commit 2ab9dfaee5c791e0d7b2c0893a55ced8677e2e2a
author wenson_hsieh@apple.com <wenson_hsieh@apple.com@268f45cc-cd09-0410-ab3c-d52691b4dbfc> Sun Jul 14 03:36:19 2019 +0000
committer wenson_hsieh@apple.com <wenson_hsieh@apple.com@268f45cc-cd09-0410-ab3c-d52691b4dbfc> Sun Jul 14 03:36:19 2019 +0000
tree 6936c86a80ffd1dc85c58199703671c1faf7dbeb
parent 615a3e0bc7e49bb529533f6adc6ffb4ce43c6e9b

[Text autosizing] [iPadOS] Further adjust our heuristics to determine text autosizing candidates
https://bugs.webkit.org/show_bug.cgi?id=199780
<rdar://problem/52289088>

Reviewed by Simon Fraser.

Source/WebCore:

Our current idempotent text autosizing candidate heuristic makes the right judgment call most of the time, but
there is still a large batch of text autosizing bugs left unfixed by the first iteration of the heuristic added
in r246781. This patch attempts to address most of these bugs by adjusting the decision-tree-based heuristic
once again, mostly with improvements to the model generation pipeline.

During the first iteration, I placed emphasis on tuning the max tree depth and min leaf size hyperparameters
when coming up with my decision tree, and didn't consider the inclusion or exclusion of each feature as a
hyperparameters. As such, the trees generated using the pipeline tended to use too many features, and as a
result, tended to have cross-validation overall accuracy scores hovering around 73%.

In this revised model generation pipeline, I now consider the inclusion of each feature (along with max depth
and min leaf size, as before) as a hyperparameter. Since this increases the number of hyperparameters by many
orders of magnitude, a naive grid search (as described in the prior ChangeLog entry) is no longer a tractible
procedure for tuning hyperparameters to the training algorithm.

Instead, I now use a stochastic greedy algorithm to search for good sets of hyperparameters; this process begins
with seeding some number (usually 20-24) of "searchers" with completely randomized sets of hyperparameters (i.e.
random max depth, random leaf size, and random subsets of features). I then evaluate the average performance of
each set of hyperparameters by using them to generate 2000 decision trees over 90% of the training data, and
then cross-validating these trees against the remaining 10%. These cross-validation scores are aggregated into a
single confusion matrix, which is then passed into a loss function that computes a single value indicating how
well training with the set of hyperparameters generalized to cross-validation data. After experimenting with
various loss functions, I settled on the following:

`k(false positive rate)^2 + (false negative rate)^2`

...where a constant k is chosen to penalize false positives (i.e. broken layout) more harshly than false
negatives (small text). Additionally, squaring the false negative and false positive rates seems to help avoid
converging on solutions that heavily favor reducing only false positives or false negatives, or vice versa.

The stochastic algorithm starts by computing a loss value for the randomly generated configuration. Then, for
an indefinite number of iterations, it randomly mutates the configuration (e.g. by adding or removing features,
or changing min leaf size or max tree depth) and computes a new loss value for the mutated configuration. If the
mutated configuration performs better (i.e. achieves lower loss) than the current configuration, I set the
current configuration to be the mutated configuration. Otherwise, I keep the current (non-mutated) configuration
as-is. The stochastic algorithm then proceeds, ad-infinitum, with this current configuration.

Of course, since each mutation is small, this strategy so far is prone to leaving each searcher stuck in local
optima. To mitigate this, for each searcher, I keep track of a side-table of configurations that have already
been tested; when random mutations would normally lead to testing a configuration that has already been tested,
each searcher instead increases the chance of applying additional mutations. This has the effect of searchers
initially exhausting similar configurations, and expanding to test more and more dissimilar configurations as
the local alternatives all turn out to be worse. This allows searchers to effectively jump out of local optima
after being stuck for a long time.

So, using these strategies, I simultaneously ran a handful of searchers until they all appeared to converge
(a process that takes 8-12 hours on my current dataset). Many of the searchers achieved configurations with
cross-validation scores of 81% and above, up from the 73% of the previous attempt. These additionally have the
added bonus of reducing the number of features, often making the final trees themselves shallower and simpler to
understand than before.

This patch introduces one such decision tree generated using a set of hyperparameters acquired via this
stochasic search algorithm; it appears to simultaneously use fewer features, and achieve better cross-validation
performance.

Test: fast/text-autosizing/ios/idempotentmode/idempotent-autosizing-candidates.html

* css/StyleResolver.cpp:
(WebCore::StyleResolver::adjustRenderStyleForTextAutosizing):

Adjust the early return to bail if either (1) the element is a candidate and the computed size is already equal
to the boosted size, or (2) the element is not a candidate and the computed size is already equal to the
specified size. Since the autosizing candidate heuristic depends on styles specified on the element itself (as
opposed to styles on any element in the ancestor chain), a parent may be an autosizing candidate, but a child of
it may not.

* rendering/style/RenderStyle.cpp:
(WebCore::RenderStyle::isIdempotentTextAutosizingCandidate const):

Revamp the idempotent text autosizing candidate heuristic. See the explanation above for more details.

* rendering/style/RenderStyle.h:

Remove some bits from RenderStyle's autosizeStatus, now that we care about fewer bits of information from the
inherited flags.

* rendering/style/TextSizeAdjustment.cpp:
(WebCore::AutosizeStatus::updateStatus):
* rendering/style/TextSizeAdjustment.h:

LayoutTests:

Rebaseline an existing idempotent text autosizing test, and add an additional test case.

* fast/text-autosizing/ios/idempotentmode/idempotent-autosizing-candidates-expected.txt:
* fast/text-autosizing/ios/idempotentmode/idempotent-autosizing-candidates.html:


git-svn-id: http://svn.webkit.org/repository/webkit/trunk@247421 268f45cc-cd09-0410-ab3c-d52691b4dbfc