simuation result of IP different with hls_model.predict

[Feng Gao]

Dear all,

This is Feng from ULB,belgium. 

Following hls4ml tuturial part1,  i get a IP, https://github.com/fastmachinelearning/hls4ml-tutorial/blob/main/part1_getting_started.ipynb.

But i met a problem when i simulate it using vivado. the prediction result is really bad. the detail is :

• The backend : vivado 2018.3
• The ML model: 1layer MLP.
• The configuration i change a bit:  change the input precision from ap_fix<16,6> to ap_fix<8,8> because my data is integer smaller than 128. the left precision is changed to ap_fixed<16,8>
• the accuracy of y_hls( y_hls = hls_model.predict(X_test)) is 92%,
• then i simulate the result of the IP using same testdata, but the result show the  accuracy is only ~51%

I will really appreciate for your help.

Best regards,

Feng