search

espressif / esp-dl

Espressif deep-learning library for AIoT applications
MIT License
541 stars 117 forks source link

Quantization Layer Not supported: Softmax and constant (AIV-702) #168

Open raprakashvi opened 1 month ago

raprakashvi commented 1 month ago

Hi, I am following the tutorial https://blog.espressif.com/hand-gesture-recognition-on-esp32-s3-with-esp-deep-learning-176d7e13fd37 and have been getting issue on quantization step.

My python version is 3.7 and whenever I get to quantization , I get the error: Generating the quantization table: Constant is not supported on esp-dl yet LogSoftmax is not supported on esp-dl yet

My layers are simple and as shown below: model = Sequential() model.add(Conv2D(32, (5, 5), activation='relu', input_shape=(96, 96, 1))) model.add(MaxPooling2D((2, 2))) model.add(Dropout(0.2)) model.add(Conv2D(64, (3, 3), activation='relu')) model.add(MaxPooling2D((2, 2))) model.add(Dropout(0.2)) model.add(Conv2D(64, (3, 3), activation='relu')) model.add(MaxPooling2D((2, 2))) model.add(Flatten()) model.add(Dense(128, activation='relu')) model.add(Dense(6, activation='softmax'))

I have checked that softmax is supported in the layers. I am using calibrator.so file on linux. Should I be using convert.py to work around?

raprakashvi commented 1 month ago

@robinvanemden @kedars any thoughts?

BlueSkyB commented 1 month ago

For the constant operator, you can add 'extract_constant_to_initializer' in the onnxoptimizer.optimize() within the optimize_fp_model function. Refer to this commit for the modification: https://github.com/espressif/esp-dl/commit/020231ea04ef27e8cf23471a15365513a46eeeeb.

For the LogSoftmax operator, esp-dl does not currently support it; you can check your ONNX file again.