Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

DL4J: Weight initialization for embedding layers needs review #8663

Closed
AlexDBlack opened this issue Feb 3, 2020 · 0 comments · Fixed by KonduitAI/deeplearning4j#279
Closed

DL4J: Weight initialization for embedding layers needs review #8663

AlexDBlack opened this issue Feb 3, 2020 · 0 comments · Fixed by KonduitAI/deeplearning4j#279
Assignees
@AlexDBlack
Labels
DL4J General DeepLearning4j issues UX Issues related to User Experience

Comments

@AlexDBlack
Copy link
Contributor

Currently users can use standard Xavier etc initialization, which as it stands seems to use fanIn == vocabSize, which results in very small initializations.

This should be checked, and a better scheme implemented (maybe treat fanIn as 1, or warn the user or something)
@AlexDBlack AlexDBlack added DL4J General DeepLearning4j issues UX Issues related to User Experience labels Feb 3, 2020
@AlexDBlack AlexDBlack self-assigned this Feb 28, 2020
AlexDBlack added a commit to KonduitAI/deeplearning4j that referenced this issue Feb 28, 2020
@AlexDBlack
deeplearning4j#8663 DL4J embedding layer weight init - don't depend o…
d82650c 
…n vocab size

Signed-off-by: Alex Black <blacka101@gmail.com>
@AlexDBlack AlexDBlack mentioned this issue Feb 28, 2020
Merged
AlexDBlack added a commit to KonduitAI/deeplearning4j that referenced this issue Mar 2, 2020
@AlexDBlack
Assorted SameDiff/DL4J fixes (#279)
483c3d7 
* deeplearning4j#8565 Normalizer toString/hashcode

Signed-off-by: Alex Black <blacka101@gmail.com>

* deeplearning4j#8731 ImagePreProcessingScaler lables/segmentation fix

Signed-off-by: Alex Black <blacka101@gmail.com>

* deeplearning4j#8691 Fix SameDiffLayer/Vertx finetuning and parameter setting support

Signed-off-by: Alex Black <blacka101@gmail.com>

* deeplearning4j#8663 DL4J embedding layer weight init - don't depend on vocab size

Signed-off-by: Alex Black <blacka101@gmail.com>

* EmbeddingLayer test tweak

Signed-off-by: Alex Black <blacka101@gmail.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@AlexDBlack AlexDBlack

Labels
DL4J General DeepLearning4j issues UX Issues related to User Experience
Projects
None yet
Milestone
No milestone
Development

Successfully merging a pull request may close this issue.

Assorted SameDiff/DL4J fixes KonduitAI/deeplearning4j
1 participant
@AlexDBlack