CODEWITHSUNDEEP
machine-learningartificial-intelligenceconv-neural-networkpytorch
LogicShy
LogicShy

Reputation: 11

Using MIT Indoor scene database in CNN

I'm an engineering student and kind of a noob to programming. I'm taking an AI course, currently trying to do my final project.

I have to create a CNN net, I have to use de MIT Indoor scene database (it can be found here: http://web.mit.edu/torralba/www/indoor.html). I don't have a problem doing the CNN since I've done a few before in the semester using CIFAR10, but I'm having trouble with this one since I don't know who to use this set of images.
I think I need to create a dataset of my own, I've tried with PyTorch using https://pytorch.org/tutorials/beginner/data_loading_tutorial.html, but I get confused because I don't have a .csv with features, I have a lot of .xml files with several features for each picture. Also, I don't have a file just saying "bedroom, bar, etc" as I've seen in other tutorials.

I would rather use PyTorch since I can use the "train_test_split" function, but if anyone could help me to understand how to make those 15620 my input to the net, I would really appreciate it.

Upvotes: 1

Views: 237

Answers (1)

Mircea
Mircea

Reputation: 1257

you can generate your own csv files, although, you might not need it. There is a good tutorial on pytorch website https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html#load-data, which is very similar or easily applicable to your case.

MIT Indoor dataset has the images one folder per class, and the txt files mentioned on the website are the train / test splits.

So, if you create the following folder structure:

train
|- class 1
|- class 2
...
|- class n

and the same for val / test, it should be straight-foward to use (adapt) the datasets.ImageFolder example for your case.

Upvotes: 1

Related Questions