BasStein's picture
Set `library_name` to `tf-keras`. (#1)
4727d14 verified
metadata
datasets:
  - BasStein/250000-randomfunctions-10d
language:
  - en
library_name: tf-keras
license: apache-2.0
metrics:
  - mse
tags:
  - doe2vec
  - exploratory-landscape-analysis
  - autoencoders
co2_eq_emissions:
  emissions: 0.0363
  source: code carbon
  training_type: pre-training
  geographical_location: Leiden, The Netherlands
  hardware_used: 1 Tesla T4

Model description

DoE2Vec model that can transform any design of experiments (function landscape) to a feature vector.
For different input dimensions or sample size you require a different model.
Each model name is build up like doe2vec-d{dimension}-m{sample size}-ls{latent size}-{AE or VAE}-kl{Kl loss weight}

Example code of loading this huggingface model using the doe2vec package.

First install the package

pip install doe2vec

Then import and load the model.

from doe2vec import doe_model

obj = doe_model(
    10,
    8,
    latent_dim=24,
    kl_weight=0.001,
    model_type="VAE"
)
obj.load_from_huggingface()
#test the model
obj.plot_label_clusters_bbob()

Intended uses & limitations

The model is intended to be used to generate feature representations for optimization function landscapes. The representations can then be used for downstream tasks such as automatic optimization pipelines and meta-learning.

Training procedure

The model is trained using a weighed KL loss and mean squared error reconstruction loss. The model is trained using 250.000 randomly generated functions (see the dataset) over 100 epochs.

  • Hardware: 1x Tesla T4 GPU
  • Optimizer: Adam