Unable to Reproduce Testing Results

[adthorson]

Heewon Chung,

To my understanding, I have implemented your NN architecture using keras 2.2.4. I was able to reach and exceed your training accuracy; however, I'm unable to reproduce your testing results (BAMI-2).

It would be greatly appreciated if you provided the source code. I would like to verify if I am somehow incorrectly implementing your NN architecture.

A few questions as well:

• How do you choose the HR value from the222-output layer? Do you choose the associated HR value with the max "FFT" bin / top layer node?
• How do you label your training frames? Average HR within the 8 s window?
• Do you do any dataset leveling?

[adthorson]

I have reached out to Prof. Jinseok Lee, who I believe to be a part of your BAMI research lab; however, he has yet to respond. Thank you!

[HeewonChung92]

Hello, Thank you for your interest.

I can't give you all the code, but I can give you the code for the models. If you write your email address, I will send you a zip file. (related my models)

1. The 222-output layer meanings: It is the frequency about HR. It has the 'myFrequency' in my_data.mat files. and the variable myHR means the HR. The HR can be found by multiplying the frequency by 60. (myHR = myFrequency * 60).

2. Parameter input_dim = 222 input_rows = 2 output_dim = 222 time_steps = 6 batch_size = 1

total_epochs = 50 learning_rate = 0.0001 dropRate = 0.3

Have good luck for a new year 😃

[adthorson]

Thank you for responding!

My email address is adthorson@wisc.edu

[mintisan]

@HeewonChung92 Could you send a copy to me ? My email address is mint@mail.ustc.edu.cn. Thanks.