Tutorials on retrain a MobileNet v2 object detection model for Edge TPU #46
-
|
There is a Retrain SSD MobileNet V1 detector for the Edge TPU (TF1) tutorial, also there is Retrain MobileNet V2 classifier for the Edge TPU (TF2) tutorial Are there any tutorials for MobileNet V2 detector for the Edge TPU retraining? |
Beta Was this translation helpful? Give feedback.
Replies: 3 comments 6 replies
-
|
Hello Igro. It depends if you are using an artpec7 or artpec8 camera. For more details about how to retrain and quantize a model, you can also follow these two guides for artpec7 and artpec8 |
Beta Was this translation helpful? Give feedback.
-
|
Hi Angelo, It took me some time to build tensorflow_to_larod_a8_app_1_0_0_aarch64.eap basing on ensorflow-to-larod-artpec8 tutorial. I was able to load it to the cameras and judging from /var/volatile/log/info.log log file everything seems to be working ok. However, now, when I build the application, I don't think this is what I was asking about. I wanted to check if there is a way to use a different model in acap-dl-models module in Object Detector Python example. acap-dl-models does the following: ADD https://github.com/google-coral/edgetpu/raw/master/test_data/ssd_mobilenet_v2_coco_quant_postprocess_edgetpu.tflite models/ My question is - can I replace these links with my own tflite modles and labels, and if yes, is there a tutorial on how to re-train models to be compatible with acap-dl-models? |
Beta Was this translation helpful? Give feedback.
-
|
One of the output in tensorflow-to-larod-artpec8 is a tflite model in app folder - /env/app/coverted_model.tflite. Can I use it in acap-dl-models instead of the two tflite models from google-coral? |
Beta Was this translation helpful? Give feedback.
-
|
Hi Igor Yes, if you have your own .tflite model, you can just replace those lines in the dockerfile with the path of your model. |
Beta Was this translation helpful? Give feedback.
-
|
Hi Angelo, Can you please point me in the right direction on how to trace the issue (please see below) I used the following tutorial: After I updated object-detector-python example with re-trained model, I received the following error message: [+] Running 5/5 |
Beta Was this translation helpful? Give feedback.
-
|
I also tried floating point model, it resulted in slightly different error: object-detector-python-acap_dl-models-1 | detect4.tflite |
Beta Was this translation helpful? Give feedback.
-
|
Keep using the int quantized version. |
Beta Was this translation helpful? Give feedback.
-
|
Hi Angelo, Thank you, I re-tested is with int quantized version. This time it displayed the error messages above for about a minute, but then inference server started working. |
Beta Was this translation helpful? Give feedback.
-
|
Yeah, it happens on artpec8 cameras because before the first inference an internal conversion is executed, it can take up to 1 min. |
Beta Was this translation helpful? Give feedback.
Hello Igro.
It depends if you are using an artpec7 or artpec8 camera.
If you have a quantized model in the .tflite format, with an artpec7 you can just use the edgetpu-compiler to convert the model and use it on the TPU.
For artpec8 this step is not needed, you can directly load the .tflite model.
For more details about how to retrain and quantize a model, you can also follow these two guides for artpec7 and artpec8