Switch to BERTIN model for training script and testing on image

Use BERTIN model

Keep only best performing model in main (all models still available in develop branch). Best performing model is clip_spanish_141230_samples with a loss of 2.231235980987549

Add model using BERTIN language encoder trained on 146364 examples and obtaining a loss of 2.231235980987549

Refactor test_on_image.py

Update README and add a training doc

Add actual files instead of symlinks

Copy latest model to repo root

Add model trained on 141230 samples, achieving a validation loss of 2.1384739875793457

Add all necessary files to replicate training run

Add model trained on 72972 samples

Add new model trained on the spanish subset of suitable images of the 20% of the WIT dataset, using a 998/1/1 train/valid/test split with a validation loss of 22.3439

Add new model trained on the spanish subset of suitable images of the 20% of the WIT dataset, with a validation loss of 2.4001

Use 1 caption per image

Add README

Update downloading and training scripts

Resolve merge conflict in .gitattributes

Add training scripts and initial model trained on 1% of the data.

Update .gitattributes

initial commit