Skip to content

bingli-borland/Transformer-from-scratch

 
 

Repository files navigation

Transformer from scratch

This is a Transformer based Large Language Model (LLM) training demo with only ~240 lines of code.

Inspired by nanoGPT, I wrote this demo to show how to train a LLM from scratch using PyTorch. The code is very simple and easy to understand. It's a good start point for beginners to learn how to train a LLM.

The demo is trained on a 450Kb sample textbook dataset, and the model size is about 51M. I trained on a single i7 CPU, and the training time takes about 20 minutes, result in approximately ~1.3M parameters.

Get Started

  1. Install dependencies
pip install numpy requests torch tiktoken
  1. Run model.py

First time when you run it, the program will download the dataset and save to data folder. Then the model will start training on the dataset. Training & validation losses will be printed on the console screen, something like:

Step: 0 Training Loss: 11.68 Validation Loss: 11.681
Step: 20 Training Loss: 10.322 Validation Loss: 10.287
Step: 40 Training Loss: 8.689 Validation Loss: 8.783
Step: 60 Training Loss: 7.198 Validation Loss: 7.617
Step: 80 Training Loss: 6.795 Validation Loss: 7.353
Step: 100 Training Loss: 6.598 Validation Loss: 6.789
...

The training loss will decrease as the training goes on. After 5000 iterations, the training will stop and the losses are down to around 2.807. The model will be saved under name model-ckpt.pt.

Then a sample text will be generated and pop to the console screen from the model we just trained, something like:

The salesperson to identify the other cost savings interaction towards a nextProps audience, and interactive relationships with them. Creating a genuine curiosityouraging a persuasive knowledge, focus on the customer's strengths and responding, as a friendly and thoroughly authority. 
Encouraging open communication style to customers that their values in the customer's individual finding the conversation.2. Addressing a harmoning ConcernBIG: Giving and demeanor is another vital aspect of practicing a successful sales interaction. By sharing case studies, addressing any this compromising clearly, pis

It looks pretty descent!

Feel free to change some of the hyperparameters on the top of the model.py file, and see how it affects the training process.

  1. Step-by-step Jupyter Notebook

I also provide a step-by-step Jupyter Notebook step-by-step.ipynb to help you understand the architecture logic. To run this, you also need to insall:

pip install matplotlib pandas

This notebook prints out the intermediate results of each step followed by Transformer architecture from original paper, but only the Decoder part (Since GPT only use the decoder). So you can see how the model is trained each single step. For examples:

  • what a [4,16] matrix of input embedding looks like:
      0     1      2      3     4      5      6      7      8      9      10     11     12     13     14     15
0    627  1383  88861    279  1989    315  25607  16940  65931    323  32097     11    584  26458  13520    449
1  15749   311   9615   3619   872   6444      6   3966     11  10742     11    323  32097     13   3296  22815
2  13189   315   1701   5557   304   6763    374  88861   7528  10758   7526     13   4314   7526   2997   2613
3    323  6376   2867  26470  1603  16661    264  49148    627     18     13  81745  48023  75311   7246  66044
  • the positional encoding plot of the input sequence:

  • the attention matrix of the first Q * K layer:

  • after applying Mask attention of the above matrix:

Other contents in this repo

Under /GPT2 directory, I put some sample code to show how to fine-tune a pre-trained GPT2 model, as well as inference from it.

If you want to dive deeper

As if you're new to LLM, I recommend you to read my blog post Transformer Architecture: LLM From Zero-to-Hero , which breaks down the concepts of a Transformer architecture.

References

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 80.6%
  • Python 19.4%