I'd say these techniques will allow you to choose cheaper hardware, which will essentially reduce your cloud server costs.
@rasbt
-
Fabric’s Intuitive API for PyTorch Script Modifications
By
–
I can't say much about accelerate since I never really used it, but regarding Fabric, I like the pretty intuitive API. It's basically just a minimal number of lines of code to modify a PyTorch script.
-
Vision Transformers: Tokenizing Images Like Language Models
By
–
In a tweet: similar to LLMs (language transformers) but tokenizing images instead of text:
-
Modern Classification: Multiple Output Nodes vs Sigmoid Threshold
By
–
So, in this case, rather than using a threshold on a logistic sigmoid score [0, 1], in most modern architectures, you usually define at least 2 output nodes. Then, you determine the class based on which output node has the highest score.
-
Memory-Saving Techniques for Tabular Data Transformers
By
–
but yeah, I am training classifiers in both cases, the same memory-saving techniques could potentially be applied to the recent transformers for tabular data.
-
Vision Transformers for Images and LLMs for Text
By
–
Oh this post is about vision transformers for images and LLMs for text
-
Reducing GPU Memory for LLMs and Vision Transformers in PyTorch
By
–
One of the big bottlenecks with LLMs & Vision Transformers is GPU memory on consumer devices. Wrote about my favorite techniques for reducing peak memory in PyTorch: https://
lightning.ai/pages/communit
y/tutorial/pytorch-memory-vit-llm/
… Focused on techniques that don't require architecture changes! Suggestions welcome! -
ChatGPT’s Knowledge Cutoff Limitation and Synthetic Data Challenges
By
–
3/3 Intuitively one major criticism of ChatGPT is it's 2020 knowledge cut-off. If you are only using synthetic data from old models, it'd be challenging to get more contemporary knowledge into the model.
-
Synthetic Data Generation Limitations in Predictive Model Improvement
By
–
2/2 I think there's been a lot of research on generating synthetic data to boost predictive models (incl. classifiers). The concern is that you'll never be able to generate anything "new" by recycling models (trained on a fixed dataset) that generate synthetic data.
-
Synthetic Data: Predictive Modeling vs Generative Modeling Distinction
By
–
I haven't heard of it yet, but it looks like it's related to generating medical images for image segmentation tasks. Looks like a solid approach. I think a crucial difference though is to distinguish between synthetic data for predictive modeling vs generative modeling 1/2.