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

[raprakashvi]

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]

@robinvanemden @kedars any thoughts?

[BlueSkyB]

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: 020231e.

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