Yolov3 object detection Inference Atlas200DK Error!
Hello,
I trained a tensorflow Yolov3 implementation from here: https://github.com/YunYang1994/tensorflow-yolov3 and I followed all the steps for generated the .pb model for inference on coco dataset.
After that, with the ATC model conversion tool I converted the .pb to .om model with the followed command.
atc --model=yolov3_coco_shape.pb --framework=3 --output=yolov3_coco_shape --soc_version=Ascend310 --input_format=NHWC --output_type=FP32 --input_shape="input/input_data:1,416,416,3"
Then, for the inference process with the .om model I used the same pre-process and pos-process functions from the original tensorflow implementation. My script is attached with this message. (inference_yolov3.txt).
The implementation run without errors but the inference results with the .om model is unsatisfied.
I noted that the outputs of the model.execute function for the om model has the same dimensions that the original tensorflow implementation, however the arrays values are different. The image above shows an example
Could someone explain to me where is the error??
Thanks in advance for your support
- x
- convention：
- Emotion(0)
|
|
|
|
Hello, after a segmented investigation, we found that there is a problem in the preprocessing part of the model. The input data should be of float32 type. Tf may automatically convert the input data type according to the input data type of the model during model inference, but the om model does not Doing this step according to the input data type of the model resulted in the deviation of the input data and no error was reported. The inference code of the om model we wrote based on the original model pre-processing and post-processing code is modified as shown in the figure, and I hope it can help you:
|
|
Comment
- Politically sensitive content
- Content concerning pornography, gambling, and drug abuse
- Content that may disclose or infringe upon others ' commercial secrets, intellectual properties, including trade marks, copyrights, and patents, and personal privacy
