Train­ing and retr ain­ing

Linux Format - - Tutorials -

A trained neu­ral net­work is able to take in a piece of data, such as an image, that it hasn’t seen be­fore and out­put a pre­dic­tion about what the data is, such as a cat­e­gory of flower, fish, car or what­ever it’s been trained on. The process of train­ing takes heaps of train­ing data, which in this case is im­ages la­belled by what’s in them and tunes all the nodes in the net­work to ac­cu­rately cat­e­gorise all that train­ing data and new data.

This is done by turn­ing the image into a huge num­ber of pa­ram­e­ters, which you can think of as each pixel or groups of pix­els. In­side the net­work, there are weights and bi­ases that de­ter­mine which pix­els and groups of pix­els are used to iden­tify which cat­e­gory the image is ex­pected to be in.

Train­ing a net­work re­quires lots of data and com­put­ing power, which you don’t have on a Rasp­berry Pi. Yet it turns out that you can re­train a net­work by tak­ing an ex­ist­ing trained net­work, which uses much less power and data.

Re­train­ing en­ables you to take an ex­ist­ing net­work, keep­ing most of the net­work but re­plac­ing the fi­nal cat­e­gori­sa­tion layer to suit your new train­ing data (the la­belled im­ages). You can think of this as keep­ing the parts of the net­work that can iden­tify shapes while re­plac­ing the fi­nal layer which cat­e­gorises im­ages, and us­ing that abil­ity to iden­tify shapes.

Newspapers in English

Newspapers from Australia

© PressReader. All rights reserved.