Generalize reshape layer to higher dimensions (Kotlin#249)

michalharakal · Nov 30, 2021 · ac1c0b4 · ac1c0b4
1 parent daa3eea
commit ac1c0b4
Show file tree

Hide file tree

Showing 2 changed files with 97 additions and 36 deletions.
diff --git a/api/src/main/kotlin/org/jetbrains/kotlinx/dl/api/core/layer/reshaping/Reshape.kt b/api/src/main/kotlin/org/jetbrains/kotlinx/dl/api/core/layer/reshaping/Reshape.kt
@@ -36,47 +36,18 @@ public class Reshape(
     override fun build(tf: Ops, kGraph: KGraph, inputShape: Shape) {
         val tensorShape = TensorShape(inputShape)
         val amountOfNeuronsInFlattenLayer = (tensorShape.numElements() / abs(tensorShape.size(0))).toInt()
-        units = when (targetShape.size) {
-            1 -> tf.constant(intArrayOf(-1, targetShape[0]))
-            2 -> tf.constant(intArrayOf(-1, targetShape[0], targetShape[1]))
-            3 -> tf.constant(intArrayOf(-1, targetShape[0], targetShape[1], targetShape[2]))
-            4 -> tf.constant(intArrayOf(-1, targetShape[0], targetShape[1], targetShape[2], targetShape[3]))
-            5 -> tf.constant(
-                intArrayOf(
-                    -1,
-                    targetShape[0],
-                    targetShape[1],
-                    targetShape[2],
-                    targetShape[3],
-                    targetShape[4]
-                )
-            )
-            else -> throw UnsupportedOperationException("Reshaping for ${targetShape.size} is not supported yet!")
-        }
+
+        units = tf.constant(IntArray(targetShape.size + 1) {
+            if (it == 0) -1 else targetShape[it - 1]
+        })
 
         fanIn = tensorShape.numElements().toInt()
         fanOut = amountOfNeuronsInFlattenLayer
     }
 
-    override fun computeOutputShape(inputShape: Shape): Shape {
-        // leaves unknown dimensions unknown
-        val tensorShape = TensorShape(inputShape)
-        return when (targetShape.size) {
-            3 -> Shape.make(
-                tensorShape.head(),
-                targetShape[0].toLong(),
-                targetShape[1].toLong(),
-                targetShape[2].toLong()
-            )
-            2 -> Shape.make(
-                tensorShape.head(),
-                targetShape[0].toLong(),
-                targetShape[1].toLong(),
-            )
-            1 -> Shape.make(tensorShape.head(), targetShape[0].toLong())
-            else -> throw UnsupportedOperationException("Input shape with ${targetShape.size} dimensions is not supported.")
-        }
-    }
+    // leaves unknown dimensions unknown
+    override fun computeOutputShape(inputShape: Shape): Shape =
+        Shape.make(inputShape.size(0), *targetShape.map { it.toLong() }.toLongArray())
 
     override fun forward(
         tf: Ops,

diff --git a/api/src/test/kotlin/org/jetbrains/kotlinx/dl/api/core/layer/reshaping/ReshapeTest.kt b/api/src/test/kotlin/org/jetbrains/kotlinx/dl/api/core/layer/reshaping/ReshapeTest.kt
@@ -0,0 +1,90 @@
+package org.jetbrains.kotlinx.dl.api.core.layer.reshaping
+
+import org.jetbrains.kotlinx.dl.api.core.layer.LayerTest
+import org.jetbrains.kotlinx.dl.api.core.layer.RunMode
+import org.junit.jupiter.api.Test
+
+internal class ReshapeTest : LayerTest() {
+    @Test
+    fun default() {
+        val layer = Reshape(targetShape = listOf(2, 5))
+
+        val input = arrayOf(
+            FloatArray(10) { it + 10.0f },
+            FloatArray(10) { it + 30.0f }
+        )
+        val expected = arrayOf(
+            arrayOf(
+                FloatArray(5) { it + 10.0f },
+                FloatArray(5) { it + 15.0f }
+            ),
+            arrayOf(
+                FloatArray(5) { it + 30.0f },
+                FloatArray(5) { it + 35.0f }
+            )
+        )
+        assertLayerOutputIsCorrect(layer, input, expected)
+        assertLayerOutputIsCorrect(layer, input, expected, RunMode.GRAPH)
+    }
+
+    @Test
+    fun flatten() {
+        val layer = Reshape(targetShape = listOf(60))
+
+        var i = 0.0f
+        val input = Array(10) { Array(2) { Array(3) { Array(2)  { FloatArray(5) { i++ } } } } }
+
+        i = 0.0f
+        val expected = Array(10) { FloatArray(60) { i++ } }
+
+        assertLayerOutputIsCorrect(layer, input, expected)
+        assertLayerOutputIsCorrect(layer, input, expected, RunMode.GRAPH)
+    }
+
+    @Test
+    fun fromFlatToStructured() {
+        val layer = Reshape(targetShape = listOf(2, 3, 2, 5))
+
+        var i = 0.0f
+        val input = Array(10) { FloatArray(60) { i++ } }
+
+        i = 0.0f
+        val expected = Array(10) { Array(2) { Array(3) { Array(2)  { FloatArray(5) { i++ } } } } }
+
+        assertLayerOutputIsCorrect(layer, input, expected)
+        assertLayerOutputIsCorrect(layer, input, expected, RunMode.GRAPH)
+    }
+
+    @Test
+    fun computeOutputShape() {
+        assertLayerComputedOutputShape(
+            layer = Reshape(targetShape = listOf(1, 2, 3)),
+            inputShapeArray = longArrayOf(100, 3, 2, 1),
+            expectedOutputShape = longArrayOf(100, 1, 2, 3)
+        )
+
+        assertLayerComputedOutputShape(
+            layer = Reshape(targetShape = listOf(1, 2, 3)),
+            inputShapeArray = longArrayOf(-1, 3, 2, 1),
+            expectedOutputShape = longArrayOf(-1, 1, 2, 3)
+        )
+
+        assertLayerComputedOutputShape(
+            layer = Reshape(targetShape = listOf(6)),
+            inputShapeArray = longArrayOf(100, 3, 2, 1),
+            expectedOutputShape = longArrayOf(100, 6)
+        )
+
+        assertLayerComputedOutputShape(
+            layer = Reshape(targetShape = listOf(4, 5)),
+            inputShapeArray = longArrayOf(100, 20),
+            expectedOutputShape = longArrayOf(100, 4, 5)
+        )
+
+        assertLayerComputedOutputShape(
+            layer = Reshape(targetShape = listOf(4, 5, 6, 7, 8, 9, 10)),
+            inputShapeArray = longArrayOf(100, 4 * 5 * 6, 7 * 10, 8 * 9),
+            expectedOutputShape = longArrayOf(100, 4, 5, 6, 7, 8, 9, 10)
+        )
+    }
+}