Tflite prediction is totally different than frozen inference graph prediction

问题

I worked on eye region localisation project and I trained my own custom dataset to create a model using Tensorflow library.

I produce .ckpts files ( model ), I get acceptable results, I convert this model to .pb frozen inference model and I test the accuracy of the frozen model on my webcam and it works fine.

The problem is when I convert .pb model to tflite model. I get so bad results using an Android application and MLkit firebase custom model.

I have posted this issue on GitHub ( Tensorflow repo and firebase repo ) Github link but I didn't get any answer, I really need to solve this problem.

Result using frozen inference( python script ):

Result using tflite model ( Android device )

That was some java code ( Android ):

 private void useInferenceResult(float[] probabilities) throws IOException {
        // [START mlkit_use_inference_result]
        String[] result=new String[80];
        String x="";
        String y="";
        ArrayList<Point> listpoint= new ArrayList<Point>();
        double viewWidth = canvas.getWidth();
        double viewHeight = canvas.getHeight();
        double imageWidth = mutableBitmap.getWidth();
        double imageHeight = mutableBitmap.getHeight();
        Log.i("viewWidth","viewwidth "+viewWidth);
        Log.i("viewHeight","viewheight "+viewHeight);
        Log.i("imagewidth","imagewidth "+imageWidth);
        Log.i("imaageHeigh","imageheigh "+imageHeight);
        double scale = Math.min(viewWidth / imageWidth, viewHeight / imageHeight);
        Log.i("Scale","Scale"+scale);
        try {
            for (int i = 0; i < probabilities.length; i++) {

                Log.i("MLKit", String.format("%1.8f", probabilities[i]));
                float i1 = probabilities[i];
                Log.i("floaaat", "" + i1);
                x = String.format("%1.8f", probabilities[i]);
                y = String.format("%1.8f", probabilities[i + 1]);
                Point p = new Point(x, y);
                i = i + 1;
                p.setX(x);
                p.setY(y);
                listpoint.add(p);              
            }
        }
        catch(Exception exc){
            Log.e("Exception","Error:  "+exc);
        }
        for(int j=0;j<listpoint.size();j++){
            try {

                String xx = listpoint.get(j).getX();
                String yy = listpoint.get(j).getY();
                xx=xx.replace(",",".");
                yy=yy.replace(",",".");          
                float xx1 = Float.parseFloat(xx);
                float yy1 = Float.parseFloat(yy);
                Log.i("Float results", "point_" + j + "(" + xx1 + ", " + yy1 + ")");
                Log.i("Scale","Scale  "+scale);
                drawpoint(image2, (xx1*(float)scale*293) , (yy1*(float)scale*293) , 1);
            }
            catch(Exception esa){
                Log.e("Exception","Exception:  "+esa);
                Toast.makeText(this, "Exception"+esa, Toast.LENGTH_SHORT).show();
            }

        }

        }

// drawbitmap function

  private double drawBitmap(Canvas canvas) {
        double viewWidth = canvas.getWidth();
        double viewHeight = canvas.getHeight();
        double imageWidth = mutableBitmap.getWidth();
        double imageHeight = mutableBitmap.getHeight();
        double scale = Math.min(viewWidth / imageWidth, viewHeight / imageHeight);

        Rect destBounds = new Rect(0, 0, (int)(imageWidth * scale), (int)(imageHeight * scale));
        canvas.drawBitmap(mutableBitmap, null, destBounds, null);
        return scale;
    }

How can I solve this error?

来源：https://stackoverflow.com/questions/57442372/tflite-prediction-is-totally-different-than-frozen-inference-graph-prediction

标签

android

machine-learning

tensorflow-lite