tarekziade/distilvit

This model is a variation of https://huggingface.co/nlpconnect/vit-gpt2-image-captioning

• Read the blog post here https://ziade.org/2024/03/17/distilvit-image-captioning-model
• The training code is here: https://github.com/tarekziade/distilvit

Results after after 3 epochs (and ~45 hours of training)

• eval_loss: 0.19939416646957397
• eval_rouge1: 43.006
• eval_rouge2: 16.9939
• eval_rougeL: 38.8923
• eval_rougeLsum: 38.8877
• eval_gen_len: 11.327256736227712
• eval_runtime: 1816.5255
• eval_samples_per_second: 13.77
• eval_steps_per_second': 1.721
• train_runtime: 46263.3695
• train_samples_per_second: 38.373
• train_steps_per_second: 4.797
• train_loss: 0.05974134062104816