A Rust port of google-research/nasbench.

Of course, the primary purpose of this crate is to make NASBench dataset available in Rust. Besides, another aim is to reduce dataset loading time. By using a compact binary data format, this crate can reduce the loading time drastically. For example, on my laptop, google-research/nasbench requires about 200 seconds for loading the full dataset. By contrast, this crate only needs a few seconds to complete the loading.

First of all, you have to convert a NASBench dataset to this crate's format as follows:

$ wget https://storage.googleapis.com/nasbench/nasbench_full.tfrecord
$ nasbench convert nasbench_full.tfrecord nasbench_full.bin
$ ls -lh
-rw-rw-rw- 1 foo foo 328M May 12 16:47 nasbench_full.bin
-rw-rw-rw- 1 foo foo 2.0G May 12 16:45 nasbench_full.tfrecord

Then, you can query the evaluation metrics associated with a model (ops and adjacency):

$ nasbench query nasbench_full.bin \
                 --adjacency 0100110001000000010010000010000001000000010000000 \
                 --ops input conv3x3-bn-relu maxpool3x3 conv3x3-bn-relu \
                       conv3x3-bn-relu conv1x1-bn-relu output
EvaluationMetrics {
    training_time: 1769.1279296875,
    training_accuracy: 1.0,
    validation_accuracy: 0.9241786599159241,
    test_accuracy: 0.9211738705635071
}

Rust code corresponded to the above command:

use nasbench::{AdjacencyMatrix, ModelSpec, NasBench, Op};

// Loads the dataset.
let nasbench = NasBench::new("nasbench_full.bin")?;

// Queries a model.
let ops = vec![Op::Input, Op::Conv3x3, Op::MaxPool3x3, Op::Conv3x3,
               Op::Conv3x3, Op::Conv1x1, Op::Output];
let adjacency = "0100110001000000010010000010000001000000010000000".parse()?;
let model_spec = ModelSpec::new(ops, adjacency)?;
println!("{:?}", nasbench.models().get(&model_spec));

google-research/nasbench provides NASBench.evaluate() method to train and evaluate models from scratch, but this crate does not.

• NAS-Bench-101: Towards Reproducible Neural Architecture Search