Using a custom SSD model in tf-lite-unity-sample

I've been trying to use my own custom SSD model in the SSD sample of this repo: https://github.com/asus4/tf-lite-unity-sample, however I've been running into issues getting it to work.

Under Assets/Samples/SSD/SsdSample.cs

Code (CSharp):

1. public class SsdSample : MonoBehaviour
2. {
3.     [SerializeField, FilePopup("*.tflite")] string fileName = "coco_ssd_mobilenet_quant.tflite";
4.     [SerializeField] RawImage cameraView = null;
5.     [SerializeField] Text framePrefab = null;
6.     [SerializeField, Range(0f, 1f)] float scoreThreshold = 0.5f;
7.     [SerializeField] TextAsset labelMap = null;

Simply changing the fileName to my own model's name (my model is in the same folder as the default model), does not work as removing the default model results in an error.

Under Issue #121, the owner mentioned changing the BaseImagePredictor in Under Assets/Samples/SSD/SSD.cs from sbyte to float if your model is quantized differently. I checked my model using netron, and it uses float instead of byte, so I changed it.

Code (CSharp):

1.     public class SSD : BaseImagePredictor<float>
2.     {
3.         public struct Result
4.         {
5.             public int classID;
6.             public float score;
7.             public Rect rect;
8.         }

However, this still does not work, giving me the error "Failed to create TensorflowLite Model" from Packages/com.github.asus4.tflite/Runtime/Interpreter

Code (CSharp):

1.         public Interpreter(byte[] modelData, InterpreterOptions options)
2.         {
3.             GCHandle modelDataHandle = GCHandle.Alloc(modelData, GCHandleType.Pinned);
4.             IntPtr modelDataPtr = modelDataHandle.AddrOfPinnedObject();
5.             model = TfLiteModelCreate(modelDataPtr, modelData.Length);
6.             if (model == IntPtr.Zero) throw new Exception("Failed to create TensorFlowLite Model");
7.  
8.             this.options = options ?? new InterpreterOptions();
9.  
10.             interpreter = TfLiteInterpreterCreate(model, options.nativePtr);
11.             if (interpreter == IntPtr.Zero) throw new Exception("Failed to create TensorFlowLite Interpreter");
12.         }
13.  

In Issue #89, the owner made mention of changing the input shape, but did respond as to where this might be located (the default model shape is 1x300x300x3, but mine is 1x640x640x3).

The author of the issue also trained a model with the same input shape as the default, and he says that it still didn't work, but also did not get a response.

Any help is appreciated!

 

hello, please did you solve this I am in the same situation

 

do anyone have solution to import custom tflite models in ssd template