add more parameters for text embedding model (opensearch-project#640)

* add more parameters for text embedding model Signed-off-by: Yaliang Wu <[email protected]> * upgrade junit version to 4.13.2 Signed-off-by: Yaliang Wu <[email protected]> * address comments Signed-off-by: Yaliang Wu <[email protected]> Signed-off-by: Yaliang Wu <[email protected]>
ylwu-amzn · Feb 28, 2023 · bef97c8 · bef97c8
1 parent de60f53
commit bef97c8
Show file tree

Hide file tree

Showing 20 changed files with 313 additions and 80 deletions.
diff --git a/.gitignore b/.gitignore
@@ -5,3 +5,4 @@ client/build/
 common/build/
 ml-algorithms/build/
 plugin/build/
+.DS_Store
diff --git a/client/build.gradle b/client/build.gradle
@@ -15,7 +15,7 @@ plugins {
 dependencies {
     implementation project(':opensearch-ml-common')
     compileOnly group: 'org.opensearch', name: 'opensearch', version: "${opensearch_version}"
-    testImplementation group: 'junit', name: 'junit', version: '4.13.1'
+    testImplementation group: 'junit', name: 'junit', version: '4.13.2'
     testImplementation group: 'org.mockito', name: 'mockito-core', version: '4.4.0'
 
 }

diff --git a/common/build.gradle b/common/build.gradle
@@ -13,7 +13,7 @@ plugins {
 dependencies {
     compileOnly group: 'org.opensearch', name: 'opensearch', version: "${opensearch_version}"
     implementation group: 'org.reflections', name: 'reflections', version: '0.9.12'
-    testImplementation group: 'junit', name: 'junit', version: '4.13.1'
+    testImplementation group: 'junit', name: 'junit', version: '4.13.2'
     compileOnly "org.opensearch.client:opensearch-rest-client:${opensearch_version}"
     compileOnly "org.opensearch:common-utils:${common_utils_version}"
     testImplementation group: 'org.mockito', name: 'mockito-core', version: '4.4.0'

diff --git a/common/src/main/java/org/opensearch/ml/common/CommonValue.java b/common/src/main/java/org/opensearch/ml/common/CommonValue.java
@@ -9,6 +9,9 @@
 import static org.opensearch.ml.common.model.MLModelConfig.MODEL_TYPE_FIELD;
 import static org.opensearch.ml.common.model.TextEmbeddingModelConfig.EMBEDDING_DIMENSION_FIELD;
 import static org.opensearch.ml.common.model.TextEmbeddingModelConfig.FRAMEWORK_TYPE_FIELD;
+import static org.opensearch.ml.common.model.TextEmbeddingModelConfig.MODEL_MAX_LENGTH_FIELD;
+import static org.opensearch.ml.common.model.TextEmbeddingModelConfig.NORMALIZE_RESULT_FIELD;
+import static org.opensearch.ml.common.model.TextEmbeddingModelConfig.POOLING_METHOD_FIELD;
 
 public class CommonValue {
 
@@ -87,6 +90,9 @@ public class CommonValue {
             + MODEL_TYPE_FIELD + "\":{\"type\":\"keyword\"},\""
             + EMBEDDING_DIMENSION_FIELD + "\":{\"type\":\"integer\"},\""
             + FRAMEWORK_TYPE_FIELD + "\":{\"type\":\"keyword\"},\""
+            + POOLING_METHOD_FIELD + "\":{\"type\":\"keyword\"},\""
+            + NORMALIZE_RESULT_FIELD + "\":{\"type\":\"boolean\"},\""
+            + MODEL_MAX_LENGTH_FIELD + "\":{\"type\":\"integer\"},\""
             + ALL_CONFIG_FIELD + "\":{\"type\":\"text\"}}},\n"
             + "      \""
             + MLModel.MODEL_CONTENT_HASH_VALUE_FIELD

diff --git a/common/src/main/java/org/opensearch/ml/common/model/TextEmbeddingModelConfig.java b/common/src/main/java/org/opensearch/ml/common/model/TextEmbeddingModelConfig.java
@@ -33,12 +33,19 @@ public class TextEmbeddingModelConfig extends MLModelConfig {
 
     public static final String EMBEDDING_DIMENSION_FIELD = "embedding_dimension";
     public static final String FRAMEWORK_TYPE_FIELD = "framework_type";
+    public static final String POOLING_METHOD_FIELD = "pooling_method";
+    public static final String NORMALIZE_RESULT_FIELD = "normalize_result";
+    public static final String MODEL_MAX_LENGTH_FIELD = "model_max_length";
 
-    private Integer embeddingDimension;
-    private FrameworkType frameworkType;
+    private final Integer embeddingDimension;
+    private final FrameworkType frameworkType;
+    private final PoolingMethod poolingMethod;
+    private final boolean normalizeResult;
+    private final Integer modelMaxLength;
 
     @Builder(toBuilder = true)
-    public TextEmbeddingModelConfig(String modelType, Integer embeddingDimension, FrameworkType frameworkType, String allConfig) {
+    public TextEmbeddingModelConfig(String modelType, Integer embeddingDimension, FrameworkType frameworkType, String allConfig,
+                                    PoolingMethod poolingMethod, boolean normalizeResult, Integer modelMaxLength) {
         super(modelType, allConfig);
         if (embeddingDimension == null) {
             throw new IllegalArgumentException("embedding dimension is null");
@@ -48,13 +55,23 @@ public TextEmbeddingModelConfig(String modelType, Integer embeddingDimension, Fr
         }
         this.embeddingDimension = embeddingDimension;
         this.frameworkType = frameworkType;
+        if (poolingMethod != null) {
+            this.poolingMethod = poolingMethod;
+        } else {
+            this.poolingMethod = PoolingMethod.MEAN;
+        }
+        this.normalizeResult = normalizeResult;
+        this.modelMaxLength = modelMaxLength;
     }
 
     public static TextEmbeddingModelConfig parse(XContentParser parser) throws IOException {
         String modelType = null;
         Integer embeddingDimension = null;
         FrameworkType frameworkType = null;
         String allConfig = null;
+        PoolingMethod poolingMethod = PoolingMethod.MEAN;
+        boolean normalizeResult = false;
+        Integer modelMaxLength = null;
 
         ensureExpectedToken(XContentParser.Token.START_OBJECT, parser.currentToken(), parser);
         while (parser.nextToken() != XContentParser.Token.END_OBJECT) {
@@ -74,12 +91,21 @@ public static TextEmbeddingModelConfig parse(XContentParser parser) throws IOExc
                 case ALL_CONFIG_FIELD:
                     allConfig = parser.text();
                     break;
+                case POOLING_METHOD_FIELD:
+                    poolingMethod = PoolingMethod.from(parser.text().toUpperCase(Locale.ROOT));
+                    break;
+                case NORMALIZE_RESULT_FIELD:
+                    normalizeResult = parser.booleanValue();
+                    break;
+                case MODEL_MAX_LENGTH_FIELD:
+                    modelMaxLength = parser.intValue();
+                    break;
                 default:
                     parser.skipChildren();
                     break;
             }
         }
-        return new TextEmbeddingModelConfig(modelType,  embeddingDimension, frameworkType, allConfig);
+        return new TextEmbeddingModelConfig(modelType,  embeddingDimension, frameworkType, allConfig, poolingMethod, normalizeResult, modelMaxLength);
     }
 
     @Override
@@ -91,13 +117,19 @@ public TextEmbeddingModelConfig(StreamInput in) throws IOException{
         super(in);
         embeddingDimension = in.readInt();
         frameworkType = in.readEnum(FrameworkType.class);
+        poolingMethod = in.readEnum(PoolingMethod.class);
+        normalizeResult = in.readBoolean();
+        modelMaxLength = in.readOptionalInt();
     }
 
     @Override
     public void writeTo(StreamOutput out) throws IOException {
         super.writeTo(out);
         out.writeInt(embeddingDimension);
         out.writeEnum(frameworkType);
+        out.writeEnum(poolingMethod);
+        out.writeBoolean(normalizeResult);
+        out.writeOptionalInt(modelMaxLength);
     }
 
     @Override
@@ -115,13 +147,31 @@ public XContentBuilder toXContent(XContentBuilder builder, Params params) throws
         if (allConfig != null) {
             builder.field(ALL_CONFIG_FIELD, allConfig);
         }
+        if (modelMaxLength != null) {
+            builder.field(MODEL_MAX_LENGTH_FIELD, modelMaxLength);
+        }
+        builder.field(POOLING_METHOD_FIELD, poolingMethod);
+        builder.field(NORMALIZE_RESULT_FIELD, normalizeResult);
         builder.endObject();
         return builder;
     }
 
+    public enum PoolingMethod {
+        MEAN,
+        CLS;
+
+        public static PoolingMethod from(String value) {
+            try {
+                return PoolingMethod.valueOf(value);
+            } catch (Exception e) {
+                throw new IllegalArgumentException("Wrong pooling method");
+            }
+        }
+    }
     public enum FrameworkType {
         HUGGINGFACE_TRANSFORMERS,
-        SENTENCE_TRANSFORMERS;
+        SENTENCE_TRANSFORMERS,
+        HUGGINGFACE_TRANSFORMERS_NEURON;
 
         public static FrameworkType from(String value) {
             try {

diff --git a/common/src/test/java/org/opensearch/ml/common/model/TextEmbeddingModelConfigTests.java b/common/src/test/java/org/opensearch/ml/common/model/TextEmbeddingModelConfigTests.java
@@ -51,7 +51,8 @@ public void toXContent() throws IOException {
         XContentBuilder builder = XContentBuilder.builder(XContentType.JSON.xContent());
         config.toXContent(builder, EMPTY_PARAMS);
         String configContent = TestHelper.xContentBuilderToString(builder);
-        assertEquals("{\"model_type\":\"testModelType\",\"embedding_dimension\":100,\"framework_type\":\"SENTENCE_TRANSFORMERS\",\"all_config\":\"{\\\"field1\\\":\\\"value1\\\",\\\"field2\\\":\\\"value2\\\"}\"}", configContent);
+        System.out.println(configContent);
+        assertEquals("{\"model_type\":\"testModelType\",\"embedding_dimension\":100,\"framework_type\":\"SENTENCE_TRANSFORMERS\",\"all_config\":\"{\\\"field1\\\":\\\"value1\\\",\\\"field2\\\":\\\"value2\\\"}\",\"pooling_method\":\"MEAN\",\"normalize_result\":false}", configContent);
     }
 
     @Test

diff --git a/common/src/test/java/org/opensearch/ml/common/transport/upload/MLUploadInputTest.java b/common/src/test/java/org/opensearch/ml/common/transport/upload/MLUploadInputTest.java
@@ -41,8 +41,8 @@ public class MLUploadInputTest {
     public ExpectedException exceptionRule = ExpectedException.none();
     private final String expectedInputStr = "{\"function_name\":\"LINEAR_REGRESSION\",\"name\":\"modelName\",\"version\":\"version\",\"url\":\"url\",\"model_format\":\"ONNX\"," +
             "\"model_config\":{\"model_type\":\"testModelType\",\"embedding_dimension\":100,\"framework_type\":\"SENTENCE_TRANSFORMERS\"," +
-            "\"all_config\":\"{\\\"field1\\\":\\\"value1\\\",\\\"field2\\\":\\\"value2\\\"}\"}," +
-            "\"load_model\":true,\"model_node_ids\":[\"modelNodeIds\"]}";
+            "\"all_config\":\"{\\\"field1\\\":\\\"value1\\\",\\\"field2\\\":\\\"value2\\\"}\"" +
+            ",\"pooling_method\":\"MEAN\",\"normalize_result\":false},\"load_model\":true,\"model_node_ids\":[\"modelNodeIds\"]}";
     private final FunctionName functionName = FunctionName.LINEAR_REGRESSION;
     private final String modelName = "modelName";
     private final String version = "version";

diff --git a/...test/java/org/opensearch/ml/common/transport/upload_chunk/MLCreateModelMetaInputTest.java b/...test/java/org/opensearch/ml/common/transport/upload_chunk/MLCreateModelMetaInputTest.java
@@ -40,7 +40,8 @@ public class MLCreateModelMetaInputTest {
 
 	@Before
 	public void setup() {
-		config = new TextEmbeddingModelConfig("Model Type", 123, FrameworkType.SENTENCE_TRANSFORMERS, "All Config");
+		config = new TextEmbeddingModelConfig("Model Type", 123, FrameworkType.SENTENCE_TRANSFORMERS, "All Config",
+				TextEmbeddingModelConfig.PoolingMethod.MEAN, true, 512);
 		mLCreateModelMetaInput = new MLCreateModelMetaInput("Model Name", FunctionName.BATCH_RCF, "1.0",
 				"Model Description", MLModelFormat.TORCH_SCRIPT, MLModelState.LOADING, 200L, "123", config, 2);
 	}
@@ -74,8 +75,8 @@ public void testToXContent() throws IOException {
 		XContentBuilder builder = XContentBuilder.builder(XContentType.JSON.xContent());
 		mLCreateModelMetaInput.toXContent(builder, EMPTY_PARAMS);
 		String mlModelContent = TestHelper.xContentBuilderToString(builder);
-		final String expected = "{\"name\":\"Model Name\",\"function_name\":\"BATCH_RCF\",\"version\":\"1.0\",\"description\":\"Model Description\",\"model_format\":\"TORCH_SCRIPT\",\"model_state\":\"LOADING\",\"model_content_size_in_bytes\":200,\"model_content_hash_value\":\"123\",\"model_config\":{\"model_type\":\"Model Type\","
-				+ "\"embedding_dimension\":123,\"framework_type\":\"SENTENCE_TRANSFORMERS\",\"all_config\":\"All Config\"},\"total_chunks\":2}";
+		final String expected = "{\"name\":\"Model Name\",\"function_name\":\"BATCH_RCF\",\"version\":\"1.0\",\"description\":\"Model Description\",\"model_format\":\"TORCH_SCRIPT\",\"model_state\":\"LOADING\",\"model_content_size_in_bytes\":200,\"model_content_hash_value\":\"123\",\"model_config\":{\"model_type\":\"Model Type\"," +
+				"\"embedding_dimension\":123,\"framework_type\":\"SENTENCE_TRANSFORMERS\",\"all_config\":\"All Config\",\"model_max_length\":512,\"pooling_method\":\"MEAN\",\"normalize_result\":true},\"total_chunks\":2}";
 		assertEquals(expected, mlModelContent);
 	}
 }
diff --git a/...st/java/org/opensearch/ml/common/transport/upload_chunk/MLCreateModelMetaRequestTest.java b/...st/java/org/opensearch/ml/common/transport/upload_chunk/MLCreateModelMetaRequestTest.java
@@ -30,7 +30,8 @@ public class MLCreateModelMetaRequestTest {
 
 	@Before
 	public void setUp() {
-		config = new TextEmbeddingModelConfig("Model Type", 123, FrameworkType.SENTENCE_TRANSFORMERS, "All Config");
+		config = new TextEmbeddingModelConfig("Model Type", 123, FrameworkType.SENTENCE_TRANSFORMERS, "All Config",
+				TextEmbeddingModelConfig.PoolingMethod.MEAN, true, 512);
 		mlCreateModelMetaInput = new MLCreateModelMetaInput("Model Name", FunctionName.BATCH_RCF, "1.0",
 				"Model Description", MLModelFormat.TORCH_SCRIPT, MLModelState.LOADING, 200L, "123", config, 2);
 	}

diff --git a/ml-algorithms/build.gradle b/ml-algorithms/build.gradle
@@ -29,7 +29,7 @@ dependencies {
     implementation group: 'io.protostuff', name: 'protostuff-runtime', version: '1.8.0'
     implementation group: 'io.protostuff', name: 'protostuff-api', version: '1.8.0'
     implementation group: 'io.protostuff', name: 'protostuff-collectionschema', version: '1.8.0'
-    testImplementation group: 'junit', name: 'junit', version: '4.13.1'
+    testImplementation group: 'junit', name: 'junit', version: '4.13.2'
     testImplementation group: 'org.mockito', name: 'mockito-core', version: '4.4.0'
     testImplementation group: 'org.mockito', name: 'mockito-inline', version: '4.4.0'
     implementation group: 'com.google.guava', name: 'guava', version: '31.0.1-jre'
@@ -62,7 +62,7 @@ jacocoTestCoverageVerification {
             }
             limit {
                 counter = 'BRANCH'
-                minimum = 0.80 //TODO: increase coverage to 0.85
+                minimum = 0.79 //TODO: increase coverage to 0.85
             }
         }
     }

diff --git a/...rg/opensearch/ml/engine/algorithms/text_embedding/HuggingfaceTextEmbeddingTranslator.java b/...rg/opensearch/ml/engine/algorithms/text_embedding/HuggingfaceTextEmbeddingTranslator.java
@@ -22,6 +22,7 @@
 import ai.djl.translate.Batchifier;
 import ai.djl.translate.Translator;
 import ai.djl.translate.TranslatorContext;
+import org.opensearch.ml.common.model.TextEmbeddingModelConfig;
 
 import java.io.IOException;
 import java.util.Map;
@@ -33,10 +34,18 @@ public class HuggingfaceTextEmbeddingTranslator implements Translator<String, fl
 
     private HuggingFaceTokenizer tokenizer;
     private Batchifier batchifier;
+    private TextEmbeddingModelConfig.PoolingMethod poolingMethod;
+    private boolean normalizeResult;
+    private String modelType;
+    private boolean neuron;
 
-    HuggingfaceTextEmbeddingTranslator(HuggingFaceTokenizer tokenizer, Batchifier batchifier) {
+    HuggingfaceTextEmbeddingTranslator(HuggingFaceTokenizer tokenizer, Batchifier batchifier, TextEmbeddingModelConfig.PoolingMethod poolingMethod, boolean normalizeResult, String modelType, boolean neuron) {
         this.tokenizer = tokenizer;
         this.batchifier = batchifier;
+        this.poolingMethod = poolingMethod;
+        this.normalizeResult = normalizeResult;
+        this.modelType = modelType;
+        this.neuron = neuron;
     }
 
     /** {@inheritDoc} */
@@ -56,13 +65,39 @@ public NDList processInput(TranslatorContext ctx, String input) {
         NDList ndList = new NDList(2);
         ndList.add(manager.create(indices));
         ndList.add(manager.create(attentionMask));
+        if (neuron && ("bert".equalsIgnoreCase(modelType) || "albert".equalsIgnoreCase(modelType))) {
+            long[] tokenTypeIds = encoding.getTypeIds();
+            ndList.add(manager.create(tokenTypeIds));
+        }
         return ndList;
     }
 
     /** {@inheritDoc} */
     @Override
     public float[] processOutput(TranslatorContext ctx, NDList list) {
+        NDArray embeddings = null;
+        switch (this.poolingMethod) {
+            case MEAN:
+                embeddings = meanPooling(ctx, list);
+                break;
+            case CLS:
+                embeddings = list.get(0).get(0);
+                break;
+            default:
+                throw new IllegalArgumentException("Unsupported pooling method");
+        }
+
+        if (normalizeResult) {
+            embeddings = embeddings.normalize(2, 0);
+        }
+        return embeddings.toFloatArray();
+    }
+
+    private static NDArray meanPooling(TranslatorContext ctx, NDList list) {
         NDArray embeddings = list.get("last_hidden_state");
+        if (embeddings == null) {
+            embeddings = list.get(0);
+        }
         Encoding encoding = (Encoding) ctx.getAttachment("encoding");
         long[] attentionMask = encoding.getAttentionMask();
         NDManager manager = ctx.getNDManager();
@@ -73,11 +108,9 @@ public float[] processOutput(TranslatorContext ctx, NDList list) {
         NDArray clamp = inputAttentionMaskSum.clip(1e-9, 1e12);
         NDArray prod = embeddings.mul(inputAttentionMask);
         NDArray sum = prod.sum(AXIS);
-        embeddings = sum.div(clamp).normalize(2, 0);
-
-        return embeddings.toFloatArray();
+        embeddings = sum.div(clamp);
+        return embeddings;
     }
-
     /**
      * Creates a builder to build a {@code TextEmbeddingTranslator}.
      *
@@ -107,6 +140,10 @@ public static final class Builder {
 
         private HuggingFaceTokenizer tokenizer;
         private Batchifier batchifier = Batchifier.STACK;
+        private TextEmbeddingModelConfig.PoolingMethod poolingMethod;
+        private boolean normalizeResult;
+        private String modelType;
+        private boolean neuron;
 
         Builder(HuggingFaceTokenizer tokenizer) {
             this.tokenizer = tokenizer;
@@ -140,7 +177,27 @@ public void configure(Map<String, ?> arguments) {
          * @throws IOException if I/O error occurs
          */
         public HuggingfaceTextEmbeddingTranslator build() throws IOException {
-            return new HuggingfaceTextEmbeddingTranslator(tokenizer, batchifier);
+            return new HuggingfaceTextEmbeddingTranslator(tokenizer, batchifier, poolingMethod, normalizeResult, modelType, neuron);
+        }
+
+        public Builder poolingMethod(TextEmbeddingModelConfig.PoolingMethod poolingMethod) {
+            this.poolingMethod = poolingMethod;
+            return this;
+        }
+
+        public Builder normalizeResult(boolean normalizeResult) {
+            this.normalizeResult = normalizeResult;
+            return this;
+        }
+
+        public Builder modelType(String modelType) {
+            this.modelType = modelType;
+            return this;
+        }
+
+        public Builder neuron(boolean neuron) {
+            this.neuron = neuron;
+            return this;
         }
     }
 }