Identifying Statistical Bias in Dataset Replication

Code for the paper "Identifying Statistical Bias in Dataset Replication." The original paper is here (blog post) .

Annotation Data

Annotation data not included in the Git repository is necessary to run any analyses (approx 2 GB). The annotation data can be found here. There are two Pandas dataframes saved with pytorch (e.g. load them with torch.load(path)) that correspond to both the raw annotation data and the data cleaned version (data cleaning described in paper). The is also a dataframe containing raw data from the original ImageNet-v2 study, which is required by run_orig_data.py.

Beta-Binomial fitting

To adjust for accuracy using parametric modelling (cf. Section 5.3), run run_betabinom.py while inputting the path of a dataframe retrieved from the "Annotation Data" section above to --df-path.

usage: run_betabinom.py [-h] --out-dir OUT_DIR --df-path DF_PATH [--debug]

optional arguments:
  -h, --help         show this help message and exit
  --out-dir OUT_DIR  Out directory to save results to
  --df-path DF_PATH  Input dataframe to draw annotations from
  --debug

For example:

python run_betabinom.py --df-path df_clean.pt --out-dir output/

After this program runs, to visualize/parse the results you will need to open up the iPython notebook src/Beta-Binomial Model Analysis.ipynb and replace the INPUT_DATA variable as instructed by the printout at the end of run_betabinom.py before running all the cells. The notebook will output both a visualization as well as a Pandas dataframe with accuracies and adjusted accuracies for each model.

Jackknife fitting

Similarly, run_jackknife.py will run the statistical jackknife correction (cf. Section 5.2) using the dataframe retrieved from the "Annotation Data" section above.

usage: run_jackknife.py [-h] --out-dir OUT_DIR --df-path DF_PATH [--debug]

optional arguments:
  -h, --help          show this help message and exit
  --out-dir OUT_DIR   Out directory to save results to
  --df-path DF_PATH   Input dataframe to draw annotations from
  --delete-d D        Runs the delete-D jackknife (default 1)
  --num-replicates N  Number of replicates to use to estimate expectations (default 100)
  --workers W         Number of threads to use (default 2)
  --save-justification    Also output plot of bias vs 1/n (cf. Figure 12)

For example:

python run_jackknife.py --df-path df_clean.pt --out-dir output/ --save-justification

Original data analysis

Finally, to run the analysis in Appendix D (original data analysis), run analyze_orig_data.py using the dataframe dowloaded in the first section.

usage: analyze_orig_data.py [-h] --out-dir OUT_DIR --df-path DF_PATH --experiment {naiveest, heldout, ezflickr}

optional arguments:
  -h, --help         show this help message and exit
  --out-dir OUT_DIR  Out directory to save results to
  --df-path DF_PATH  Input dataframe to draw annotations from
  --experiment       Which Appendix D experiment to run (D.1, D.2, D.3)
  --workers          Number of threads to use (default 2)
  --trials           Number of trials to average over (default 10)