Accuracy Pokemon Classification

#1

Followed all the steps for data preparation and model yet getting a train accuracy of 2% only. I m not able to find the error.
I am working with complete pokemon dataset i.e. 10120 images and 149 classes

Here is my code:

#2

Can any TA please reply???

#3

hey @devanshmarwaha27 ,
first of all i am really sorry to reply so late.
Now comming to you doubt,
while working on images , the way we extract information matters a lot in neural networks .
And for this purpose we mostly times use pre built layers from keras , Like CNN ,etc.

You are currently using simple Dense layers to get data from those images , which doesn’t work well.
and you are providing an input of 4800 and converting it to 100 , means thinking it like as its gonna convert data from such large number of nodes into just 100 nodes.
and at the end you want to get results for 149 classes.
Currently the model , 4800 -> 100 -> 50 -> 149 .
Such structure isn’t able to learn data and hence performs very very poor. This is the reason your getting only 2 % accuracy .

To improve , you need to use convolutions majorly to learn from images , and use that data extracted to get predictions.
and just think normally , can you define between 149 objects based on 50 features , i guess yes not exactly. If a human mind is incapable of doing this properly then how can network much better on this.

Suggestion : First try working with the small dataset the sir had used in video with 3 output classes. Experiment on it , and once you understand those concepts properly then move to this.
But, let me tell you that for such big dataset with so many output classes , you need to get features from images extracted much more properly else , it will not perform well. You can try increasing your model size , but its just an experiment.

I hope this helps you :slightly_smiling_face.
Thank You :slightly_smiling_face:

1 Like
#4

Thanks alot sir. You can close this doubt diiscussion here. :smile:

#5

hey @devanshmarwaha27 ,
Its good that you have understand the problem.
But actually we are asked to get doubts closed by the student itself , so it would be great if you can yourself mark this doubt as resolved in your course doubt section.

Thank You and Happy Learning :slightly_smiling_face:.