Concatenation has inconsistent input parameters quantization parameters about tinyneuralnetwork

[LynnL4]

I tried to add a quantization operator below the nodes with different quantization parameters, and this solved the problem I encountered. But there are still some details that are not taken care of.
LynnL4@6f62020

[peterjc123]

Looks like a duplicate of #95 (comment)

[peterjc123]

I think this part of the model should not be quantized, not only because of this, but also there are something that cannot be quantized like pow.

[LynnL4]

If not quantized, it would be unfriendly for devices like MCU, such as esp32s3. When exporting the YOLOv5 project to int8 TFLite, a quantization operator will also be added in this case. However, it is important to ensure that the values of the input tensor are within the same magnitude to avoid losing a significant amount of precision. Tomorrow, I will try exporting YOLOv8 to TFLite to see what the situation will be. The YOLOv8 project uses onnx2tf.

[peterjc123]

If not quantized, it would be unfriendly for devices like MCU, such as esp32s3.

Yes, but it seems the Pow op has Quantize and Dequantize ops around it, which is already not optimal. Even if #95 (comment) is resolved, we will add three more Quantize ops to perform re-quantization before the Concatenation op. Actually, I don't know how is the performance of the those ops using Float32 kernels on your hardware.

However, it is important to ensure that the values of the input tensor are within the same magnitude to avoid losing a significant amount of precision.

Yes, I agree with you. Otherwise, we can just use the q params of the output of the Logistics op as the unified one.

The YOLOv8 project uses onnx2tf.

I suppose they will add Quantize nodes before the Concatenation nodes, which may lead to the same result.

[peterjc123]

Let's continue the discussion in #95.