Add non-linearity to hidden state in char-RNN generation tutorial - #3956
Add non-linearity to hidden state in char-RNN generation tutorial#3956adityasharmaaaaa wants to merge 1 commit into
Conversation
Sorry, something went wrong.
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/tutorials/3956
Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. |
Sorry, something went wrong.
|
Thank you for your pull request and welcome to our community. Action RequiredIn order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you. ProcessIn order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA. Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it. If you have received this in error or have any questions, please contact us at cla@meta.com. Thanks! |
Sorry, something went wrong.
|
Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Meta Open Source project. Thanks! |
Sorry, something went wrong.
Fixes #3953
Problem
The RNN.forward() in this tutorial passed the hidden state straight
through a linear layer with no activation:
This makes h_t a linear combination of h_{t-1}, the input, and the
category tensor at every timestep — nn.LogSoftmax on the output was
the only non-linear operation anywhere in the recurrence.
Fix
Added a nn.Tanh() on the hidden path:
This is a minimal, targeted fix for the specific issue raised — it
doesn't change i2o, o2o, or dropout, so the rest of the
architecture and surrounding explanation stay valid. Also added one
sentence to the "Creating the Network" section explaining why the
non-linearity is there.
Testing
Ran the tutorial locally for the full 100,000 iterations. Loss trends
down as expected (starts ~2.98, settles in the 1.4–2.4 range with
normal SGD noise), and sampled names after training look reasonable
(e.g. Rovakin, Sarana, Chan), confirming the change doesn't
break training.