# Does my idea make sense? (Image Regression?)

Hi guys. I’m really super beginer for Learning.
I took a basic course in FAST AI for 2 weeks.

Now I major in physics and about to write a paper.

My objective is that Predict a Image with dataset(which is numeric value)

I have dataset which is atomic coordinate(ex. x : 0.5, y : 0.3, z : 0.7…), lattice parameter ( a : 0.2 , b : 0.6 …) and each dataset has a graph(Image) that corresponds to dataset.

My Idea is … ** Assumed the graph image is 20x20

(1) make each column(coordinate x, coordinate y, … lattice parameter a …) 20x1 embedding.
(2) make weights vector(20x1)
(3) matrix multiplication (1) and (2). then I get 20x20 matrix.
(4) Compare that 20x20 matrix with corresponding graph_Image
(5) Use GDS ~~ also neural network.

The idea that make just raw numerical value dataset into matrix by embedding.

Does it make sense? or my method is wrong in someway?

thanks a lot.

Have a good day

Could you share a few real examples from your dataset? I’m having a hard time visualizing what you’re describing. Are you wanting to generate an image based on the tabular data or figure out if an image is a match for the set of tabular data, generate tabular data from an image? What is GDS?

1 Like

Hi Mat
My example is

_cell_length_a 4.91981800
_cell_length_b 5.11699600
_cell_length_c 5.43134400
_cell_angle_alpha 90.00000000
_cell_angle_beta 90.00000000
_cell_angle_gamma 90.00000000

I’m gonna use this 6 parameters and yes I’m trying to generate an image based on tabular data.

I think we’re doing something similar

GDS was wrong. I was meaning Stochastic Gradient Descend SGD.

Sorry about wrong name.

There are a number of techniques for generating images using neural nets. ‘Papers with code’ has several of examples of people using neural nets for image generation: Image Generation | Papers With Code . This is likely going to be a very challenging exercise unless you can find an implementation that does almost exactly what you’re trying to do that is pre-built that just needs to be adapted. I’m not aware of any projects doing something like this that are built using fastai 2, which is the current version of the fastai library and is that the course is based on, but you can do some searching in the forums to see if anything else looks similar. It’s always best to work on projects you’re interested in because that interest will help you work through the frustrating times, but this is likely going to be very challenging and is probably going to take quite a lot of time and effort to get working if you’ve only been working with fast.ai/AI in general for only 2 weeks. It might be a good idea to work on this along with some other challenges/projects that are a little less challenging so you don’t get too frustrated when you run into roadblocks or get stuck on this project. Those other projects will also help bolster your overall knowledge and may be able to help you learn the skills required to complete this project. If you’re just looking for a ‘beginner’ project to get your hands dirty and learn but are not passionate about making it work, this may not be the best fit for your first ‘beginner’ project.

Sorry for being late.

Thanks for your reply and tips.

I’ll try and let you know further ongoings.

have a good day

1 Like