Address #252 - Make soft ranking notebook self contained #301
Conversation
AWehenkel
commented
Feb 15, 2023
AWehenkel
commented
- Remove the examples folder.
- Remove the fairness notebook.
- Make the notebook for soft ranking self-contained.
- Switch for an example with MNIST to make things runnable in a decent amount of time on CPU.
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
|
View / edit / reply to this conversation on ReviewNB michalk8 commented on 2023-02-15T18:14:32Z Line #14. import scipy.ndimage Use |
|
View / edit / reply to this conversation on ReviewNB michalk8 commented on 2023-02-15T18:14:33Z Line #15. from flax.training import common_utils Seems like common_utils are not used, please remove |
|
View / edit / reply to this conversation on ReviewNB michalk8 commented on 2023-02-15T18:14:34Z Line #20. import ott For consistency with other tutorials, please use |
|
View / edit / reply to this conversation on ReviewNB michalk8 commented on 2023-02-15T18:14:35Z Line #1. url = "https://raw.githubusercontent.com/matplotlib/matplotlib/master/doc/_static/stinkbug.png" Since we removed
with urllib.request.urlopen(url) as resp: image = plt.imread(io.BytesIO(resp.read())) |
|
View / edit / reply to this conversation on ReviewNB michalk8 commented on 2023-02-15T18:14:36Z Line #1. def salt_and_pepper(im, amount=0.05): Completely optional, but if you want, would be great to add type hints to all the functions. |
|
View / edit / reply to this conversation on ReviewNB michalk8 commented on 2023-02-15T18:14:37Z Optax with capital "O" ... neural network with Jax, as well PyTorch dataset and dataloaders. |
|
View / edit / reply to this conversation on ReviewNB michalk8 commented on 2023-02-15T18:14:38Z Please move this function to the cell where you define |
|
View / edit / reply to this conversation on ReviewNB michalk8 commented on 2023-02-15T18:14:39Z Also add these functions to the cell where |
|
View / edit / reply to this conversation on ReviewNB michalk8 commented on 2023-02-15T18:14:40Z Line #1. @flax.struct.dataclass For consistency, use and then |
|
View / edit / reply to this conversation on ReviewNB michalk8 commented on 2023-02-15T18:14:41Z Move this below the cell below.
|
|
View / edit / reply to this conversation on ReviewNB michalk8 commented on 2023-02-15T18:14:42Z Use pass instead of |
|
View / edit / reply to this conversation on ReviewNB michalk8 commented on 2023-02-15T18:14:43Z Line #60. log(results, step + 1, summary, train=True) I'd either log every n-th iteration (add a new config value) or use |
|
View / edit / reply to this conversation on ReviewNB michalk8 commented on 2023-02-15T18:14:44Z Instead of "loss", would use "soft error loss" |
|
View / edit / reply to this conversation on ReviewNB michalk8 commented on 2023-02-15T18:14:45Z I'd add legend for all 3 axes. Would it be possible to train it on GPU for a bit longer and see if CE increases, as in the original notebook? Would be nice to see if we can replicate the trend in the original notebook. |
# Conflicts: # examples/fairness/losses.py # examples/fairness/main.py # examples/fairness/train.py # examples/soft_error/data.py # examples/soft_error/losses.py # examples/soft_error/main.py # examples/soft_error/train.py
|
Thanks @AWehenkel ! |
* Test commit on personal fork. * Test commit on personal fork. * Update the soft sort notebook to make it self-contained. Remove the `examples` folder. * pre-commit the soft sort notebook * runnable on GPUs + Michaal comments on PR * Clean the NB. * Clean the NB. --------- Co-authored-by: antoinewehenkel <[email protected]>
