Since the creation of this repository, the world of AI has been evolving at full speed. As such, this repository has been continually updated and upgraded. This changelog contains the notable updates to the Transformers for NLP and Computer Vision 3rd Edition repository to adapt to the fast-evolving Generative AI market after October 30, 2024.
Before October 30, 2024, the repository was also continually updated. You can rely on the icons in front of the notebooks in the README file to find the evolutions:
Look for 🐬 to explore new bonus notebooks such as OpenAI o1's reasoning models, Midjourney's API, Google Vertex AI Gemini's API, OpenAI asynchronous batch API calls!
Look for 🎏 to explore existing notebooks for the latest model or platform releases, such as OpenAI's latest GPT-4o and GPT-4o-mini models.
Look for 🛠 to run existing notebooks with new dependency versions and platform API constraints and tweaks.
Chapter06/KantaiBERT.ipynb and Chapter06/Customer_Support_for_X.ipynb enhancement:
wandb.ai training metrics added while training.
You can now view your model's metrics on Hugging Face.
Chapter14/Google_Vertex_AI.ipynb has now been upgraded to Chapter14/Google_Vertex_AI_Gemini.ipynb with the gemini-1.5-flash-001 model.
A timer has been added to adapt to the API call rates.
It is recommended to read Google_Vertex_AI.ipynb with the chapter in the book to understand the process but then to run Google_Vertex_AI_Gemini.ipynb which now calls 20 LLM tasks and a computer vision task.
Open and Read Chapter17/Stable_Diffusion_Keras.ipynb with the book to understand the process. However, this notebook has been replaced by
Chapter17/Stable_Diffusion_Hugging_Face.ipynb.
Reason for Change: The initial implementation using keras_cv for Stable Diffusion was unstable due to compatibility issues between Keras and TensorFlow. Specifically, there were multiple shape mismatch errors, difficulties in managing precision (mixed vs float32), and challenges in manually encoding prompts due to limitations in keras_cv's API. These issues led to unreliable behavior and a frustrating development experience in Google Colab.
New Approach: We transitioned to using Hugging Face's Diffusers library for Stable Diffusion because it provides a more stable and intuitive workflow for generating images from text prompts. The diffusers library integrates well with PyTorch and has comprehensive support for text-to-image generation. It allowed us to seamlessly encode prompts and use GPU acceleration without running into the issues present in the Keras implementation.
Benefits:
- Improved compatibility and stability.
- Easier to manage the text encoding and image generation workflow.
- Robust GPU utilization through PyTorch, leading to faster and more reliable results.
- Clearer API with better community support for addressing issues.
This change ensures a smoother experience when working with Stable Diffusion in Google Colab, resulting in faster image generation and more predictable outputs.
Diffusion models have evolved tremendously in a very short time, from the ability to create animations to generating videos and more. The code in this Chapter17/Stable__Vision_Stability_AI_Animation.ipynb notebook contains a Stability SDK, which is interesting to read along with the book.
Then, you can access (open source and free, of course!) the recent developments of text-to-video technology in the RAG-Driven Generative AI repo starting with the first notebook of Chapter 10, RAG for Video Stock Production with Pinecone and OpenAI
Fixed typo in Chapter04/WMT_translations.ipynb:
Now, "smoothing" is correctly written in one of the print instructions.
Chapter 16/ViT_CLIP.ipynb:
The Feature extractor simulator section automatically displays all the patches to give a full view of feature extraction.