Vislab provides support for several vision datasets out of the box.
- VOC2012, ~10K images tagged with 20 object classes. Multi-label.
- ~250K images with aesthetic ratings
- ~20K images from AVA that also have style labels
- ~50K images with style labels
- ~100K images with style, genre, artist labels
Adding your own
Adding your own dataset is super simple.
- Add the file
vislab/datasets/your_dataset.py that will contain a function to load a
- unique string-based index, with name
image_filename in columns
- a column for whatever boolean label you care about
vislab/dataset.py to map a name to your new function.
Dataset for classification
The classifier expects DataFrames with only two columns: ‘label’ and ‘importance’.
The ‘label’ column can contain:
- real values (regression)
- -1/1 (binary classification)
- or positive ints (multiclass classification).