Lost 40% accuracy after converting to tflite model

[rudikx6]

Hi Nam Vu,

first thank you much for that nice instruction and the amazing possability through Colab :)

I used your colab notebook to train my own coustom model. Only the *.record files i created by another way. Edje Electronics' tutorial

Could be this the reason why get lost 40% of the accuracy after converting to tflite model? I got 5 from 5 correct detection of my own 5 Images after 10000steps training from the frozen inference graph (Step.8 in your tutorial) after converting to tflite only 3 from 5....

I have only one class in my trained model.

Greetings from Germany, Rudi

[Namburger]

@rudikx6 hummnn, I'm able to use this same pipeline for multiple dataset without this issues so my suggestion would be to add more images to your dataset, the tfrecord should be the same so that shouldn't matter

[tino926]

@Namburger, I encountered the same problem with the notebook provided in this repo. I did little modifications:

1. mount google drive to save the training progress, because free google colab may interrupt the training.
2. add ngrok token since ngrok now requires it
3. set pipeline.train_config.num_steps = 10200
4. install tflite_runtime-2.5.0-cp37-cp37m-linux_x86_64.whl instead of tflite_runtime-2.5.0-cp36-cp36m-linux_x86_64.whl

(the modified notebook can be seen at https://colab.research.google.com/drive/17sgw8hODl9zQt1689Y8bZLFQBk5sh165?usp=sharing )

this is what I got:
result from frozen_inference_graph.pb

result from ssdlite_mobiledet_dog_vs_cat.tflite

If there is no such inconsistency in your environment, could you kindly provide the version of tensorflow and commit of models you used.