search

espressif / esp-dl

Espressif deep-learning library for AIoT applications
MIT License
516 stars 115 forks source link

quantization tool failed at eva.generate_quantized_model (AIV-552) #103

Open skodippily opened 1 year ago

skodippily commented 1 year ago

Hello, Error while executing the line below in example.py. eva.generate_quantized_model(model_proto, pickle_file_path) I slightly modify example.py to load yolov2tiny model to it.

error: image

Optimized model file satisfied all the layers defined in the link here

Could you please help me with this problem.

Auroragan commented 1 year ago

HI, what is scalarPreprocessor_scale/bias? Can you find them in the onnx graph?

skodippily commented 1 year ago

Yes, those are under MUL and ADD layers image

I'm trying to do a project base on object detection using ESP-DL. Most of the pre-trained models are not supported with built-in operators right? So the only model that I could find is this tinyyolov2 model. If you know any other pre-trained model supported ESP-DL, please let me know.

Auroragan commented 1 year ago

are these two nodes for normalization?

skodippily commented 1 year ago

Not quite sure about that. I cannot find any description of each layer and its task. It's a pre-trained model, and haven't modified anything on it. The only thing that I noticed is when I do the optimization, BatchNormalization layer will fuse to CONV layer.