https://github.com/kyegomez/simplifiedtransformers
SimplifiedTransformer simplifies transformer block without affecting training. Skip connections, projection parameters, sequential sub-blocks, and normalization layers are removed. Experimental results confirm similar training speed and performance.
https://github.com/kyegomez/simplifiedtransformers
artificial-intelligence gpt4 machine-learning machine-learning-algorithms neural-architecture-search neural-networks swarms
Last synced: 5 months ago
JSON representation
SimplifiedTransformer simplifies transformer block without affecting training. Skip connections, projection parameters, sequential sub-blocks, and normalization layers are removed. Experimental results confirm similar training speed and performance.
- Host: GitHub
- URL: https://github.com/kyegomez/simplifiedtransformers
- Owner: kyegomez
- License: mit
- Created: 2023-12-02T18:54:59.000Z (almost 2 years ago)
- Default Branch: main
- Last Pushed: 2025-04-11T12:59:05.000Z (6 months ago)
- Last Synced: 2025-04-15T11:46:46.284Z (6 months ago)
- Topics: artificial-intelligence, gpt4, machine-learning, machine-learning-algorithms, neural-architecture-search, neural-networks, swarms
- Language: Python
- Homepage: https://discord.gg/GYbXvDGevY
- Size: 2.16 MB
- Stars: 14
- Watchers: 2
- Forks: 3
- Open Issues: 2
-
Metadata Files:
- Readme: README.md
- Funding: .github/FUNDING.yml
- License: LICENSE
Awesome Lists containing this project
README
[](https://discord.gg/qUtxnK2NMf)
# SimplifiedTransformers
The author presents an implementation for Simplifying Transformer Blocks. The standard transformer blocks are complex and can lead to architecture instability. In this work, the author investigates how the standard transformer block can be simplified. Through signal propagation theory and empirical observations, the author proposes modifications that remove several components without sacrificing training speed or performance. The simplified transformers achieve the same training speed and performance as standard transformers, while being 15% faster in training throughput and using 15% fewer parameters.
# Install
```
pip3 install --upgrade simplified-transormer-torch
```
--------
## Usage
```python
import torch
from simplified_transformers.main import SimplifiedTransformers
model = SimplifiedTransformers(
dim=4096,
depth=6,
heads=8,
num_tokens=20000,
)
x = torch.randint(0, 20000, (1, 4096))
out = model(x)
print(out.shape)
```
# License
MIT