Update README.md

README.md

@@ -11,13 +11,10 @@ pipeline_tag: image-classification
 
 An advanced deep learning model for automatically tagging anime/manga illustrations with relevant tags across multiple categories, achieving **58.1% micro F1 score** (31.5% macro F1 score using the balanced threshold preset) across 70,527 possible tags on a test set of 20,116 samples. Trained on a single 3060.
 
-## 🚀 Updates (March 2025)
+## 🚀 Updates (April-May 2025)
+Apologies for the lack of updates. I've been busy with Uni and unfortunately the extra training (1 epoch) I made the mistake of only loading the model weights and not the optimizer state as well. This is important because the optimizer state maintains a tuned learning rate for each parameter. The reason I had loaded weights only was because training time was much slower so I overlooked the importance of loading the optimizer state.
 
-- **ONNX Model**: Added ONNX export support for better cross-platform compatibility, deployment options and inference speed
-- **Save to TXT**: New feature to save tag predictions directly to text files
-- **Batch Inference**: Added support for processing multiple images in a single run
-- **Tagging Game**: Introducing a gamified tagging experience - scan images, discover tags, earn TagCoins, and generate essence!
-- **Training Progress**: The model shows clear signs of being undertrained, with significant room for improvement (see Performance Notes below)
+The good news is that I should soon be getting an RTX 3090. This means that I should be able to train the entire model in the GPU VRAM (no offloading) and take advantage of the 3x memory bandwidth the 3090 provides. This should mean massive speedups for training time!
 
 ## 🔑 Key Highlights
 
@@ -66,7 +63,7 @@ The significant improvement in Macro-F1 (+4% per epoch) suggests that longer tra
 
 I plan to continue training the model to further improve performance, especially for rare tags. However, each epoch takes approximately 1.5-2 weeks of overnight training on my current hardware.
 
-**If you'd like to support further training on the complete dataset or my future projects, consider [buying me a coffee](https://www.buymeacoffee.com/camais). Your support will directly enable longer training runs and better models!**
+**If you'd like to support further training on the complete dataset or my future projects, consider supporting me here:(https://ko-fi.com/camais). Your support will directly enable longer training runs and better models!**
 
 After this project, I plan to move onto LLMs as I have lots of ideas on how to improve upon them. I will update this model based on community attention.