Calculate a negative value in bias_shift (shift right)

[ccnankai]

At the final fc layer bias dec bit =10. I calculated the last layer, wx=8 ， bias=10 ， bias_shit=-2，bias need to be shifted right 。 But there is no option for the right shift 。 what should I do with this situation.

[ccnankai]

@saichand07

90

last inputlayer * w =8bit dec bias=10 ====> bias_shit=-2

[ccnankai]

@saichand07 Sorry, I am busy recently. The calculation step has been told in this question #51 .Check out the quant_guide and re-read the quantization section again.The steps are simple.Discuss problems together. swdtianxia@gmail.com

[saichand07]

@ccnankai

I have calculated all the bias_shifts and out_shifts layer by layer by adjusting the input and output format (using act_max ), like they explained in the quant guide. I put those values in ds_cnn.h and I also tuned the recording window, averaging window len and detection threshold. but still deployed model on board is not detecting words correctly.

could you please tell me what else I need to taken care of. how did you deal with negative shift ?

Thank you very much

[ccnankai]

@saichand07 Did you calculate the MFCC_DEC_BITS ? How did you calculate it.Are you calculating the other *shift parameters correctly?

[saichand07]

yes, I calculated shift parameters as they explained in the quant guide, How to calculate MFCC_DEC_BITS ? Actually, I didn't change its value, I put MFCC_DEC_BITS = 1 and also changed it to 2, but still its detecting wrong words

[saichand07]

@ccnankai my first layer Q5.2 weights Q-3.10 bias Q1.6 output layer Q3.4

BIAS_RSHIFT = 2+10-6 = 6 OUT_LSHIFT = 2+10-4 = 8

I calculated all layers like this,

last layer I got negative shift and I don't know how to deal with it, Pool layer act_max I took Q3.4 and last layer Q5.2 so shift of pool layer is 4-2 = 2

could you tell me whatelse I need to consider ?

[ccnankai]

@saichand07 I think you still have the wrong calculation of the parameters.I can see good results.

[ccnankai]

Weights Q-3.10 Wrong, how did you get it?

[saichand07]

after running Quant_test.py the weights are generated randomly, my parameters are window size_ms 40 window_stride_ms 20 dct_coefficient_count 10

[ccnankai]

Weights Q-3.10 This parameter is definitely wrong. No need to change any thing in fold_batchnorm.Are you a student, is this algorithm used for experimentation??

[saichand07]

yes, how can we deal with this situation ?

[ccnankai]

How did you get this result Weights Q-3.10，This calculation step is wrong.

[saichand07]

@ccnankai thanks, I found the problem, weights were wrong now its working

[weishu20]

@ccnankai thanks, I found the problem, weights were wrong now its working

hello，can you tell me how to calculate Q of input and output？I have no idea....

[ccnankai]

@weishu20 the first layer is input layer(mfcc)