Upgraded to sentis 1.4.0-pre3

DebertaV3.cs

@@ -13,33 +13,52 @@ public sealed class DebertaV3 : MonoBehaviour
     public string hypothesisTemplate = "This example is about {}";
     public string[] classes = { "politics", "economy", "entertainment", "environment" };
 
-    Ops ops;
     IWorker engine;
-    ITensorAllocator allocator;
     string[] vocabularyTokens;
 
     const int padToken = 0;
     const int startToken = 1;
     const int separatorToken = 2;
     const int vocabToTokenOffset = 260;
-    const BackendType backend = BackendType.GPUCompute;
 
     void Start()
     {
-        vocabularyTokens = vocabulary.text.Replace("\r", "").Split("\n");
-
-        allocator = new TensorCachingAllocator();
-        ops = WorkerFactory.CreateOps(backend, allocator);
-
-        Model loadedModel = ModelLoader.Load(model);
-        engine = WorkerFactory.CreateWorker(backend, loadedModel);
-
         if (classes.Length == 0)
         {
             Debug.LogError("There need to be more than 0 classes");
             return;
         }
 
+        vocabularyTokens = vocabulary.text.Replace("\r", "").Split("\n");
+
+        Model baseModel = ModelLoader.Load(model);
+        Model modelWithScoring = Functional.Compile(
+            input =>
+            {
+                // The logits represent the model's predictions for entailment and non-entailment for each example in the batch.
+                // They are of shape [batch size, 2] i.e. with two values per example.
+                // To obtain a single score per example, a softmax function is applied
+                FunctionalTensor logits = baseModel.Forward(input)[0];
+
+                if (multipleTrueClasses || classes.Length == 1)
+                {
+                    // Softmax over the entailment vs. contradiction dimension for each label independently
+                    logits = Functional.Softmax(logits);
+                }
+                else
+                {
+                    // Softmax over all candidate labels
+                    logits = Functional.Softmax(logits, 0);
+                }
+
+                // The scores are stored along the first column
+                return new []{logits[.., 0]};
+            },
+            InputDef.FromModel(baseModel)
+        );
+
+        engine = WorkerFactory.CreateWorker(BackendType.GPUCompute, modelWithScoring);
+
         string[] hypotheses = classes.Select(x => hypothesisTemplate.Replace("{}", x)).ToArray();
         Batch batch = GetTokenizedBatch(text, hypotheses);
         float[] scores = GetBatchScores(batch);
@@ -57,15 +76,15 @@ public sealed class DebertaV3 : MonoBehaviour
 
         Dictionary<string, Tensor> inputs = new()
         {
-            {"input_ids", inputIds},
-            {"attention_mask", attentionMask}
+            {"input_0", inputIds},
+            {"input_1", attentionMask}
         };
 
         engine.Execute(inputs);
-        TensorFloat logits = (TensorFloat)engine.PeekOutput("logits");
-        float[] scores = ScoresFromLogits(logits);
+        TensorFloat scores = (TensorFloat)engine.PeekOutput("output_0");
+        scores.CompleteOperationsAndDownload();
 
-        return scores;
+        return scores.ToReadOnlyArray();
     }
 
     Batch GetTokenizedBatch(string prompt, string[] hypotheses)
@@ -105,39 +124,6 @@ public sealed class DebertaV3 : MonoBehaviour
         return batch;
     }
 
-    float[] ScoresFromLogits(TensorFloat logits)
-    {
-        // The logits represent the model's predictions for entailment and non-entailment for each example in the batch.
-        // They are of shape [batch size, 2], with two values per example.
-        // To obtain a single value (score) per example, a softmax function is applied
-
-        TensorFloat tensorScores;
-        if (multipleTrueClasses || logits.shape.Length(0, 1) == 1)
-        {
-            // Softmax over the entailment vs. contradiction dimension for each label independently
-            tensorScores = ops.Softmax(logits, -1);
-        }
-        else
-        {
-            // Softmax over all candidate labels
-            tensorScores = ops.Softmax(logits, 0);
-        }
-
-        tensorScores.MakeReadable();
-        float[] tensorArray = tensorScores.ToReadOnlyArray();
-
-        tensorScores.Dispose();
-
-        // Select the first column which is the column where the scores are stored
-        float[] scores = new float[tensorArray.Length / 2];
-        for (int i = 0; i < scores.Length; i++)
-        {
-            scores[i] = tensorArray[i * 2];
-        }
-
-        return scores;
-    }
-
     List<int> Tokenize(string input)
     {
         string[] words = input.Split(null);
@@ -164,12 +150,7 @@ public sealed class DebertaV3 : MonoBehaviour
         return ids;
     }
 
-    void OnDestroy()
-    {
-        engine?.Dispose();
-        allocator?.Dispose();
-        ops?.Dispose();
-    }
+    void OnDestroy() => engine?.Dispose();
 
     struct Batch
     {

README.md

@@ -14,7 +14,7 @@ There are more models of different sizes that are compatible made by [MoritzLaur
 
 ## How to Use
 
-* Ensure Sentis version >= 1.3.0-pre.3
+* Ensure Sentis version is 1.4.0-pre.3
 * Create a new scene in Unity 2023
 * Add the DebertaV3.cs file to a GameObject in the scene
 * Assign model and vocabulary

deberta-v3-xsmall-zeroshot-v1.1-all-33.sentis

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c847e53df0918823cd4b5dcb73c9cca343a3b7d9d7d6f528cf685e02b5ed9efe
-size 302793605
+oid sha256:2d1eec4f37a0c222acb32d6ae67945c1099019cac8bd22cdc9ff7b548adcaf9d
+size 301898184