Support Tensorflow model file read/write #800

yiheng · 2017-04-19T09:47:30Z

What changes were proposed in this pull request?

Support Tensorflow model file read/write

How was this patch tested?

manual test, unit test

yiheng · 2017-05-26T09:48:27Z

Let me break the change into smalls to check in.

i8run · 2017-06-13T02:58:41Z

spark/dl/src/main/scala/com/intel/analytics/bigdl/utils/tf/TensorflowToBigDL.scala

+    require(
+      tfTensor.getDtype == DataType.DT_FLOAT ||
+        tfTensor.getDtype == DataType.DT_FLOAT ||
+        tfTensor.getDtype == DataType.DT_INT32,


Double DT_FLOAT.

i8run · 2017-06-13T03:55:34Z

spark/dl/src/main/scala/com/intel/analytics/bigdl/utils/tf/TensorflowToBigDL.scala

+          tmp(j) = params.get(j)
+          j += 1
+        }
+        Tensor(Storage(tmp), 1, shape).asInstanceOf[Tensor[T]]


It seems a little strange, there are five similar code snippets. Does this have much higher performance?

I don't have an idea to refine this. Can you provide a specific example?

yangw1234 · 2017-06-13T06:44:56Z

spark/dl/src/main/scala/com/intel/analytics/bigdl/utils/tf/TensorflowLoader.scala

+            val posGraph = { if (direction == 0) i else graphNode.prevNodes.length - 1 - j}
+            val pn = patternNode.prevNodes(posPattern)
+            val gn = graphNode.prevNodes(posGraph)
+            if (patternToGraph.keySet.contains(pn)) {


Why not using patternToGraph.contains(pn) directly?

yangw1234 · 2017-06-13T07:00:31Z

spark/dl/src/main/scala/com/intel/analytics/bigdl/utils/tf/TensorflowLoader.scala

+        // Normal operation node
+        if (patternToGraph.get(patternNode).isEmpty) return (util.Collections.emptyList(), Seq())
+
+        val graphNode = patternToGraph.get(patternNode).get


Why not use patternToGraph(patternNode) directly?

yangw1234 · 2017-06-13T07:20:35Z

spark/dl/src/main/scala/com/intel/analytics/bigdl/utils/tf/TensorflowToBigDL.scala

+        }
+        Tensor(Storage(tmp), 1, shape).asInstanceOf[Tensor[T]]
+      } else {
+        throw new IllegalArgumentException("Data type ${tfTensor.getDtype} is not supported now")


lacking the leading "s", s"Data type ${}"

yangw1234 · 2017-06-13T07:36:20Z

spark/dl/src/main/scala/com/intel/analytics/bigdl/utils/tf/TensorflowToBigDL.scala

+
+    val shape = tfTensor.getTensorShape.getDimList.asScala.map(_.getSize.toInt).toArray
+
+    if (shape.product == 1) {


Just a thought. Maybe it is not necessary to treat this case specially? If the tensor is a scalar and shape is an empty array, we can simply change the shape to Array(1), then the following code will be able to handle the scalar.

The following code can not get the only 1 element. Leave a comment.

yangw1234 · 2017-06-13T07:50:24Z

spark/dl/src/main/scala/com/intel/analytics/bigdl/utils/tf/TensorflowToBigDL.scala

+   * Sort the pattern list to make sure the graph match first should not be a sub-graph of the graph
+   * match later
+   */
+  private def sortPattern() : Unit = {


There is a potential issue about this. MulTF and ElementWiseMulTF pattern have the exact same number of nodes and edges, but MulTF should comes before ElementWiseMulTF.

This has been addressed by removing the wildcards then compare the node number. After removed wildcards, MulTF has two nodes and ElementWiseMulTf only has one, so MulTF come first.

verified by exchange the order.

yangw1234 · 2017-06-13T07:54:37Z

spark/dl/src/test/scala/com/intel/analytics/bigdl/utils/tf/TensorflowLoaderSpec.scala

+    val BigDLResult = model.forward(input)
+
+    tfResult.map( BigDLResult.toTensor, (v1, v2) => {
+      assert(abs(v1 - v2) < 1e-7);


Maybe we should use relative error here?

…n in BigDL

* add unit test of lenet backward * add some print * add backward test in lenet and alexnet * seperate testModel into forward and backward methods

jason-dai · 2017-06-28T05:50:30Z

pyspark/bigdl/util/tf_utils.py

+
+import tempfile
+
+import tensorflow as tf


Why import tensorflow? I don't think we want to do that. @yiheng

jason-dai · 2017-06-28T05:51:10Z

pyspark/example/tf_example.py

+# limitations under the License.
+#
+
+import tensorflow as tf


Why add such an example here? It should go to unit test if needed.

jason-dai · 2017-06-28T07:53:36Z

pyspark/bigdl/nn/layer.py

        return Layer.of(jmodel)

+    @staticmethod
+    def load_tensorflow(path, inputs, outputs, byte_order = "little_endian", bigdl_type="float"):


Do you have a test case to cover this function?

jason-dai · 2017-06-29T12:36:11Z

We need to add some examples to show how to load tensorflow model (maybe in the load_model example) and save to tensorflow mode.

In addition, we should add some utilities to convert between tensorflow and bigdl models.

jason-dai · 2017-06-29T14:16:45Z

spark/dl/src/main/scala/com/intel/analytics/bigdl/utils/tf/TensorflowLoader.scala

+        })
+
+        // These two pieces of code are all necessary
+        val nextNodes = n.nextNodes.filter(


Shouldn't we apply this every node in the match subgraph, not just n?

yiheng force-pushed the tfpb branch from ab75349 to 4bbfd96 Compare April 25, 2017 02:07

helenlly assigned i8run and zhichao-li Apr 25, 2017

yiheng force-pushed the tfpb branch 3 times, most recently from 230aad4 to 8e26d65 Compare April 28, 2017 09:28

yiheng force-pushed the tfpb branch from d8ab57b to b763a1b Compare May 5, 2017 07:06

yiheng force-pushed the tfpb branch 2 times, most recently from b8d35c9 to e2d093b Compare May 26, 2017 09:20

yiheng force-pushed the tfpb branch from e2d093b to 34bd374 Compare May 31, 2017 06:38

yiheng force-pushed the tfpb branch 7 times, most recently from 3a2f024 to a873fdc Compare June 12, 2017 16:19

i8run reviewed Jun 13, 2017

View reviewed changes

yangw1234 reviewed Jun 13, 2017

View reviewed changes

yiheng changed the title ~~[WIP] Support Tensorflow model file read/write~~ Support Tensorflow model file read/write Jun 15, 2017

yiheng-wang-intel and others added 8 commits June 20, 2017 16:17

nn refactor

bff76f5

fix code style issue

269681c

change back the layers

27362e8

nn refactor

fd0b347

code refactor

1f5edcf

change tests to automatic test

7df9a62

add more test for model save

0ab6677

remove some useless unit test

bf56fe3

yiheng-wang-intel and others added 13 commits June 20, 2017 16:17

add more save test

db5f385

add more writer test

d481afc

rnn test case automation

bc4fd87

refine save test

190ecf6

refine save unit test

fab5c47

remove NHWC

adcc53f

meet code review

01837b2

meet code review

3f8a858

use MulConst in MulTF

ca105c1

add a flatten node for tf 1.1

50c8d6c

fix code style and failed unit test

44b66b8

mv tf model layers to another package

886f912

add a python example to show how to define model in tensorflow and ru…

11f6ad4

…n in BigDL

yiheng force-pushed the tfpb branch from 6b35a2c to 11f6ad4 Compare June 20, 2017 08:20

yiheng-wang-intel and others added 2 commits June 20, 2017 16:29

move tf example python code to example folder

82bba8d

Add backward test in LeNet and AlexNet (#19)

ce5905e

* add unit test of lenet backward * add some print * add backward test in lenet and alexnet * seperate testModel into forward and backward methods

yiheng merged commit 6cf1f6a into intel:master Jun 21, 2017

jason-dai reviewed Jun 28, 2017

View reviewed changes

jason-dai reviewed Jun 29, 2017

View reviewed changes

yiheng deleted the tfpb branch August 23, 2017 02:23


		val shape = tfTensor.getTensorShape.getDimList.asScala.map(_.getSize.toInt).toArray

		if (shape.product == 1) {

Support Tensorflow model file read/write #800

Support Tensorflow model file read/write #800

Uh oh!

Conversation

yiheng commented Apr 19, 2017

What changes were proposed in this pull request?

How was this patch tested?

Uh oh!

yiheng commented May 26, 2017

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jason-dai commented Jun 29, 2017

Uh oh!

jason-dai Jun 29, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

jason-dai Jun 29, 2017 •

edited

Loading