bknyaz/deepnets1m

This is a copy of the DeepNets-1M dataset originally released at https://github.com/facebookresearch/ppuda under the MIT license.

The dataset presents diverse computational graphs (1M training and 1402 evaluation) of neural network architectures used in image classification. See detailed description in the Parameter Prediction for Unseen Deep Architectures paper.

There are four files in this dataset:

• deepnets1m_eval.hdf5; # 16 MB (md5: 1f5641329271583ad068f43e1521517e)
• deepnets1m_meta.tar.gz; # 35 MB (md5: a42b6f513da6bbe493fc16a30d6d4e3e), run tar -xf deepnets1m_meta.tar.gz to unpack it before running any code reading the dataset
• deepnets1m_search.hdf5; # 1.3 GB (md5: 0a93f4b4e3b729ea71eb383f78ea9b53)
• deepnets1m_train.hdf5; # 10.3 GB (md5: 90bbe84bb1da0d76cdc06d5ff84fa23d)

GHN-2

• paper: Parameter Prediction for Unseen Deep Architectures
• code: https://github.com/facebookresearch/ppuda
• models: https://github.com/facebookresearch/ppuda/tree/main/checkpoints

GHN-3

• paper: Can We Scale Transformers to Predict Parameters of Diverse ImageNet Models? training an improved GHN on this dataset
• code: https://github.com/SamsungSAILMontreal/ghn3
• models: https://huggingface.co/SamsungSAILMontreal/ghn3

Citation

If you use this dataset, please cite it as:

@inproceedings{knyazev2021parameter,
  title={Parameter Prediction for Unseen Deep Architectures},
  author={Knyazev, Boris and Drozdzal, Michal and Taylor, Graham W and Romero-Soriano, Adriana},
  booktitle={Advances in Neural Information Processing Systems},
  year={2021}  
}