Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Generative Code Modeling with Graphs

This is the code required to reproduce experiments in two of our papers on modeling of programs, composed of three major components:


If you want to cite this work for the encoder part (i.e., our ICLR'18 paper), please use this bibtex entry:

  title={Learning to Represent Programs with Graphs},
  author={Allamanis, Miltiadis
          and Brockschmidt, Marc
          and Khademi, Mahmoud},
  booktitle={International Conference on Learning Representations (ICLR)},

If you want to cite this work for the generative model (i.e., our ICLR'19 paper), please use this bibtex entry:

  title={Generative Code Modeling with Graphs},
  author={Brockschmidt, Marc
          and Allamanis, Miltiadis
          and Gaunt, Alexander~L. 
          and Polozov, Oleksandr},
  booktitle={International Conference on Learning Representations (ICLR)},

Running the Code

The released code provides two components:

  • Data Extraction: A C# project extracting graphs and expressions from a corpus of C# projects. The sources for this are in DataExtraction/.
  • Modelling: A Python project learning model of expressions, conditionally on the program context. The sources for this are in Models/.

Note that the code is a research prototype; the documentation is generally incomplete and code quality is varying.

Data Extraction

Building the data extractor

To build the data extraction, you need a .NET development environment (i.e., a working dotnet executable). Once this is set up, you can build the extractor as follows:

DataExtraction$ dotnet build
    ExpressionDataExtractor -> ExpressionDataExtractor\bin\Debug\net472\ExpressionDataExtractor.exe

Build succeeded.

Using the data extractor

You can then use the resulting binary to extract contexts and expressions from a C# project:

DataExtraction$ ExpressionDataExtractor/bin/Debug/net472/ExpressionDataExtractor.exe TestProject outputs/{graphs,types}
Writing graphs at outputs/graphs
Writing type hierarchies at outputs/types
[11/01/2019 14:07:05] Starting building all solutions in TestProject
[11/01/2019 14:07:05] Restoring packages for: TestProject/TinyTest.sln
[11/01/2019 14:07:05] In dir TestProject running nuget restore TinyTest.sln -NonInteractive -source
[11/01/2019 14:07:05] Nuget restore took 0 minutes.
[11/01/2019 14:07:06] Starting build of TestProject/TinyTest.sln
Compilations completed, completing extraction jobs...
Opening output file outputs/graphs/exprs-graph.0.jsonl.gz.
[11/01/2019 14:07:09] Extracted 15 expressions from TestProject/Program.cs.

Now, outputs/graphs/exprs-graph.0.jsonl.gz will contain (15) samples consisting of a context graph and a target expression in tree form. ExpressionDataExtractor.exe --help provides some information on additional options.

Note: Building C# projects is often non-trivial (requiring NuGet and other libraries in the path, preparing the build by running helper scripts, etc.). Roughly, data extraction from a solution Project.sln will only succeed if running MSBuild Project.sln succeeds as well.

Extractor Structure

Data extraction is split into two projects:

  • ExpressionDataExtractor: This is the actual command-line utility with some code to find and build C# projects in a directory tree.

  • SourceGraphExtractionUtils: This project contains the actual extraction logic. Almost all of the interesting logic is in GraphDataExtractor, which is in dire need of a refactoring. This class does four complex things:

    • Identify target expressions to extract (SimpleExpressionIdentifier).
    • Turn expressions into a simplified version of the C# syntax tree (TurnIntoProductions). This is needed because Roslyn does not expose an /abstract/ syntax tree, but a /full/ ST with all surface code artifacts.
    • Construction of a Program Graph as in "Learning to Represent Programs with Graphs", ICLR'18 (ExtractSourceGraph).
    • Extraction of a subgraph of limited size around a target expression, removing the target expression in the process (CopySubgraphAroundHole).

    There is some bare-bones documentation for these components, but if you are trying to understand them and are stuck, open an issue with concrete questions and better documentation will magically appear.


First, run pip install -r requirements.txt to download the needed dependencies. Note that all code is written in Python 3.

As the preprocessing of graphs into tensorised form is relatively computationally expensive, we use a preprocessing step to do this. This computes vocabularies, the grammar required to produce the observed expressions and so on, and then transforms node labels from string form into tensorised form, etc.:

$ utils/ test_data/tensorised test_data/exprs-types.json.gz test_data/graphs/
Imputed grammar:
  Expression  -[00]->  ! Expression
  Expression  -[01]->  - Expression
  Expression  -[02]->  -- Expression
  Expression  -[03]->  CharLiteral
  Expression  -[04]->  Expression * Expression
  Expression  -[05]->  Expression + Expression
  Expression  -[06]->  Expression ++
  Expression  -[07]->  Expression . IndexOf ( Expression )
  Expression  -[08]->  Expression . IndexOf ( Expression , Expression , Expression )
  Expression  -[09]->  Expression . StartsWith ( Expression )
  Expression  -[10]->  Expression < Expression
  Expression  -[11]->  Expression > Expression
  Expression  -[12]->  Expression ? Expression : Expression
  Expression  -[13]->  Expression [ Expression ]
  Expression  -[14]->  IntLiteral
  Expression  -[15]->  StringLiteral
  Expression  -[16]->  Variable
Known literals:
  IntLiteral: ['%UNK%', '0', '1', '2', '4', '43']
  CharLiteral: ['%UNK%', "'-'"]
  StringLiteral: ['"foobar"', '%UNK%']
Tensorised 15 (15 before filtering) samples from 'test_data/graphs/' into 'test_data/tensorised/'.

If you want to use a given vocabulary/grammar (e.g., to prepare validation data), you can use the computed metadata from another folder:

$ utils/ --metadata-to-use test_data/tensorised/metadata.pkl.gz test_data/tensorised_valid test_data/exprs-types.json.gz test_data/graphs/
Tensorised 15 (15 before filtering) samples from 'test_data/graphs/' into 'test_data/tensorised_valid/'.


To test if everything works, training on a small number of examples should work:

% utils/ trained_models/overtrain test_data/tensorised/{,}
Starting training run NAG-2019-01-11-18-21-14 of model NAGModel with following hypers:
{"optimizer": "Adam", "seed": 0, "dropout_keep_rate": 0.9, "learning_rate": 0.00025, "learning_rate_decay": 0.98, "momentum": 0.85, "gradient_clip": 1, "max_epochs": 100, "patience": 5, "max_num_cg_nodes_in_batch": 100000, "excluded_cg_edge_types": [], "cg_add_subtoken_nodes": true, "cg_node_label_embedding_style": "Token", "cg_node_label_vocab_size": 10000, "cg_node_label_char_length": 16, "cg_node_label_embedding_size": 32, "cg_node_type_vocab_size": 54, "cg_node_type_max_num": 10, "cg_node_type_embedding_size": 32, "cg_ggnn_layer_timesteps": [3, 1, 3, 1], "cg_ggnn_residual_connections": {"1": [0], "3": [0, 1]}, "cg_ggnn_hidden_size": 64, "cg_ggnn_use_edge_bias": false, "cg_ggnn_use_edge_msg_avg_aggregation": false, "cg_ggnn_use_propagation_attention": false, "cg_ggnn_graph_rnn_activation": "tanh", "cg_ggnn_graph_rnn_cell": "GRU", "eg_token_vocab_size": 100, "eg_literal_vocab_size": 10, "eg_max_variable_choices": 10, "eg_propagation_substeps": 50, "eg_hidden_size": 64, "eg_edge_label_size": 16, "exclude_edge_types": [], "eg_graph_rnn_cell": "GRU", "eg_graph_rnn_activation": "tanh", "eg_use_edge_bias": false, "eg_use_vars_for_production_choice": true, "eg_update_last_variable_use_representation": true, "eg_use_literal_copying": true, "eg_use_context_attention": true, "eg_max_context_tokens": 500, "run_id": "NAG-2019-01-11-18-21-14"}
==== Epoch 0 ====
  Epoch 0 (train) took 0.26s [processed 58 samples/second]
 Training Loss: 10.622053
  Epoch 0 (valid) took 0.11s [processed 132 samples/second]
 Validation Loss: 9.516558
  Best result so far -- saving model as 'trained_models/overtrain/NAGModel_NAG-2019-01-11-18-21-14_model_best.pkl.gz'.
==== Epoch 100 ====
  Epoch 100 (train) took 0.22s [processed 69 samples/second]
 Training Loss: 0.650332
  Epoch 100 (valid) took 0.10s [processed 146 samples/second]
 Validation Loss: 0.637806

We can then evaluate the model:

$ utils/ trained_models/overtrain/NAGModel_NAG-2019-01-11-18-21-14_model_best.pkl.gz test_data/graphs/ trained_models/overtrain/test_results/
Groundtruth: b ? 1 : - i
  @1 Prob. 0.219: b ? 1 : i
  @2 Prob. 0.095: b ? 1 : - i
  @3 Prob. 0.066: b ++
  @4 Prob. 0.041: b ? 2 : i
  @5 Prob. 0.040: b ? 1 : arr
Num samples: 15 (15 before filtering)
Avg Sample Perplexity: 1.51
Std Sample Perplexity: 0.27
Accuracy@1: 60.0000%
Accuracy@5: 86.6667%

Model Variations

There are four different model types implemented:

  • NAG: The main model presented in "Generative Code Modeling with Graphs" (ICLR'19), representing the program context by a graph and using the graph-structured decoding strategy discussed in the paper.
  • seq2graph: An ablation that uses the graph-structured decoder, but represents the context using a sequence model. Concretely, a window of tokens around the hole to fill is fed into a two-layer BiGRU to obtain a representation for the program context. Additionally, for each variable in scope, a number of token windows around usages are encoded with a second BiGRU, and their representation is averaged.
  • graph2seq: An ablation that uses a graph to represent the program context, but then relies on a 2-layer GRU to construct the target expression.
  • seq2seq: An ablation using both a sequence encoder for the program context as well as a sequence decoder.

All models have a wide range of different hyperparameters. As these choices influence the format of tensorised data, both and need to be re-run for every variation:

$ utils/ test_data/tensorised_seq2graph test_data/exprs-types.json.gz test_data/graphs/ --model seq2graph --hypers-override '{"eg_hidden_size": 32, "cx_token_representation_size": 64}'
$ utils/ test_data/tensorised_seq2graph/{,} --model seq2graph --hypers-override '{"eg_hidden_size": 32, "cx_token_representation_size": 64}'

Model Structure

Roughly, the model code is split into three main components:

  • Infrastructure: The Model class (in exprsynth/ implements the usual general infrastructure bits; and models are expected to implement certain hooks in it. All of these are documented individually.
    • Saving and loading models, hyperparameters, training loop, etc.

    • Construction of metadata such as vocabularies: This code is parallelised and implementations need to extend three core methods (_init_metadata, _load_metadata_from_sample, _finalise_metadata) to use this code. Intuitively, init_metadata prepares a dict to store raw information (e.g., token counters) and _load_metadata_from_sample processes a single datapoint to update this raw data. These two are usually parallelised, and _finalise_metadata has to combine all raw metadata dictionaries to obtain one metadata dictionary, containing for example a vocabulary (in a MapReduce style).

    • Tensorising raw samples: _load_data_from_sample needs to be extended for this, and implementors can use the computed metadata.

    • Minibatch construction: We build minibatches by growing a batch until we reach a size limit (e.g., because we hit the maximal number of nodes per batch). This is implemented in the methods _init_minibatch (creating a new dictionary to hold data), _extend_minibatch_by_sample to add a new sample to a batch, and _finalise_minibatch, which can do final flattening operations and turn things into a feed dict.

      Note 1: This somewhat complicated strategy is required for two reasons. First, the sizes of graphs can vary substantially, and so picking a fixed number of graphs may yield a minibatch that is very small or large (in number of nodes). At the same time, our strategy of treating graph batches as one large graph requires regular shifting of node indices of samples, which is easiest to implement correctly in this incremental fashion.

      Note 2: In principle, these methods should be executed on another thread, so that a new minibatch can be constructed while the GPU is computing. Code for this exists, but was taken out for simplicity here.

  • Context Models: Two context models are implemented:
    • ContextGraphModel (in exprsynth/ This is the code implementing the modeling of a program graph, taking types, labels, dataflow, etc. into account. It produces a representation of all nodes in the input graphs as model.ops['cg_node_representations'], which can then be used in downstream models.
    • ContextTokenModel (in exprsynth/ This implements a simple BiGRU model over the tokens around the target expression, taking types into account. It produces a representation of all tokens in these contexts as model.ops['cx_hole_context_representations'].
  • Decoder Models: Two decoder models are implemented:
    • NAGDecoder (in exprsynth/ This is the code implementing the modeling of program generation as a graph.

      First, a representation of all nodes in the expansion graph is computed using scheduled message passing (implemented using AsyncGGNN at training time and step-wise use of get_node_attributes at test time). The schedule is determined by the __load_expansiongraph_training_data_from_sample method (and is the core of our paper).

      Second, a number of expansion decisions are made. The modeling of grammar productions is in __make_production_choice_logits_model, variables are chosen in __make_variable_choice_logits_model and literals are produced or copied in __make_literal_choice_logits_model.

    • SeqDecoder (in exprsynth/ A simple sequence decoder.

  • Glue code: Context models and decoders are combined using the actual models we instantiate. For example, NAGModel extends the ContextGraphModel and instantiates a NAGDecoder, contributing only some functionality to forward data from the encoder to the decoder.


This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit

When you submit a pull request, a CLA-bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact with any additional questions or comments.


Code for "Generative Code Modeling with Graphs" (ICLR'19)



Code of conduct


No releases published


No packages published