|
--- |
|
datasets: |
|
- competitions/aiornot |
|
language: |
|
- en |
|
metrics: |
|
- accuracy |
|
pipeline_tag: image-classification |
|
--- |
|
|
|
# Model Card for Model Soups on AirorNot Dataset |
|
|
|
## Model Details |
|
|
|
### Model Description |
|
|
|
Code implementation of the paper [Model soups: averaging weights of multiple fine-tuned models |
|
improves accuracy without increasing inference time](https://arxiv.org/abs/2203.05482). |
|
|
|
In recent years, finetuning large models has been proving to be an excellent strategy to achieve high-performances in downstream tasks. |
|
The conventional recipe to do so, it's to fine-tune models with different hyperparameters and select the one achieving the highest accuracy. However Wortsman *et. al* proved that averaging the weights of multiple models finetuned with different hyperparameter configurations can actually inprove accuracy and robustness. |
|
|
|
I read this paper recently and I felt intrigued by the powerful yet simple idea (achieving a SOTA result on Imagenet of s 90.94%) so I decided that this could be an opportunity to get my hands dirty and dive into the code and...try the soup! |
|
|
|
I started by using the official [code implementation](https://github.com/mlfoundations/model-soups) with CLIP ViT-B/32 and finetuned only 5 of their models on AiorNot. I used a simple strategy with minimal modifications. Mainly, I finetuned the models for 8 epochs with a batch size of 56 samples. |
|
|
|
The tricky part was that I had to modify the baseline to use it with our custom dataset. |
|
|
|
|
|
- **Developed by:** HuggingSara |
|
- **Model type:** Computer Vision |
|
- **Language :** Python |
|
|
|
|
|
|
|
|
