isl-img2text

Author: Sigurdur Haukur Birgisson

This model is a fine-tuned version of microsoft/git-base on Sigurdur/isl-image-captioning. It achieves the following results on the evaluation set:

• eval_loss: 0.0983
• eval_wer_score: 0.7295
• eval_runtime: 20.5346
• eval_samples_per_second: 7.792
• eval_steps_per_second: 0.974
• epoch: 15.0
• step: 150

It appears that the model heavilly overfitted to the dataset. Also, something I failed to consider was that the base model can't write any Icelandic characters and was thus not suited for this task. Future works might want to add the capability of writing Icelandic characters to the model.

Model description

More information needed

Intended uses & limitations

Image captioning in Icelandic

Training and evaluation data

Scraped images and their descriptions/captions from the Icelandic wikipedia.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 50
• mixed_precision_training: Native AMP

Metrics

Epoch	Training Loss	Validation Loss	Wer Score
1	10.096300	8.690205	102.247536
2	8.268200	7.655295	97.659365
3	7.298000	6.679112	95.714129
4	6.319800	5.673368	2.136911
5	5.317500	4.656871	22.439211
6	4.315600	3.667494	1.001095
7	3.340000	2.722741	1.063527
8	2.417700	1.852253	0.944140
9	1.593900	1.136962	0.949617
10	0.944900	0.638581	0.933187
11	0.516200	0.355187	0.955093
12	0.281600	0.215951	0.822563
13	0.167500	0.148763	0.773275
14	0.111700	0.116783	0.792990
15	0.080800	0.098261	0.729463

Framework versions

• Transformers 4.42.3
• Pytorch 2.0.1
• Datasets 2.20.0
• Tokenizers 0.19.1