Help regarding CNN training

Hi,

I am currently working on some signal processing techniques involving time-frequency analysis. Most of the cases, researchers are plotting the scalogram and feeding it as an image to train a pretrained CNN network by transfer learning process. Is there any way to train the network using some standard statistical features (e.g. mean, std, kurtosis, skewness etc.) instead of using the image of the scalogram? I am new in the field of deep learning so it will be much help if anyone can give any recommendation.

Thanks in advance.

Answer 1

[Shounak Mitra]

Hi,

O fcourse, check this out. https://www.mathworks.com/help/audio/examples/speaker-identification-using-pitch-and-mfcc.html

[Kaniska Samanta]

Hi Shounak,

Thanks for the assistance, but the link you shared is an example of training a kNN classifier. I am looking for a way to train a pretraind CNN (e.g. GoogleNet or ResNet50) with extracted features. The main problem is, pretrained CNN networks have image input layer which will not accept any feature vector of 500x20 where 500 is the number of datasamples and 20 is the number of features. If you can share any standard code, it will be great help.

Thank you again..

[Kaniska Samanta]

Hi,

I am currently trying to train GoogleNet network with 7 statistical features. I have converted the feature vector of 180x7 (double) into table and added the corresponding Labels. I am getting the following error,

7 sub-columns of the column 1 are the seven features. What is wrong? Please let me know if I am doing any fundamental mistake. I am quit new in this field so any recommendation is welcome.