[api] Allows to configure custom batchifer

api/src/main/java/ai/djl/translate/Batchifier.java

@@ -13,6 +13,7 @@
 package ai.djl.translate;
 
 import ai.djl.ndarray.NDList;
+import ai.djl.util.ClassLoaderUtils;
 
 import java.io.Serializable;
 import java.util.Arrays;
@@ -47,7 +48,12 @@ static Batchifier fromString(String name) {
             case "none":
                 return null;
             default:
-                throw new IllegalArgumentException("Invalid batchifier name");
+                ClassLoader cl = ClassLoaderUtils.getContextClassLoader();
+                Batchifier b = ClassLoaderUtils.initClass(cl, Batchifier.class, name);
+                if (b == null) {
+                    throw new IllegalArgumentException("Invalid batchifier name: " + name);
+                }
+                return b;
         }
     }
 

api/src/test/java/ai/djl/translate/BatchifierTest.java

@@ -22,6 +22,17 @@ public void testBatchifier() {
         Batchifier batchifier = Batchifier.fromString("stack");
         Assert.assertEquals(batchifier, Batchifier.STACK);
 
+        batchifier = Batchifier.fromString("none");
+        Assert.assertNull(batchifier);
+
+        batchifier = Batchifier.fromString("padding");
+        Assert.assertNotNull(batchifier);
+        Assert.assertEquals(batchifier.getClass(), SimplePaddingStackBatchifier.class);
+
+        batchifier = Batchifier.fromString("ai.djl.translate.SimplePaddingStackBatchifier");
+        Assert.assertNotNull(batchifier);
+        Assert.assertEquals(batchifier.getClass(), SimplePaddingStackBatchifier.class);
+
         Assert.assertThrows(() -> Batchifier.fromString("invalid"));
     }
 }

basicdataset/src/test/java/ai/djl/basicdataset/MnistTest.java

@@ -157,7 +157,7 @@ public void testBulkEqualsNonBulk() throws IOException, TranslateException {
                             .optUsage(Dataset.Usage.TEST)
                             .optRepository(repository)
                             .setSampling(32, false)
-                            .optLabelBatchifier(new StackBatchifier() {})
+                            .optLabelBatchifier(new StackBatchifier())
                             .build();
 
             try (Trainer trainer = model.newTrainer(config)) {