboard/FileWriter.scala
new file mode 100644
index 00000000000..80ef4eba4ca
--- /dev/null
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/tensorboard/FileWriter.scala
@@ -0,0 +1,81 @@
+/*
+ * Copyright 2016 The BigDL Authors.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package com.intel.analytics.bigdl.dllib.visualization.tensorboard
+
+import com.intel.analytics.bigdl.utils.Engine
+import org.apache.hadoop.conf.Configuration
+import org.apache.hadoop.fs.Path
+import org.tensorflow
+import org.tensorflow.framework.GraphDef
+import org.tensorflow.util.Event
+
+/**
+ * Writes Summary protocol buffers to event files.
+ * @param logDirectory Support local directory and HDFS directory
+ * @param flushMillis Interval to flush events queue.
+ */
+private[bigdl] class FileWriter(val logDirectory : String, flushMillis: Int = 1000) {
+  private val logPath = new Path(logDirectory)
+  // write to local disk by default
+  private val fs = logPath.getFileSystem(new Configuration(false))
+
+  require(!fs.exists(logPath) || fs.isDirectory(logPath), s"FileWriter: can not create $logPath")
+  if (!fs.exists(logPath)) fs.mkdirs(logPath)
+
+  private val eventWriter = new EventWriter(logDirectory, flushMillis, fs)
+  Engine.default.invoke(() => eventWriter.run())
+
+  /**
+   * Adds a Summary protocol buffer to the event file.
+   * @param summary a Summary protobuf String generated by bigdl.utils.Summary's
+   *                scalar()/histogram().
+   * @param globalStep a consistent global count of the event.
+   * @return
+   */
+  def addSummary(summary: tensorflow.framework.Summary, globalStep: Long): this.type = {
+    val event = Event.newBuilder().setSummary(summary).build()
+    addEvent(event, globalStep)
+    this
+  }
+
+  def addGraphDef(graph: GraphDef): this.type = {
+    val event = Event.newBuilder().setGraphDef(graph.toByteString).build()
+    eventWriter.addEvent(event)
+    this
+  }
+
+  /**
+   * Add a event protocol buffer to the event file.
+   * @param event A event protobuf contains summary protobuf.
+   * @param globalStep a consistent global count of the event.
+   * @return
+   */
+  def addEvent(event: Event, globalStep: Long): this.type = {
+    eventWriter.addEvent(
+      event.toBuilder.setWallTime(System.currentTimeMillis() / 1e3).setStep(globalStep).build())
+    this
+  }
+
+  /**
+   * Close file writer.
+   * @return
+   */
+  def close(): Unit = {
+    eventWriter.close()
+    fs.close()
+  }
+}
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/tensorboard/RecordWriter.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/tensorboard/RecordWriter.scala
new file mode 100644
index 00000000000..067e41c18ea
--- /dev/null
+++ b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/visualization/tensorboard/RecordWriter.scala
@@ -0,0 +1,57 @@
+/*
+ * Copyright 2016 The BigDL Authors.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package com.intel.analytics.bigdl.dllib.visualization.tensorboard
+
+import java.io.{File, FileOutputStream}
+
+import com.google.common.primitives.{Ints, Longs}
+import com.intel.analytics.bigdl.dllib.utils.Crc32
+import netty.Crc32c
+import org.apache.hadoop.fs.{FSDataOutputStream, FileSystem, Path}
+import org.tensorflow.util.Event
+
+/**
+ * A writer to write event protobuf to file by tensorboard's format.
+ * @param file Support local path and HDFS path
+ */
+private[bigdl] class RecordWriter(file: Path, fs: FileSystem) {
+  val outputStream = if (file.toString.startsWith("hdfs://")) {
+    // FSDataOutputStream couldn't flush data to localFileSystem in time. So reading summaries
+    // will throw exception.
+    fs.create(file, true, 1024)
+  } else {
+    // Using FileOutputStream when write to local.
+    new FileOutputStream(new File(file.toString))
+  }
+  val crc32 = new Crc32c()
+  def write(event: Event): Unit = {
+    val eventString = event.toByteArray
+    val header = Longs.toByteArray(eventString.length.toLong).reverse
+    outputStream.write(header)
+    outputStream.write(Ints.toByteArray(Crc32.maskedCRC32(crc32, header).toInt).reverse)
+    outputStream.write(eventString)
+    outputStream.write(Ints.toByteArray(Crc32.maskedCRC32(crc32, eventString).toInt).reverse)
+    if (outputStream.isInstanceOf[FSDataOutputStream]) {
+      // Flush data to HDFS.
+      outputStream.asInstanceOf[FSDataOutputStream].hflush()
+    }
+  }
+
+  def close(): Unit = {
+    outputStream.close()
+  }
+}
diff --git a/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/integration/HdfsSpec.scala b/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/integration/HdfsSpec.scala
index fc0ffc10e1a..a2b88ebd5c4 100644
--- a/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/integration/HdfsSpec.scala
+++ b/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/integration/HdfsSpec.scala
@@ -31,8 +31,8 @@ import com.intel.analytics.bigdl.dllib.tensor.TensorNumericMath.TensorNumeric.Nu
 import com.intel.analytics.bigdl.dllib.utils.caffe.{CaffeLoader, CaffePersister}
 import com.intel.analytics.bigdl.dllib.utils.tf._
 import com.intel.analytics.bigdl.dllib.utils.{File}
-import com.intel.analytics.bigdl.dllib.utils.visualization.Summary
-import com.intel.analytics.bigdl.dllib.utils.visualization.tensorboard.{FileReader, FileWriter}
+import com.intel.analytics.bigdl.dllib.visualization.Summary
+import com.intel.analytics.bigdl.dllib.visualization.tensorboard.{FileReader, FileWriter}
 import com.intel.analytics.bigdl.utils._
 import org.apache.commons.compress.utils.IOUtils
 import org.apache.hadoop.conf.Configuration
diff --git a/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/optim/DistriOptimizerSpec.scala b/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/optim/DistriOptimizerSpec.scala
index 05c16b0541c..496b61458d3 100644
--- a/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/optim/DistriOptimizerSpec.scala
+++ b/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/optim/DistriOptimizerSpec.scala
@@ -30,7 +30,7 @@ import com.intel.analytics.bigdl.dllib.nn.mkldnn.Phase.TrainingPhase
 import com.intel.analytics.bigdl.dllib.optim.parameters.AllReduceParameter
 import com.intel.analytics.bigdl.dllib.tensor.{DenseTensor, DnnStorage, Storage, Tensor}
 import com.intel.analytics.bigdl.dllib.utils._
-import com.intel.analytics.bigdl.dllib.utils.visualization.TrainSummary
+import com.intel.analytics.bigdl.dllib.visualization.TrainSummary
 import com.intel.analytics.bigdl.utils._
 import org.apache.log4j.{Level, Logger}
 import org.apache.spark.SparkContext
diff --git a/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/optim/DistriOptimizerV2Spec.scala b/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/optim/DistriOptimizerV2Spec.scala
index 0f33fa707d6..2789fc00b7b 100644
--- a/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/optim/DistriOptimizerV2Spec.scala
+++ b/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/optim/DistriOptimizerV2Spec.scala
@@ -30,7 +30,7 @@ import com.intel.analytics.bigdl.dllib.nn.mkldnn.Phase.TrainingPhase
 import com.intel.analytics.bigdl.dllib.optim.parameters.AllReduceParameter
 import com.intel.analytics.bigdl.dllib.tensor.{DenseTensor, DnnStorage, Storage, Tensor}
 import com.intel.analytics.bigdl.dllib.utils._
-import com.intel.analytics.bigdl.dllib.utils.visualization.TrainSummary
+import com.intel.analytics.bigdl.dllib.visualization.TrainSummary
 import com.intel.analytics.bigdl.utils._
 import org.apache.log4j.{Level, Logger}
 import org.apache.spark.SparkContext
diff --git a/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/optim/LocalOptimizerSpec.scala b/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/optim/LocalOptimizerSpec.scala
index 6fc9a72d23f..743cdbab199 100644
--- a/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/optim/LocalOptimizerSpec.scala
+++ b/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/optim/LocalOptimizerSpec.scala
@@ -26,7 +26,7 @@ import com.intel.analytics.bigdl.dllib.nn.mkldnn.HeapData
 import com.intel.analytics.bigdl.dllib.tensor.{DnnStorage, Storage, Tensor}
 import com.intel.analytics.bigdl.dllib.utils.{T}
 import com.intel.analytics.bigdl.utils._
-import com.intel.analytics.bigdl.dllib.utils.visualization.TrainSummary
+import com.intel.analytics.bigdl.dllib.visualization.TrainSummary
 import org.scalatest.{BeforeAndAfter, FlatSpec, Matchers}
 
 object DummyDataSet extends LocalDataSet[MiniBatch[Float]] {
diff --git a/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/python/api/PythonSpec.scala b/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/python/api/PythonSpec.scala
index b34cbec6b01..0435d9653f4 100644
--- a/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/python/api/PythonSpec.scala
+++ b/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/python/api/PythonSpec.scala
@@ -24,7 +24,7 @@ import com.intel.analytics.bigdl.dllib.feature.dataset.DataSet
 import com.intel.analytics.bigdl.dllib.nn._
 import com.intel.analytics.bigdl.dllib.optim._
 import com.intel.analytics.bigdl.dllib.utils.{T, Table, TestUtils}
-import com.intel.analytics.bigdl.dllib.utils.visualization.{TrainSummary, ValidationSummary}
+import com.intel.analytics.bigdl.dllib.visualization.{TrainSummary, ValidationSummary}
 import com.intel.analytics.bigdl.utils._
 import org.apache.log4j.{Level, Logger}
 import org.apache.spark.{SparkConf, SparkContext}
diff --git a/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/visualization/SummarySpec.scala b/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/visualization/SummarySpec.scala
index 17e34981018..61ea08b0629 100644
--- a/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/visualization/SummarySpec.scala
+++ b/scala/dllib/src/test/scala/com/intel/analytics/bigdl/dllib/visualization/SummarySpec.scala
@@ -14,13 +14,13 @@
  * limitations under the License.
  */
 
-package com.intel.analytics.bigdl.dllib.utils.visualization
+package com.intel.analytics.bigdl.dllib.visualization
 
 import com.intel.analytics.bigdl.dllib.example.loadmodel.AlexNet
 import com.intel.analytics.bigdl.dllib.tensor.Tensor
 import com.intel.analytics.bigdl.dllib.utils.{TestUtils}
 import Summary._
-import com.intel.analytics.bigdl.dllib.utils.visualization.tensorboard.{FileReader, FileWriter}
+import com.intel.analytics.bigdl.dllib.visualization.tensorboard.{FileReader, FileWriter}
 import com.intel.analytics.bigdl.utils._
 import org.scalatest.{BeforeAndAfter, FlatSpec, Matchers}
 import org.tensorflow.framework