diff --git a/README.md b/README.md index 62070e8b3ff6d9b18fed46aae781cb1e4e96a2a5..d68382d9ffbc9ad14b292a1eec3141542296f6c5 100644 --- a/README.md +++ b/README.md @@ -23,10 +23,10 @@ The code for fine-tuning a classification model is based on the <a href="https:/ ## Dataset -The WikiArt dataset used for fine-tuning could be found <a href="https://huggingface.co/datasets/huggan/wikiart">here</a> +The WikiArt dataset used for fine-tuning could be found <a href="https://huggingface.co/datasets/huggan/wikiart">here</a>. Before fine-tunibg, dataset should be downloaded and filtered for excluding 'Unknown genre' class for genre classification fine-tuning and for selecting artists from the list of most popular artists for artist classification fine-tuning. For the style classification fine-tuning original dataset could be used. ## Steps -To reproduce the steps first you need to finetune models for genre, style and artist classification on WikiArt dataset or used fine-tuned adapters. Then, using fine-tuned models you need to calculate embeddings for the images from WikiArt dataset and create ANNOY indexes. Due to the relatively big size, these files are available upon a request. +To reproduce the steps first you need to finetune models for genre, style and artist classification on downloaded WikiArt dataset or use fine-tuned adapters (folder LoRA adapters). Then, using fine-tuned models you need to calculate embeddings for the images from WikiArt dataset and create ANNOY indexes for each fine-tuned model. Due to the relatively big size, these files (embeddings and ANNOY indexes) are available upon a request. ## Citation ```