I trained a deeplabv3 semantic segmenation network.
I follow the example Quantization aware training in Keras example | TensorFlow Model Optimization
to quantize this model.(learn rate is 0.01, epoch is 300, dataset is 300K, traning loss: 0.172 validation loss: 0.18)
before the quantized model the miou is 77%.
after the quantized model the miou is 76.5%
it seems performed well.
but when i compare the segmenetaion image before quantization and after quantization.
I found the leaves were segmented well before quantization(pb model)
but leaves were segmented bad after quantization(Tflite model).
Can you suggest me how to improve the QAT segmentation performance in real image?
Hi @robin_peng, If possible could you please provide a stand alone code to reproduce the issue for further analysis. Thank You.