Using the pretrained SSD Mobilenet v2 model from Tensorflow Model zoo, I trained it on a custom dataset based on three classes [Non-Defective, Defect-X, Defect-Y] for 30000 steps. I have followed the tutorial to build my custom model. The regular tensorflow model performs well on the test dataset.
Through post-training quantization, I have a TFLite model, wherein the input type is uint8 and the output type is float32.
The code used for converting frozen (.pb) model into TFLite model:
converter = tf.lite.TFLiteConverter.from_saved_model("Intermediate TFLITE frozen model (.pb)")
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8, tf.lite.OpsSet.SELECT_TF_OPS]
converter.inference_input_type = tf.uint8
converter.allow_custom_ops = True
converter.representative_dataset = representative_dataset_gen
tflite_model = converter.convert()
However, on the test dataset, for every image, the TFLite interpreter always returns the same parameters (label_indices, scores, rects) in the same sequence:
e.g. For "Non-defective_image.bmp", the TFLite interpreter returns an array of class indices like [0, 2, 2, 2, 2, 2, 2, 2, 2, 2]
Array of Scores: [0.6009013 , 0.3969356 , 0.33033633, 0.19510852, 0.18718699, 0.17808083, 0.177586 , 0.17257103, 0.16170672, 0.16149464]
For "Defect-X_image.bmp", the results are same as the above.
The same problem persists for a full-precision (float32) TFLite model as well.
Any idea on how to go about this?
I have tried running TFLite model with different quantization methods, but ended up getting the same result.