Dataset
These models are trained on the 13k images dataset from the community images sourced 4/8/24. The 13k images had to have >2 reviewed labels with >50% community agreement. The dataset consists of prompts, tags, and images - although at this time the tags were neglected since there are considerable post-generation additions by the community and moderator, and this data would not be available at the time of scoring.
Training details
model: baseresNet18
- epochs: 39
- stopped due to early stopping
- train_test_split: .85/.15
- augmentations: None
model: baseresNet50
- epochs: 39
- stopped due to early stopping
- train_test_split: .85/.15
- augmentations: None
model: bestresNet50
- epochs: 25
- stopped due to max epochs
- train_test_split: .8/.2
- augmentations: RandomRotation, RandomHorizontalFlip, RandomResizedCrop
model: promptMovieBert
- epochs: 10
- test_train_split: .9/.1
model: promptMovieRoberta
- epochs: 10
- test_train_split: .9/.1