remove potential performance downgrader (#2651)

* Fix transfer learning (#2645)

* fix transfer learning
* add ParseSingleExample, DecodeBmp tf loader
* add corresponding unit tests

* remove potential performance downgrader

Author: megaSpoon

spark/dl/src/main/scala/com/intel/analytics/bigdl/nn/SpatialBatchNormalization.scala

@@ -319,7 +319,8 @@ object SpatialBatchNormalization {
       val offsetOffset = offset.storageOffset() - 1
       var isIdenticalScale = false
       var identicalScale = 0.0f
-      if (scale.stride().length == 1 && scale.stride()(0) == 0 && scaleData.length == 1) {
+
+      if (scaleData.length == 1) {
         isIdenticalScale = true
         identicalScale = scaleData(0)
       }

spark/dl/src/test/scala/com/intel/analytics/bigdl/nn/tf/DecodeBmpSpec.scala

@@ -27,6 +27,9 @@ import org.tensorflow.example.Example
 class DecodeBmpSerialTest extends ModuleSerializationTest {
   private def getInputs(name: String): Tensor[ByteString] = {
     import com.intel.analytics.bigdl.utils.tf.TFTensorNumeric.NumericByteString
+    /* since the tfrecord file is loaded into byteArrays regardless of the
+      original image type, we can map "bmp" to 0 as well
+    */
     val index = name match {
       case "png" => 0
       case "jpeg" => 1