[v0.4.0] Support Uint8List, Bytebuffer in run, bugfixes

CHANGELOG.md

@@ -1,3 +1,8 @@
+## 0.4.0 (Jun 18, 2020)
+* run supports UintList8 and ByteBuffer objects
+* Bug fix, resize input tensor
+* Improved efficiency
+
 ## 0.3.0
 * New features
     * multi-dimensional reshape with type

example/pubspec.lock

@@ -466,7 +466,7 @@ packages:
       path: ".."
       relative: true
     source: path
-    version: "0.3.0"
+    version: "0.4.0"
   typed_data:
     dependency: transitive
     description:

example/test/tflite_flutter_plugin_example_e2e.dart

@@ -81,15 +81,6 @@ void main() {
       interpreter.allocateTensors();
     });
 
-    test('allocate throws if already allocated', () {
-      interpreter.allocateTensors();
-      expect(() => interpreter.allocateTensors(), throwsA(isStateError));
-    });
-
-    test('invoke throws if not allocated', () {
-      expect(() => interpreter.invoke(), throwsA(isStateError));
-    });
-
     test('invoke throws if not allocated after resized', () {
       interpreter.allocateTensors();
       interpreter.resizeInputTensor(0, [1, 2, 4]);
@@ -171,11 +162,6 @@ void main() {
           expect(tensors[0].data, hasLength(4));
         });
 
-        test('set throws if not allocated', () {
-          expect(() => tensors[0].data = Uint8List.fromList(const [0, 0, 0, 0]),
-              throwsA(isStateError));
-        });
-
         test('set', () {
           interpreter.allocateTensors();
           tensors[0].data = Uint8List.fromList(const [0, 0, 0, 0]);

lib/src/interpreter.dart

@@ -24,7 +24,10 @@ class Interpreter {
   int get lastNativeInferenceDurationMicroSeconds =>
       _lastNativeInferenceDurationMicroSeconds;
 
-  Interpreter._(this._interpreter);
+  Interpreter._(this._interpreter) {
+    // Allocate tensors when interpreter is created
+    allocateTensors();
+  }
 
   /// Creates interpreter from model or throws if unsuccessful.
   factory Interpreter._create(Model model, {InterpreterOptions options}) {
@@ -87,7 +90,7 @@ class Interpreter {
 
   /// Updates allocations for all tensors.
   void allocateTensors() {
-    checkState(!_allocated, message: 'Interpreter already allocated.');
+//    checkState(!_allocated, message: 'Interpreter already allocated.');
     checkState(
         tfLiteInterpreterAllocateTensors(_interpreter) == TfLiteStatus.ok);
     _allocated = true;
@@ -115,25 +118,32 @@ class Interpreter {
       throw ArgumentError('Input error: Outputs should not be null or empty.');
     }
 
+    var inputTensors = getInputTensors();
+
+    for (int i = 0; i < inputs.length; i++) {
+      var tensor = inputTensors.elementAt(i);
+      final newShape = tensor.getInputShapeIfDifferent(inputs[i]);
+      if (newShape != null) {
+        resizeInputTensor(i, newShape);
+      }
+    }
+
     if (!_allocated) {
       allocateTensors();
       _allocated = true;
     }
 
-    var inputTensors = getInputTensors();
-    for (var i = 0; i < inputs.length; i++) {
-      if (inputTensors[i].shape != (inputs[i] as List).shape) {
-        resizeInputTensor(i, (inputs[i] as List).shape);
-        allocateTensors();
-        inputTensors = getInputTensors();
-      }
-      inputTensors[i].setTo(inputs[i]);
+    inputTensors = getInputTensors();
+
+    for (int i = 0; i < inputs.length; i++) {
+      inputTensors.elementAt(i).setTo(inputs[i]);
     }
 
     var inferenceStartNanos = DateTime.now().microsecondsSinceEpoch;
     invoke();
     _lastNativeInferenceDurationMicroSeconds =
         DateTime.now().microsecondsSinceEpoch - inferenceStartNanos;
+
     var outputTensors = getOutputTensors();
     for (var i = 0; i < outputTensors.length; i++) {
       outputTensors[i].copyTo(outputs[i]);

lib/src/tensor.dart

@@ -166,6 +166,8 @@ class Tensor {
     var obj;
     if (dst is Uint8List) {
       obj = bytes;
+    } else if (dst is ByteBuffer) {
+      obj = bytes.buffer;
     } else {
       obj = _convertBytesToObject(bytes);
     }
@@ -182,6 +184,9 @@ class Tensor {
     if (o is Uint8List) {
       return o;
     }
+    if (o is ByteBuffer) {
+      return o.asUint8List();
+    }
     var bytes = <int>[];
     if (o is List) {
       for (var e in o) {
@@ -327,6 +332,21 @@ class Tensor {
     }
   }
 
+  List<int> getInputShapeIfDifferent(Object input) {
+    if (input == null) {
+      return null;
+    }
+    if (input is ByteBuffer || input is Uint8List) {
+      return null;
+    }
+
+    final inputShape = computeShapeOf(input);
+    if (inputShape == shape) {
+      return null;
+    }
+    return inputShape;
+  }
+
   @override
   String toString() {
     return 'Tensor{_tensor: $_tensor, name: $name, type: $type, shape: $shape, data:  ${data.length}';

pubspec.yaml

@@ -1,6 +1,6 @@
 name: tflite_flutter
 description: TensorFlow Lite Flutter plugin provides easy, flexible and fast Dart API to integrate TFLite models in flutter apps.
-version: 0.3.0
+version: 0.4.0
 homepage: https://github.com/am15h/tflite_flutter_plugin
 
 environment: