Model not working as intended

[rancid raft]

So, You train the model from scratch with 70 exemplars? If so, that doesn't seem like nearly enough training data. You might need two orders of magnitude more.

[plush valley]

Its basically all the known "official" combinations. I would think it would be enough to at least determine presets with similarities/patterns to consider. The main problem is that it isn't giving the correct type for a preset thats already listed as well. Like Rock_Leone_145_WB as seen in the dataset gif.

I wanted to use mainly official presets which have all data about it known so that its easier for the model to understand

[dire parcel]

If you have a very small amount of discrete data and you know that all your uses will be in that data, you're better off just using a lookup table and no ai

[plush valley]

It wont all be in that dataset. Basically It needs to determine that Earth_Virgo_T125_ES is Stamina despite that specific preset not being in that dataset.

There is different presets associated with stamina to know that Earth, Virgo, T125, ES are all stamina centric. As well as the stats and totals themselves that would show that stamina is the highest stat.

So far its not able to pick up on a preset already in the dataset (Rock_Leone_145_WB) which is why i know it isn't working correctly

[rocky osprey]

Either way you still need more data
Usually datasets collect like 1000 to start off with something that barerly works
Maybe training a model on pokemon data would be better in this case

[plush valley]

There really isn't anything else I could do? Considering I don't really have much options for increasing the data if i wanted to. I could look into data augmentation if thats a good option.

[rocky osprey]

I think u could increase the dataset count artificially or something 🤔

Not really sure

[plush valley]

I'm going to attempt doing that, but based on the issue I'm facing with presets that are already in the dataset not being correctly picked up, do you think its because something in my code isn't working correctly, OR is it because of lack of data?

Also im a beginner with Tensorflow specifically

Because I feel like if its not working correctly adding more data wont necessarily change that, so before I do that I need to know for sure.

[rocky osprey]

I really don’t know too I’m a beginner x)
But what I’m sure if the dataset is imbalanced (one of the classes is the majority by a large margin) then it might be the data

U could check by the count of the rows that have the attack class

[plush valley]

That may be the case, but there is the same amount as balance class rows as there is attack class rows

but the code said there was a bias to those 2

[plush valley]

21 attack and 21 balance rows

[rocky osprey]

Seems like a balanced dataset

How many rows are there?

[plush valley]

70 rows 18 columns

[rocky osprey]

U still need more rows

70 inputs won’t gonna give it a good estimate

For the code I don’t know if it’s correct or not sry x)

[plush valley]

No worries, gonna still see how i can get more data though.

[plush valley]

Yeah I think there's a problem with the code, I increased it to 1050 rows using augmentation and the same problem still happens.