New plots and results related to adversarial training

View / edit / reply to this conversation on ReviewNB

tomtseng commented on 2023-06-06T04:27:09Z ----------------------------------------------------------------

Line #53.        plt.xlabel("Cyclic-adversary finetuning GPU days")

nit: we write "[V100] GPU-days" (hyphen) instead of "GPU days" (no hyphen) elsewhere in this notebook and in the paper — we should be consistent

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

tomtseng commented on 2023-06-06T04:30:12Z ----------------------------------------------------------------

nit: we write "fine-tune" instead of "finetune" in the paper, we should be consistent about which one we write

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

tomtseng commented on 2023-06-09T21:21:12Z ----------------------------------------------------------------

Line #19.        continue_steps: list[int] | None = None,

maybe should comment what continue_steps means

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

tomtseng commented on 2023-06-09T21:21:13Z ----------------------------------------------------------------

Line #85.                dict(

maybe remove opt_return and always return the dict? the caller can ignore the return value if they don't care about it

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

tomtseng commented on 2023-06-09T21:21:14Z ----------------------------------------------------------------

this comment is in the wrong location

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

tomtseng commented on 2023-06-09T21:21:14Z ----------------------------------------------------------------

Line #8.    df_ft["adv_orig_steps"] = df_ft.adv_name.str.split("-cs").str[0].str.split("-os").str[-1].astype(int)

may be helpful to add a comment about what the naming convention for adversaries was for these files — what do cs and os mean?

also, what does cml mean? cumulative? not immediately clear to me what these columns are (though maybe once I see the naming convention for the adversaries it will be easier to figure out)

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

tomtseng commented on 2023-06-09T21:21:15Z ----------------------------------------------------------------

Line #22.        drop_steps = steps[drop_left : len(steps) - drop_right]

why do we need to drop items from the beginning and end? a comment on what this cell is doing would be helpful

why drop the change from v32 to v64 at s26985728 for tony-cyc-adv-ft-vs-b60-s7702m-20230518-185923?

_ed1d1a8d commented on 2023-06-20T17:12:58Z_ ----------------------------------------------------------------

Added a comment.

review-notebook-app[bot] commented 1 year ago

View / edit / reply to this conversation on ReviewNB

tomtseng commented on 2023-06-09T21:21:16Z ----------------------------------------------------------------

text here seems outdated

ed1d1a8d commented on 2023-06-20T17:14:27Z ----------------------------------------------------------------

Updated text.

ed1d1a8d commented 1 year ago

Added a comment.

View entire conversation on ReviewNB

ed1d1a8d commented 1 year ago

Updated text.

View entire conversation on ReviewNB

AlignmentResearch / KataGoVisualizer

New plots and results related to adversarial training #87