juandes/pikachu-detection

ISSUE when deployed on app and tested

NidhiSultan opened this issue · 5 comments

NidhiSultan commented

Hi @juandes , I tried the android api from tensorflow site with tflite version model. But not able to get detect any box accurately when trying to test it using the app. Did you face any issue similar to this?
Though my model is working fine when tested on images using postman.

juandes commented

Hi @NidhiSultan. My model is quite outdated and not made for tflite. That's probably why it isn't working. Let me know if you find a fix. Or even better, if you manage to convert it to tflife.

NidhiSultan commented

I managed to make it work. Thanks

  • 👍1
juandes commented

Hi @NidhiSultan , do you mind explaining how? It might help others in the future. Thanks.

NidhiSultan commented

Hi @juandes , The model we train and convert to tflite is either Quantized or non quantized. The Android application provided by tensorflow has a parameter, "isQuantized".
We have 2 options.

  1. Quantize our model while converting it to tflite for it to support the application, https://www.tensorflow.org/lite/performance/post_training_quantization.
  2. Change the parameter "isQuantized" to False, assuming the android application here is the one take from official tensorflow site mentioned here. https://github.com/tensorflow/models/tree/master/research/object_detection

Thanks Nidhi

  • 👍1
juandes commented

Hi @NidhiSultan , amazing explanation. I really appreciate the effort. Thanks!