deep-learning-stack-overflow

Dataset and replication package for the empirical study of programming challenges of building deep learning applications (ISSRE 2019)

Usage

Run the code with

$ python run_model.py

You can change the configures in run_model.py, including the data_path, log_dir, max_utt_len, etc. The important parameters are

add_manual -- defines the weights of manual keywords in class prediction
top_keyword -- means the number of top words to consider (ranked by tf-idf)
only_title -- means whether or not involve post content or only title
use_major_label -- indicates whether use the first label or all the labels for a post
lemm -- whether or not do lemmatization
remove_other -- whether remove the posts labeled with "Other" category

Dataset

All the data used in this study can be downloaded from this link. You can put the folder in the master directory, i.e., ../data. The dl_framework_stackoverflow_questions file contains all the 39,628 studied deap-learning-related questions on Stack Overflow. The Manual_Inspection.xlsx file iincludes the labeled 715 questions on three deep learning platforms. The meta-data.csv file saves the meta data for each question, such as the response time. The manually labeled keywords are listed in ../data/sto/keyword.csv. Other files are generated files during running.

Citation

Please cite our ISSRE paper if you use our method / datasets in your work

@inproceedings{zgmlk2019stoissre,
    title={An empirical study of common challenges in developing deep learning applications},
    author={Tianyi Zhang and Cuiyun Gao and Lei Ma and Michael R. Lyu and Miryung Kim},
    booktitle={Proceedings of the 30th International Symposium on Software Reliability Engineering, {ISSRE} 2019},
    year={2019}
}