Fixing the numerical instability when calculating the loss of the cri… #501
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
minjiaz
requested review from
jeffra,
samyam,
tjruwase,
ShadenSmith,
conglongli,
awan-10,
cli99,
eltonzheng,
RezaYazdaniAminabadi,
duli2012,
mrwyattii,
yaozhewei,
arashb and
xiaoxiawu-microsoft
as code owners
yaozhewei
approved these changes
May 9, 2023
Syulin7
pushed a commit
to Syulin7/DeepSpeedExamples
that referenced
this pull request
May 15, 2023
microsoft#501) * Fixing the numerical instability when calculating the loss of the critic model * formatting * formatting --------- Co-authored-by: Michael Wyatt <michaelwyatt@microsoft.com> Co-authored-by: Michael Wyatt <mrwyattii@gmail.com>
leocnj
pushed a commit
to leocnj/DeepSpeedExamples
that referenced
this pull request
May 27, 2023
microsoft#501) * Fixing the numerical instability when calculating the loss of the critic model * formatting * formatting --------- Co-authored-by: Michael Wyatt <michaelwyatt@microsoft.com> Co-authored-by: Michael Wyatt <mrwyattii@gmail.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
When training the step2 critic model, our current implementation has an issue where the loss calculation using -log(sigmoid(c_truncated_reward - r_truncated_reward)) could result in infinite loss when sigmoid returns 0. To fix this issue, I replaced it with -torch.nn.functional.logsigmoid(c_truncated_reward - r_truncated_reward), which is a numerically stable version of log(sigmoid(x)) implemented in PyTorch. I also tested other implementations, such as nn.functional.softplus( r_truncated_reward - c_truncated_reward), but so far logsigmoid provides the closest value to our current implementation while avoiding infinite loss.
Testing results:
i, -torch.log(torch.sigmoid(x)), torch.nn.functiona.softplus(-x), -torch.nn.functional.logsigmoid(x)
-100 , tensor(inf) tensor(100.) tensor(100.)
-90.0 , tensor(inf) tensor(90.) tensor(90.)
-80.0 , tensor(80.) tensor(80.) tensor(80.)
-70.0 , tensor(70.) tensor(70.) tensor(70.)
-60.0 , tensor(60.) tensor(60.) tensor(60.)
-50.0 , tensor(50.) tensor(50.) tensor(50.)
-40.0 , tensor(40.) tensor(40.) tensor(40.)
-30.0 , tensor(30.) tensor(30.) tensor(30.)
-20.0 , tensor(20.) tensor(20.) tensor(20.)
-10.0 , tensor(10.0000) tensor(10.0000) tensor(10.0000)
0.0 , tensor(0.6931) tensor(0.6931) tensor(0.6931)
10.0 , tensor(4.5420e-05) tensor(4.5399e-05) tensor(4.5418e-05)
20.0 , tensor(-0.) tensor(2.0612e-09) tensor(0.)
30.0 , tensor(-0.) tensor(9.3576e-14) tensor(0.)
40.0 , tensor(-0.) tensor(4.2484e-18) tensor(0.)
50.0 , tensor(-0.) tensor(1.9287e-22) tensor(0.)
60.0 , tensor(-0.) tensor(8.7565e-27) tensor(0.)
70.0 , tensor(-0.) tensor(3.9754e-31) tensor(0.)
80.0 , tensor(-0.) tensor(1.8049e-35) tensor(0.)
90.0 , tensor(-0.) tensor(8.1940e-40) tensor(0.)
100.0 , tensor(-0.) tensor(3.7835e-44) tensor(0.)