Minor fixes to improve Apple Silicon MPS support (#2873)

* Support 32-bit toTensor() This fixes an issue where calling NDArrayEx.toTensor() fails on Apple Silicon due to a lack of support for float64. * Avoid float64 conversion in Classifications constructor Don't convert probabilities from float32 to float64, because this causes a failure on Apple Silicon. Co-authored-by: Frank Liu <frankfliu2000@gmail.com>
deepjavalibrary · Nov 28, 2023 · 36d4aec · 36d4aec
1 parent ce441c1
commit 36d4aec
Show file tree

Hide file tree

Showing 5 changed files with 70 additions and 14 deletions.
diff --git a/api/src/main/java/ai/djl/modality/Classifications.java b/api/src/main/java/ai/djl/modality/Classifications.java
@@ -88,10 +88,18 @@ public Classifications(List<String> classNames, NDArray probabilities) {
      */
     public Classifications(List<String> classNames, NDArray probabilities, int topK) {
         this.classNames = classNames;
-        NDArray array = probabilities.toType(DataType.FLOAT64, false);
-        this.probabilities =
-                Arrays.stream(array.toDoubleArray()).boxed().collect(Collectors.toList());
-        array.close();
+        if (probabilities.getDataType() == DataType.FLOAT32) {
+            // Avoid converting float32 to float64 as this is not supported on MPS device
+            this.probabilities = new ArrayList<>();
+            for (float prob : probabilities.toFloatArray()) {
+                this.probabilities.add((double) prob);
+            }
+        } else {
+            NDArray array = probabilities.toType(DataType.FLOAT64, false);
+            this.probabilities =
+                    Arrays.stream(array.toDoubleArray()).boxed().collect(Collectors.toList());
+            array.close();
+        }
         this.topK = topK;
     }
 

diff --git a/api/src/main/java/ai/djl/ndarray/internal/NDArrayEx.java b/api/src/main/java/ai/djl/ndarray/internal/NDArrayEx.java
@@ -434,7 +434,12 @@ default NDArray toTensor() {
             if (dim == 3) {
                 result = result.expandDims(0);
             }
-            result = result.div(255.0).transpose(0, 3, 1, 2);
+            // For Apple Silicon MPS it is important not to switch to 64-bit float here
+            if (result.getDataType() == DataType.FLOAT32) {
+                result = result.div(255.0f).transpose(0, 3, 1, 2);
+            } else {
+                result = result.div(255.0).transpose(0, 3, 1, 2);
+            }
             if (dim == 3) {
                 result = result.squeeze(0);
             }

diff --git a/basicdataset/src/main/java/ai/djl/basicdataset/cv/classification/FashionMnist.java b/basicdataset/src/main/java/ai/djl/basicdataset/cv/classification/FashionMnist.java
@@ -30,6 +30,7 @@
 
 import java.io.IOException;
 import java.io.InputStream;
+import java.nio.ByteBuffer;
 import java.util.Map;
 
 /**
@@ -118,8 +119,9 @@ private NDArray readData(Artifact.Item item, long length) throws IOException {
             byte[] buf = Utils.toByteArray(is);
             try (NDArray array =
                     manager.create(
-                            new Shape(length, IMAGE_WIDTH, IMAGE_HEIGHT, 1), DataType.UINT8)) {
-                array.set(buf);
+                            ByteBuffer.wrap(buf),
+                            new Shape(length, IMAGE_WIDTH, IMAGE_HEIGHT, 1),
+                            DataType.UINT8)) {
                 return array.toType(DataType.FLOAT32, false);
             }
         }
@@ -132,8 +134,8 @@ private NDArray readLabel(Artifact.Item item) throws IOException {
             }
 
             byte[] buf = Utils.toByteArray(is);
-            try (NDArray array = manager.create(new Shape(buf.length), DataType.UINT8)) {
-                array.set(buf);
+            try (NDArray array =
+                    manager.create(ByteBuffer.wrap(buf), new Shape(buf.length), DataType.UINT8)) {
                 return array.toType(DataType.FLOAT32, false);
             }
         }

diff --git a/basicdataset/src/main/java/ai/djl/basicdataset/cv/classification/Mnist.java b/basicdataset/src/main/java/ai/djl/basicdataset/cv/classification/Mnist.java
@@ -30,6 +30,7 @@
 
 import java.io.IOException;
 import java.io.InputStream;
+import java.nio.ByteBuffer;
 import java.util.Map;
 
 /**
@@ -111,8 +112,9 @@ private NDArray readData(Artifact.Item item, long length) throws IOException {
             }
 
             byte[] buf = Utils.toByteArray(is);
-            try (NDArray array = manager.create(new Shape(length, 28, 28, 1), DataType.UINT8)) {
-                array.set(buf);
+            try (NDArray array =
+                    manager.create(
+                            ByteBuffer.wrap(buf), new Shape(length, 28, 28, 1), DataType.UINT8)) {
                 return array.toType(DataType.FLOAT32, false);
             }
         }
@@ -123,10 +125,9 @@ private NDArray readLabel(Artifact.Item item) throws IOException {
             if (is.skip(8) != 8) {
                 throw new AssertionError("Failed skip data.");
             }
-
             byte[] buf = Utils.toByteArray(is);
-            try (NDArray array = manager.create(new Shape(buf.length), DataType.UINT8)) {
-                array.set(buf);
+            try (NDArray array =
+                    manager.create(ByteBuffer.wrap(buf), new Shape(buf.length), DataType.UINT8)) {
                 return array.toType(DataType.FLOAT32, false);
             }
         }

diff --git a/engines/pytorch/pytorch-engine/src/test/java/ai/djl/pytorch/integration/MpsTest.java b/engines/pytorch/pytorch-engine/src/test/java/ai/djl/pytorch/integration/MpsTest.java
@@ -13,6 +13,7 @@
 package ai.djl.pytorch.integration;
 
 import ai.djl.Device;
+import ai.djl.modality.Classifications;
 import ai.djl.ndarray.NDArray;
 import ai.djl.ndarray.NDManager;
 import ai.djl.ndarray.types.Shape;
@@ -21,6 +22,10 @@
 import org.testng.SkipException;
 import org.testng.annotations.Test;
 
+import java.util.Arrays;
+import java.util.Collections;
+import java.util.List;
+
 public class MpsTest {
 
     @Test
@@ -36,4 +41,39 @@ public void testMps() {
             Assert.assertEquals(array.getDevice().getDeviceType(), "mps");
         }
     }
+
+    private static boolean checkMpsCompatible() {
+        return "aarch64".equals(System.getProperty("os.arch"))
+                && System.getProperty("os.name").startsWith("Mac");
+    }
+
+    @Test
+    public void testToTensorMPS() {
+        if (!checkMpsCompatible()) {
+            throw new SkipException("MPS toTensor test requires Apple Silicon macOS.");
+        }
+
+        // Test that toTensor does not fail on MPS (e.g. due to use of float64 for division)
+        try (NDManager manager = NDManager.newBaseManager(Device.fromName("mps"))) {
+            NDArray array = manager.create(127f).reshape(1, 1, 1, 1);
+            NDArray tensor = array.getNDArrayInternal().toTensor();
+            Assert.assertEquals(tensor.toFloatArray(), new float[] {127f / 255f});
+        }
+    }
+
+    @Test
+    public void testClassificationsMPS() {
+        if (!checkMpsCompatible()) {
+            throw new SkipException("MPS classification test requires Apple Silicon macOS.");
+        }
+
+        // Test that classifications do not fail on MPS (e.g. due to conversion of probabilities to
+        // float64)
+        try (NDManager manager = NDManager.newBaseManager(Device.fromName("mps"))) {
+            List<String> names = Arrays.asList("First", "Second", "Third", "Fourth", "Fifth");
+            NDArray tensor = manager.create(new float[] {0f, 0.125f, 1f, 0.5f, 0.25f});
+            Classifications classifications = new Classifications(names, tensor);
+            Assert.assertEquals(classifications.topK(1), Collections.singletonList("Third"));
+        }
+    }
 }