A Convolutional Attention Network for Extreme Summarization of Source Code
Authors: Miltiadis Allamanis, Hao Peng, Charles Sutton
ICML 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate our convolutional attention neural network s performance on 10 popular Java projects showing that it achieves better performance compared to previous attentional mechanisms.Table 1 shows the F1 scores achieved by the different methods for each project while Table 2 shows a quantitative evaluation, averaged across all projects. |
| Researcher Affiliation | Academia | Miltiadis Allamanis M.ALLAMANIS@ED.AC.UK School of Informatics, University of Edinburgh, Edinburgh, EH8 9AB, United Kingdom Hao Peng PENGHAO.PKU@GMAIL.COM School of Electronics Engineering and Computer Science, Peking University, Beijing 100871, China Charles Sutton CSUTTON@INF.ED.AC.UK School of Informatics, University of Edinburgh, Edinburgh, EH8 9AB, United Kingdom |
| Pseudocode | Yes | attention_features (code tokens c, context ht 1) C LOOKUPANDPAD(c, E) L1 RELU(CONV1D(C, Kl1)) L2 CONV1D(L1, Kl2) ht 1 Lfeat L2/ L2 2 return Lfeat and attention_weights (attention features Lfeat, kernel K) return SOFTMAX(CONV1D(Lfeat, K)) |
| Open Source Code | Yes | The dataset and code can be found at groups.inf.ed.ac.uk/cup/codeattention. |
| Open Datasets | Yes | The dataset and code can be found at groups.inf.ed.ac.uk/cup/codeattention. and To collect a good dataset of good quality, we cloned 11 open source Java projects from Git Hub. |
| Dataset Splits | Yes | For each project, we split the files (top-level Java classes) uniformly at random into training (65%), validation (5%) and test (30%) sets. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU, GPU models, or memory specifications) used for running its experiments. |
| Software Dependencies | No | The paper mentions machine learning techniques and architectures like RMSProp, Nesterov momentum, dropout, and GRU, but does not specify the versions of any software libraries or frameworks used (e.g., Python, TensorFlow, PyTorch versions). |
| Experiment Setup | Yes | For conv_attention the optimized hyperparameters are k1 = k2 = 8, w1 = 24, w2 = 29, w3 = 10, dropout rate 50% and D = 128. For copy_attention the optimized hyperparameters are k1 = 32, k2 = 16, w1 = 18, w2 = 19, w3 = 2, dropout rate 40% and D = 128. |