search

STMicroelectronics / stm32ai-modelzoo

AI Model Zoo for STM32 devices
Other
278 stars 69 forks source link

Error running custom model on STM32MP257 #49

Open yoppy-tjhin opened 8 hours ago

yoppy-tjhin commented 8 hours ago

Hello,

Some progress since our issues #47. We have finished training a custom dataset for ssd_mobilenet_v2_fpn according to the wiki guide. And also finished quantizing the model, with the following parameter:

quantization: quantizer: TFlite_converter quantization_type: PTQ quantization_input_type: float quantization_output_type: int8 granularity: per_tensor #per_tensor optimize: True #can be True if per_tensor export_dir: quantized_models

But when we run the model on the STM32MP257 eval kit, got the following errors:

RuntimeError: [TFLITE] Failed reading output: Unsupported output tensor type. Traceback (most recent call last): File "/usr/local/x-linux-ai/object-detection/stai_mpu_object_detection.py", line 396, in new_sample self.app.nn_result_locations, self.app.nn_result_classes, self.app.nn_result_scores = self.nn.get_results() ^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/x-linux-ai/object-detection/ssd_mobilenet_pp.py", line 113, in get_results anchors = self.stai_mpu_model.get_output(index=2) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/site-packages/stai_mpu/stai_mpu/network.py", line 48, in get_output output_tensor: NDArray = self._exec.get_output(index) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^

What could be the problem?

Thank you.

abrSTM commented 4 hours ago

Hello,

First of all, I'd like to make sure I understand the operations you've performed:

To be able to run the model on our MP257 target using the object detection Python application, you need to modify the post-processing file to match the output of your model.

  1. Did you modify the post-processing file?
  2. In the "standard" ssd_mobilenet_v2_fpn, we have 3 outputs. Is it the same in your custom ssd_mobilenet_v2_fpn?

If the number of outputs is different, please modify the /usr/local/x-linux-ai/object-detection/ssd_mobilenet_pp.py file to get the correct number of outputs.

Regards, ABR