Augmentation methods for Image Captioning: Text, image, and joint.
Based on EDA synonym replacement (with the help of WordNet).
Translation from and to Spanish & Arabic. Needed model: argostranslate
- Create a directory
argostranslate
. - Download the models specified in
backtranslate.py
.
T5-powered paraphrasing with a finetuned model from Huggingface: tuner007/pegasus_paraphrase
Albumentations
is leveraged for image augmentation. The pipeline contains the following transformations:
- CLAHE
- RandomRotate90
- Transpose
- ShiftScaleRotate
- Blur
- OpticalDistortion
- GridDistortion
- HueSaturationValue
- HorizontalFlip.