This dataset contains dataflow computational graphs generated procedurally, intended for training and evaluating algorithms that optimize execution (e.g. placement and scheduling), in TensorFlow's CostGraphDef protocol buffer format and encoded as text.

Original paper REGAL: Transfer Learning For Fast Optimization of Computation Graphs (Paliwal, Gimeno, Nair, Li, Lubin, Kohli, Vinyals)

There are 10000 training graphs, 1000 validation graphs and 1000 test graphs. The file names follow the format of "graph_" plus a hash of the graph topology plus ".pbtxt".

For each set (train, valid, test) there are not two graphs with the same topology. We used the Biased Random Key Genetic Algorithm (BRKGA) to filter out graphs that did not have "room for improvement"

"Room for improvement" was defined as the union of two conditions:

• if BRKGA with a low fitness evaluation limit (number of calls to fitness function) did not fit the hardware constraints and BRKGA with a high number did.
• if BRKGA with a high fitness evaluation limit was 20% faster in running time that BRKGA with a low number.

node {
  name: "_SOURCE"
}
node {
  name: "node_0"
  id: 1
  control_input: 0
}
node {
  name: "node_1"
  id: 2
  output_info {
    size: 70
    alias_input_port: -1
  }
  control_input: 0
  compute_cost: 58
}
node {
  name: "node_2"
  id: 3
  output_info {
    size: 52
    alias_input_port: -1
  }
  control_input: 0
  compute_cost: 47
}
node {
  name: "node_3"
  id: 4
  input_info {
    preceding_node: 2
  }
  output_info {
    size: 55
    alias_input_port: -1
  }
  control_input: 0
  compute_cost: 58
}

The dataset is available in the following link

The following table is necessary for this dataset to be indexed by search engines such as Google Dataset Search.

This is not an officially supported Google product.