Hi. I’m a beginner on DL.
Suppose that I want to build a network for FR task on possibly my own dataset. In the assignment, I saw that we use a pre-train model to make prediction. What I want to know is that what are the possible options of network for a personal project. Few things in top of my head are:
- Use pre-train model as we did in the course. Possibly use the model given in the assignment. But seems like some people including myself cannot even load the model “keras-facenet-h5/model.h5”. I hope there’s some solution for this.
- Use another pre-train deep model. I don’t know where to get this, or even choose the right one. I have to some quick research for this. Is getting some free pre-train deep/large model possible?
- Is there another not so deep model that I can use or train that give a decent result/performance?
What kind of architecture/model should I try first? Thank you all.
Hey @kelvinn, even if you want to get a decent result for an image-related task like FR, then the best option is to use a per-trained model from my personal experience.
As far as the model you should use is concerned, you can always look for the latest implementation of the one that is used in the course, or you can create a virtual environment, in which you install the same version of packages as the model was trained on, whichever option attracts you more.
Also, you can use various other models. You just have to crawl through the web, a little bit.
Thanks @Elemento I’ll try to build a project using pre-trained model.
I saw that you used to have the struggle over keras-facenet-h5/model.h5. Do you succeed to run on your local machine?
I found the solution, which was to create a virtual environment and installing the exact versions of the packages using which the model was trained on. But then, since I only wanted to practice, I chose the simpler way. I simply made a copy in the Coursera jupyter itself, and then did the assignment.
Ok Thank you for your help @Elemento I’ll need your help later when I want to build the virtual environtment.