Load data into ImageDataBunch

(Harsha) #1

My directory is like this.
1/
Images/
1_1.jpg
1_2.jpg
2/
Images/
2_1.jpg
2_1.jpg
How to load data into databunch?

When I use from_folder all my images get tagged as ‘Images’

0 Likes

(Dominik Engel) #2

Hi there,
the from_folder() part should not be the issue since this will just grab all files with image extension from your directory. You have probably used the label_from_folder() function?? You might want to use
label_from_func(lambda fn: fn.name.split('_')[0])
The input to the lambda will be the paths to your images as pathlib object. the .name gives you the name of the file as string and .split('_') splits this name at every _. From you example it seems the very first part of the filename, before the first _ is the class, therefore take the first element of the list that is returned from split() as label.

Greetings

1 Like

(Harsha) #3

Thanks @xeTaiz.
I was able to get the training images.

But looks like validation images are named as
val_1.jpg
val_2.jpg
All the result are saved in csv
Val_1 classA
Val_2 classC

Can I get those images into ImageDataBunch?

Is there a resource on ImageDataBunch?

0 Likes

(Dominik Engel) #4

https://docs.fast.ai/data_block.html
This is probably the part of the documentation you want to look at. It has lots of examples too.
If your training and validation images have a different naming scheme this is not going to work that well. If that’s the case I would actually suggest writing a small python script to rename the validation images similar to the training images… Or just make a csv for both train and valid. Both works nicely, however trying to retrieve labels differently from train and valid is not easily doable as far as I am aware.

1 Like