07-03-2019 06:33 AM
I have compressed my model with decent_q (tensorflow). The evaluation model that is generated gives me a sufficient accuracy for my problem. When I deploy the model to the DPUs the accuracy that I measure is terrible. Is it possible that the deployed model and the evaluation model have different behavior? Or is it me, doing something wrong in the C++ code?
07-03-2019 06:43 PM
10-09-2019 10:09 AM
We are facing a similar issue. During evaluation we are getting results to be fine. During deployment in board the results are bad. We expect the eval (post quantisation) and deployment in board to be same. Is there a posibility that they might not be same ?
10-15-2019 10:56 AM
The post quantisation evaluation uses image data scaled by 255. In C++ code for DPU, we scale the image as (image-data/255 -0.5)*2*Scale.
This works for LeNET, miniVGGNet (i/e FCResults from the DPU matches with python evaluation). But for VGG16 we see divergence. Any clue what could be the issue ?
11-09-2019 09:58 AM
11-12-2019 05:58 AM
in your analysis you are measuring the accuracy. However we should also have an idea of the model robustness to quantization: how much the accuracy varies if we quantize (small changes applied to weights...) the model.
I have seen some times that a model trained with small learning rate can be more robust to quantization. I figure that the minimum of loss function J is less susceptible to small variations of weights w happening during quantization.
Vice versa, it can happen that a high LR brings the model to a lucky minimum, where a small variation of w causes a big increment of J
Are you achieving better results after quantization with models trained with small learning rate?
11-29-2019 03:21 AM
Hi, I'm facing the same problem when training a model from TF Slim on custom dataset. Which learning parameters did you change and how does it reflect on quantization eval accuracy?