joennlae / tensorli

Absolute minimalistic implementation of a GPT-like transformer using only numpy (<650 lines).

Minimalistic implementation of a GPT-like transformer using only numpy (<650 lines).

---

The implementation includes:

• Automatic differentiation
• Tensorli object (PyTorch like)
• Simple NN layers: Linearli, Embeddingli, MultiheadAttentionli, LayerNorm
• Optimizers: Adamli

All that is "needed" to train and execute a GPT-like transformer model.

...and everything else is just efficiency ~ Andrej Karpathy¹.

1: Youtube - micrograd

Example

from tensorli.tensorli import Tensorli
from tensorli.models.transformerli import Transformerli

vocb_size, embd_dim, seq_len, n_layer, n_head, batch_size = 10, 64, 10, 3, 4, 16

transformer = Transformerli(vocb_size, embd_dim, seq_len, n_layer, n_head)

x_numpy = np.random.randint(0, vocb_size - 1, size=(batch_size, seq_len))
x = Tensorli(x_numpy)

out = transformer(x)

Installation & Development

# create conda env (if you want to do it faster use mamba)
# conda env create -f environment.yml
# (if you want to do it faster use mamba)
mamba env create -f environment.yml
# activate env
conda activate tensorli
# python path need to be set to the root of the project
export PYTHONPATH=$PWD
# run all tests
pytest
# run specific test (verbose)
pytest -v -rP -k "transformer"

Naming

In the region where I grew up, a word for "little" is used as a suffix [2]. For example, "little dog" would be "dogli". I thought it would be a nice name for a minimalistic implementation of a neural network library.

Caveats

This library works, but it is NOT optimized. It is not meant to be used in production or for anything at scale. It is meant to be used as a learning tool.

Inspiration

This library is heavily inspired by the following projects:

• minGPT
• tinygrad

I highly recommend checking them out.

Training loss plot

Running the test script test/test_adder.py will produce the following loss curve:

Plans & Outlook

• Add Dropout
• Add more experimental architectures