How did you solve it?
İ didnt solve, i converted another network sample.
Maybe you can edit. Json file,and you can remove nonmaxlayer,
İ finded this, Maybe helpful to you
Mobilenet-ssd for caffe implamation
That is a possibility, but I could not find a way to use that json with Acuity tools, or just use it with the standard tools as such as export_inference_graph.py script ? Certainly the json provided on the Acuity zoo model website looks different but there is no .pb file for that https://verisilicon.github.io/acuity-models/viewer/render.html#../models/mobilenet_ssd/mobilenet_ssd.json
I tried a different approach: export_tflite_ssd_graph on check point data and then try to convert. Now I’m getting ad different error on RealDiv. Is RealDiv supported by Acuity?
D Try match RealDiv scale_logits
W Not match node scale_logits RealDiv
Traceback (most recent call last):
File “convertensorflow.py”, line 62, in
File “convertensorflow.py”, line 58, in main
File “acuitylib/app/importer/import_tensorflow.py”, line 150, in run
File “acuitylib/converter/convert_tf.py”, line 584, in match_paragraph_and_param
File “acuitylib/converter/convert_tf.py”, line 483, in _tf_push_ready_node
TypeError: ‘NoneType’ object is not iterable
Awesome! Congrats
Is SSD detection next ?
@blitzfr1 I will consider training an SSD model. yolo model is not bad either. You can use Yolo model first .
got this error
Could not find a version that satisfies the requirement tensorflow==1.10.0
Install it to PC or board ?
installed on VIM3
Have installed other software on it previously
iow not a clean install.
thanks
Do you mean that there yolo darknet detection already done and trainable ? (Not classification)
The training needs to use PC. the NPU of vim3 can accelerate the operation of the model and can not be used as training. For the training on PC, you can refer to the one on DOC, which is verified by my practice. SSD and Yolo models do not classify the picture sets, but they need to label the pictures.
Got it. Thanks
What about the FPS you can achieve, in your test? Tiny yolo or YoloV3 ?
In PC, the FPS have 30 , but just 10 in board. The yoyl_v3 model is a little bit stuck, which is related to the calculation method of the model. I am still optimizing it. At the same time, I am training a yolo_v2 model, which will run higher FPS
1 Like
Thank you
It Worked!
root@Khadas:/npu/aml_npu_sdk/acuity-toolkit/conversion_scripts# ./0_import_model.sh
./0_import_model.sh: line 14: …/bin/convertensorflow: cannot execute binary file: Exec format error
root@Khadas:/npu/aml_npu_sdk/acuity-toolkit/conversion_scripts# file …/bin/convertensorflow
…/bin/convertensorflow: symbolic link to ./acuitylib/convertensorflow
root@Khadas:/npu/aml_npu_sdk/acuity-toolkit/conversion_scripts#
You can’t run this tool on a AARCH64 architecture board, you can only run it on a X86_64 PC.
2 Likes
@RichardG This tool just run on PC and tensorflow 1.10 .You can train with tensorflow-gpu 1.1x,but you need tensorflow 1.10 when you train.