See Greenformer in action
Change the parameters and see the magic of Greenformer
Your typical modeling pipeline
Greenformer in your pipeline

from your_project import load_model, train, predict

model = load_model('transformer')

model = train(model)

score = predict(model)

from your_project import load_model, train, predict, get_submodules

from greenformer import auto_fact

model = load_model('transformer')

model = auto_fact(module=model, rank=64, solver='random', num_iter=50, submodules=None)

model = train(model)

submodules = get_submodules(model,'bert-base')

model = auto_fact(module=model, rank=64, solver='svd', num_iter=50, submodules=None)

score = predict(model)

Try Greenformer for Factorization-by-design on Google Colab
Accuracy
80.25%
GPU time (ms)
50s
CPU time (ms)
55s
Memory (MB)
100
Parameters (mio)
43
Accuracy
77.94%
GPU time (ms)
40s
CPU time (ms)
45s
Memory (MB)
100
Parameters (mio)
43
Next stop: Greenformer on Google Colab

Faster and memory efficient training without sacrificing performance

Factorization-by-design

Productionize a pre-trained model in a faster, lighter, and cheaper way

Post-training

More efficient few-shot in-context learning using billions parameters models

In-context learning
More about Greenformer
Greenformer is flexible, easy-to-use, and applicable for multiple scenarios.
Gain up to 1.5x speed increase while maintaining 93% of the performance by utilising Greenformer.
Performance and efficiency trade-off
Performance and efficiency trade-off of utilizing Greenformer on:
Left: factorization-by-design, Center: post-training factorization, Right: in-context learning factorization
Greenformer performs decomposition to the weight of linear and convolution layers within the models utilizing factorization solvers like random, SVD, or SMNF. It replaces: 1) the linear layers with Linear Encoder-Decoder (LED) layers and 2) the convolution layers with Convolution Encoder-Decoder (CED) layers. Greenformer supports factorization with either a predefined static rank or dynamic ranks across all layers by computing the rank based on a ratio to the maximum rank of the corresponding layer.
LED and CED layers in action
Left: Low-rank Encoder-Decoder (LED) layer, Right: Convolution Encoder-Decoder (CED) layer.
GitHub
Visit Greenformer repository here.
PyPi
Install Greenformer from PyPi.
Cite us
You can find our paper here. If you are using any component of Greenformer for research purposes, please cite the following paper:
@misc{cahyawijaya2021greenformer,
    title={Greenformer: Factorization Toolkit for Efficient Deep Neural Networks},
    author={Samuel Cahyawijaya and Genta Indra Winata and Holy Lovenia and Bryan Wilie and Wenliang Dai and Etsuko Ishii and Pascale Fung},
    year={2021},
    eprint={2109.06762},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}