OpenMV Firmware v4.5.6 and up TensorFlow Porting Guide

[kwagyeman]

Yeah, sorry, I was on auto pilot posting that firmware. Please ignore it. I deleted the post. The issue is related to what Ibrahim said.

[iabdalkader]

Please don’t use that firmware it might be broke. As for quantization you need full int8 quantization (weights and activations) no int16. If you’re using the latest tf, see this post:

Heads up custom Tflite models with Latest versions of Tensorflow

Doing an update of all my tools for ML and upgraded to Openmv 4.5.6. In doing so I observed that Tflite models that were created and quantized with Tensorflow 2.8.2 were working but when using Latest tensorflow 2.17 I kept getting failures while loading to Openmv. With this flag in my script before quantization process I was able to successfully load the model though converter = tf.lite.TFLiteConverter.from_keras_model(model) converter._experimental_disable_per_channel_quantization_for…

For ops, you can use https://netron.app/ and check if the model has ops not supported by tflite micro

tensorflow/tflite-micro/blob/main/tensorflow/lite/micro/micro_mutable_op_resolver.h

I will test the model as soon as I can.

[iabdalkader]

Here’s the error:

tflm_backend: Failed to get p code CUSTOM

Looks like the model has a custom op, which is something that could have easily been checked using https://netron.app/. This is why we don’t enable logs because they don’t tell us anything new, but they take about 60KBs of precious flash.

[Bugrahan]

Hi Everyone,

I was testing the new API and I got an error about the ml module while calling the net.predict() method. The error message is “Could not open ml/model.py for reading. Either the file does not exist or you do not have the permissions to open it”. What might be the reason for it? I am using H7+ with Version 4.5.6

EDITED:
That was my bad I was using the method like net.predict(img), it gaves the error becasuse of it. I have changed into net.predict([img])

[paolo]

Hi,
I’m using tf.classify(model_name, snapshot), how can I continue to work with my TFlite model without modifing my code? Or do I need to downgrade the firmware?
Thanks

[kwagyeman]

You need to modify your code to the latest API or downgrade the firmware. Updating to the latest API for classify() is extremely easy per the porting guide above.

[paolo]

Hi, great, I got it working. Just need too fine-tune again… Thanks.

[sjh2100]

i have no idea about the TensorFlow . hope openmv give some link about the guide of tensorflow on openmv step by step.
if the openmv stack is 4m+ ,hope give some example about the large size of frame example .such as AprilTag on FHD.

[kwagyeman]

Hi sjh2100, you can run AprilTag on FHD images. It just runs at a very low FPS. You can also run YOLO models with the H7 Plus and RT1062. They just run at a low FPS too.

Please open a new topic for help support versus hijacking the porting guide topic.