This repository has been archived by the owner on Aug 15, 2019. It is now read-only.
Optimize tf.tidy() and tf.keep(). #1621
Merged
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
added 2 commits
March 12, 2019 16:14
dsmilkov
approved these changes
Mar 12, 2019
There was a problem hiding this comment.
Reviewed 2 of 2 files at r1.
Reviewable status:complete! 1 of 1 approvals obtained (waiting on @caisq and @dsmilkov)
dsmilkov
added a commit
that referenced
this pull request
Mar 13, 2019
BUG Improve memory management of tensors during training. Op authors now explicitly save intermediate tensors that are needed for backwards mode. This allows the engine to optimally dispose memory during forward pass, and keep only tensors needed for the backward pass. Based on the [layers benchmark](https://github.com/tensorflow/tfjs-layers/tree/v1.0.0/integration_tests/benchmarks), this change along with #1621 led to: ## 2-3X memory reduction **Before**  **After**  ## 1.5-1.7x improvement in fit() for GRU and LSTM ops. **Before**  **After**  - When the user writes the forward pass of an op, they are given a `save` function that allows them to save inputs or intermediate tensors to be reused for the backwards pass - Before this change, the `save` function was a no-op, a placeholder for when we decide to optimize disposal of tensors in the training process in the future. - However, `save` being a no-op caused a bug for existing users who rely on it (e.g. Magenta.js). - After this change, the `save` function makes a shallow copy of the tensor, and keeps it until the backwards pass is done. - `save` used to take an array of tensors. Now it takes a `NamedTensorMap` which improves code readability, and reduces chances of off-by-one index bugs. Fixes tensorflow/tfjs#1320 PERF BUG
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
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.
We do this by:
tf.keep()tracking. Instead, we add a "kept" bit to tensors and we implicitly remove kept tensors from tracking mechanisms. When we are going to track in a parent scope, we avoid doing that if the kept bit is set on the tensor.tracksets the tensor's scope ID and then when we're checking whether the tensor belongs to the current scope when tracking in a parent, we simply check that id matching the current scope ID.No unit tests added since we have pretty serious coverage over memory and this is an internal optimization.
This change is