TensorFlow內部：摘要和TensorBoard (Inside TensorFlow: Summaries and TensorBoard)

字幕列表影片播放

NICK: Hi, everyone.
My name is Nick.
I am a engineer on the TensorBoard team.
And I'm here today to talk about TensorBoard and Summaries.
So first off, just an outline of what I'll be talking about.
First, I'll give an overview of TensorBoard, what it is
and how it works, just mostly sort of as background.
Then I'll talk for a bit about the tf.summary APIs.
In particular, how they've evolved from TF 1.x to TF 2.0.
And then finally, I'll talk a little bit
about the summary data format, log directories, event files,
some best practices and tips.
So let's go ahead and get started.
So TensorBoard-- hopefully, most of you
have heard of TensorBoard.
If you haven't, it's the visualization toolkit
for TensorFlow.
That's a picture of the web UI on the right.
Typically, you run this from the command line as the TensorBoard
command.
It prints out a URL.
You view it in your browser.
And from there on, you have a bunch of different controls
and visualizations.
And the sort of key selling point of TensorBoard
is that it provides cool visualizations out of the box,
without a lot of extra work.
you basically can just run it on your data
and get a bunch of different kinds of tools
and different sort of analyses you can do.
So let's dive into the parts of TensorBoard
from the user perspective a little bit.
First off, there's multiple dashboards.
So we have this sort of tabs setup
with dashboards across the top.
In the screenshot, it shows the scalers dashboard, which
is kind of the default one.
But there's also dashboards for images, histogram, graphs,
a whole bunch more are being added every month almost.
And one thing that many of the dashboards have in common
is this ability to sort of slice and dice
your data by run and by tag.
And a run, you can think of that as a sign
of a run of your TensorFlow program,
or your TensorFlow job.
And a tag corresponds to a specific named metric,
or a piece of summary data.
So here, the runs, we have a train
and evolve run on the lower left corner in the run selector.
And then we have different tags, including the cross [INAUDIBLE]
tag is the one being visualized.
And one more thing I'll mention is that one thing a lot
of TensorBoard emphasizes is seeing how
your data changes over time.
So most of the data takes the form of a time series.
And in this case, with the scalers dashboard,
the time series is sort of as a step count across the x-axis.
So we might ask, what's going on behind the scenes
to make this all come together?
And so here is our architecture diagram for TensorBoard.
We'll start over on the left with your TensorFlow job.
It writes data to disk using the tf.summary API.
And we'll talk both about the summary API and the event file
format a little more later.
Then the center component is TensorBoard itself.
We have a background thread that loads event file data.
And because the event file data itself
isn't efficient for querying, we construct a subsample
of the data and memory that we can query more efficiently.
And then the rest so TensorBoard is a web server that
has a plugin architecture.
So each dashboard on the frontend--
as a backend, it has a specific plugin
backend So for example, the scalers dashboard talks
to a scalers backend, images to an image backend.
And this allows the backends to do pre-processing or otherwise
structure the data in an appropriate way
for the frontend to display.
And then each plugin has a frontend dashboard component,
which are all compiled together by TensorBoard
and served as a single page and index.html.
And that page communicates back and forth through the backends
through standard HTTP requests.
And then finally, hopefully, we have our happy user
on the other end seeing their data,
analyzing it, getting useful insights.
And I'll talk a little more about just some details
about the frontend.
The front end is built on the Polymer web component
framework, where you define custom elements.
So the entirety of TensorBoard is one large custom element,
tf-tensorboard.
But that's just the top.
From there on, each plugin front end is--
each dashboard is its own frontend component.
For example, there's a tf-scaler dashboard.
And then all the way down to shared components
for more basic UI elements.
So we can think of this as a button, or a selector,
or a card element, or a collapsible pane.
And these components are shared across many of the dashboards.
And that's one of the key ways in which TensorBoard
achieves what is hopefully a somewhat uniform look
and feel from dashboard to dashboard.
The actual logic for these components
is implemented in JavaScript.
Some of that's actually TypeScript
that we compile to JavaScript.
Especially the more complicated visualizations,
TypeScript helps build them up as libraries
without having to worry about some of the pitfalls
you might get writing them in pure JavaScript.
And then the actual visualizations
are a mix of different implementations.
Many of them use Plottable, which
is a wrapper library over the D3, the standard JavaScript
visualization library.
Some of them use native D3.
And then for some of the more complex visualizations,
there are libraries that do some of the heavy lifting.
So the graph visualization, for example,
uses a directed graph library to do layout.
The projector uses a WebGL wrapper library
to do the 3D visualizations.
And the recently introduced What-If Tool plugin
uses the facets library from [INAUDIBLE] folks.
So we bring a whole bunch of different visualization
technologies together under one TensorBoard umbrella
is how you can think about the frontend.
So now that we have a overview of TensorBoard itself,
I'll talk about how your data actually gets to TensorBoard.
So how do you unlock all of this functionality?
And the spoiler announcement to that is the tf.summary API.
So to summarize the summary API, you
can think of it as structured logging for your model.
The goal is really to make it easy to instrument your model
code.
So to allow you to log metrics, weights,
details about predictions, input data, performance metrics,
pretty much anything that you might want to instrument.
And you can log these all, save them
to disk for later analysis.
And you won't necessarily always be calling the summary API
directly.
Some frameworks call the summary API for you.
So for examples, estimator has the summary saver hook.
Keras has a TensorBoard callback,
which takes care of some of the nitty gritty.
But underlying that is still the summary API.
So most data gets to TensorBoard in this way.
There are some exceptions.
Some dashboards have different data flows.
The debugger is a good example of this.
The debugger dashboard integrates with tfdbg.
It has a separate back channel that it uses
to communicate information.
It doesn't use the summary API.
But many of the commonly used dashboards do.
And so the summary API actually has sort of--
there's several variations.
And when talking about the variations,
it's useful to think of the API as having two basic halves.
On one half we have the instrumentation surface.
So these logging these are like logging
ops that you place in your model code.
They're pretty familiar to people
who have used the summary API, things like scaler, histogram,
image.
And then the other half of the summary API
is about writing that log data to disk.
And creating a specially formatted log
file which TensorBoard can read and extract the data from.
And so, just to give a sense of how those relate
to the different versions, there's
four variations of the summary API from TF 1.x to 2.0.
And the two key dimensions on which they vary
are the instrumentation side and the writing side.
And we'll go into this in more detail.
But first off, let's start with the most familiar summary
API from TF 1.x.
So just as a review-- again, if you've
used the summary API before, this will look familiar.
But this is kind of a code sample
of using the summary API 1.x.
The instrumentation ops, like scaler, actually output summary
protos directly.
And then those are merged together
by a merge all op that generates a combined proto output.
The combined output, you can fetch using session dot run.
And then, that output, you can write to a File Writer
for a particular log directory using
this add summary call that takes the summary proto itself
and also a step.
So this is, in a nutshell, the flow
for TF 1.x summary writing.
There's some limitations to this, which
I'll describe in two parts.
The first set of limitations has to do with the kinds of data
types that we can support.
So in TF 1.x, there's a fixed set of data types.
And adding new ones is a little involved.
It requires changes to TensorFlow in terms of you
would need a new proto definition field.
You'd need a new op definition, a new kernel, and a new Python
API symbol.
And this is a barrier to sensibility
for adding new data types to support new TensorBoard
plugins.
It's led people to do creative workarounds.
For example, like rendering a matplotlib plot
in your training code.
And then logging it as an image summary.
And the prompt here is, what if we instead
had a single op or a set of ops that could
generalize across data formats?
And this brings us to our first variation.
Which is the TensorBoard summary API,
where we try and make this extensible to new data types.
And the TensorBoard API, the mechanism
here is that we use the tensor itself as a generic data
container.
Which can correspond to--
for example, we can represent a histogram, an image,
scaler itself.
We can represent these all in certain formats as tensors.
And what this lets us do is use a shared tensor summary API
with some metadata that we can use
to describe the tensor format for our one place
to send summary data.
So TensorBoard.summary, the principle it takes
is actually that you can reimplement the tf.summary ops
and APIs as Python logic to call TensorFlow
ops for pre-processing and then a call to tensor summary.
And this is a win in the sense that you no longer need
individual C++ kernels and proto fields for each individual data
type.
So the TensorBoard plugins today actually do this.
They have for a while.
They have their own summary ops defined in TensorBoard.
And the result of this has been that for a new TensorBoard
plugins, where this is the only option,
there's been quite a bit of uptake.
For example, the pr_curve plugin has a pr_curve summary.
And that's the main route people use.
But for existing data types, there
isn't really much reason to stop using tf.summary.
And so, for those, it makes sense.
That's been what people have used.
But then tf.summary, it still has some other limitations.
And so that's what we're going to look at next.
So the second set of limitations in tf.summary
is around this requirement that the summary data
flows through the graph itself.
So merge_all uses the hidden graph collection essentially
to achieve the effect to the user
as though your summary ops have side effects of writing data.
Kind of like a conventional-- the way
you use a standard logging API.
But because it's using a graph collection,
it's not really safe for use inside control flow
and functions.
And also, with eager execution, it's very cumbersome to use.
You would have to keep track of outputs
by hand or in some way wait to send them to the writer.
And these limitations also apply to TensorBoard.summary ops
themselves.
Because they don't really change anything about the writing
structure.
And these limitations have sort of led to the prompt of,
what about if summary recording was an actual side
effect of op execution?
And so this brings us to tf.contrib summary,
which has new writing logic that achieves this.
And so here's a code sample for tf.contrib summary,
which looks pretty different from the original TF summary.
It works with eager execution.
But the change we have to make is now
we create a writer upfront via create_file_writer.
It's still tied to a specific log directory, which
we'll talk more about later.
You set the writer as the default writer in the context.
You enable summary recording.
And then the individual instrumentation ops
will actually write directly to the writer when they run.
So this gives you standard usage pattern of a logging
API that you would expect.
And it's compatible with eager execution
and also with graph execution.
So some details to how this works with contrib summaries.
The writer is backed in the TensorFlow runtime by a C++
resource called SummaryWriterInterface.
That essentially encapsulates the actual writing logic.
Which makes it possible in principle
to have different implementations of this.
The default writer, as conceived of by the Python code that's
executing, is just a handle to that resource
stored in the context.
And then instrumentation ops, like scaler and image,
now are stateful ops.
They have side effects.
And they achieve this by taking the default writer handle
as input along with the data they're supposed to write.
And then the actual op kernel implements the writing using
the C++ resource object.
And with this model, the Python writer objects
mostly manage this state.
They don't quite completely align because the C++ resource
could actually be shared across Python objects.
Which is a little bit different still
from the TensorFlow 2.0 paradigm,
where we want our Python state to reflect runtime state
1 to 1.
And this was just one example of a few things
that we're changing with TF 2.0.
And with TF 2.0, we had this opportunity
to stitch some of these features together and make
one unified new tf.summary API.
And so here we are completing our filling
out of the space of possibilities
where we have the tf.summary API bringing together features
from really all three of the existing APIs in 2.0.
So TF summary and TF 2.0, it really represents,
like I said, this unification of the three different APIs.
The instrumentation ops are actually
provided by TensorBoard.
And they use this generic Tensor data format.
That's the same format as SensorBoard.summary.
Which lets them extend to multiple different kinds
of data types.
We borrowed the implementation of the writing logic
from tf.contrib summary and some of the APIs.
But with slightly adjusted semantics in some
places, mostly just so that we align with the state
management in TF 2.0.
And then there's actually a not trivial amount
of just glue and circular import fixes
to get the two halves of both TensorBoard
and the original TF summary writing
API to talk to each other.
And I'll go into a little bit more detail about that.
So the dependency structure for tf.summary and TF 2.0
is a little bit complicated.
The actual Python module contains API symbols
from both TensorBoard and TensorFlow fused together.
But because the TensorBoard symbols also
depend on TensorFlow, this creates
this complicated dependency relationship.
And the way we linearize this dependency relationship
is that tf.summary in the original TensorFlow code
exposes the writing APIs, like create_file_writer
and the actual underlying writing logic.
Then we have what I call a shim module in TensorFlow,
that merges those symbols via wildcard import
with the regular imports of the instrumentation APIs,
like scaler and image that are now defined in TensorBoard.
And this produces the combined namespace.
But now it's a TensorBoard module.
So then TensorFlow, in it's top level init__.py where
it's assembling the API together,
imports the TensorBoard module and sets that as the new
tf.summary.
And this does mean that the API service depends directly
on TensorBoard.
But TensorFlow already has a pip dependency on TensorBoard.
So this isn't really a change in that respect.
But the API surface is now being constructed
through multiple components, where the summary component is
provided by TensorBoard.
And what that gives us is a single module that
combines both sets of symbols.
So that the delta for users is smaller.
But we can have the code live in the appropriate places for now.
And so these are some code samples for the TF 2.0 summary
API.
The first one shows it under eager execution.
It should look fairly similar to contrib.
You create the writer upfront, set it as default.
You can call the instrumentation ops directly.
So you no longer need to enable summary writing, which makes
it a little more streamlined.
And I should say that the ops, they
write when executed in theory.
But there's actually some buffering in the writer.
So usually you want to make sure to flush the writer
to ensure the data is actually written to disk.
And this example shows an explicit flush.
In eager, it will do flushing for you
when you exit the as default context.
But it's good if you care about making sure--
like, for example, after every iteration of the loop, that you
have data persisted to disk, it's good to flush the writer.
And then this is an example with the at tf.function decorator.
Again, you create the writer up front.
One important thing to note here is that the writer,
you have to maintain a reference to it
as long as you have a function that uses the writer.
And this has to do with the difference between when
the function is traced and executed.
It's a limitation that hopefully we
can improve this a little bit.
But for now, at least, that's one caveat.
So the best way to handle that is you
set the writer as default in the function itself.
And then call instrumentation ops that you need.
And these write, again, when executed,
meaning when the function is actually called.
So we can see that's happening down with the my_func call.
And then you can flush the writer again.
And then, here we have an example
with legacy graph execution, since there are still
folks who use the 2.0.
This is a little bit more verbose.
But again, you create the writer.
You set it as default. You've constructed your graph.
And then, in this case, you need to explicitly initialize
the writer if you're running init op.
We have sign of a compatibility shim
that lets you run all of the v2 summary ops.
So that you don't have to keep track of them manually.
And then, again, flushing the writer.
So this is how you would use it in legacy graph execution.
So how do you get to TF 2.0?
The contrib summary API is close enough to the 2.0 summary API
that we do--
actually, mostly, we auto-migrate this in the TF
upgrade v2 migration script.
But tf.summary in 1.0 is sufficiently
different on the writing side that we can't really
do a safe auto-migration.
So here is the three bullet version
of how to migrate by hand.
The first thing is that the writer now needs to be present,
created, and set via as default before using the summary ops.
And this is a limitation that it's a little bit tricky.
We're hoping to relax this a little bit.
So it's possible to set a writer later.
But for now, you want to have the default writer already
present.
Otherwise, the summary ops basically just become no ops
if there's no writer, since they have no where to write to.
Then each op takes its own step argument now.
This is because since there's no later step where
you add the summary to the writer, that's
where the step was previously provided.
And there's also no global step in TF 2.0.
So there isn't really a good default variable to use.
So for now, steps are being passed explicitly.
And I'll talk about this a little more later
on the next slide.
And then the function signatures for the instrumentation ops,
like scaler and image, have change slightly.
The most obvious thing being that they no longer
return an output.
Because they write via side effect.
But also there's slight differences
in the keyword arguments that won't affect most people.
But it's something good to know about.
And these details will all be in the external migration guide
soon.
And so the other changes--
and this and some of the other stuff I was mentioning.
One change is with graph writing.
Since there's no default global graph in 2.0,
there's no direct instrumentation op
to write the graph.
Instead, the approach here is there's
a set of tracing style APIs to enable and disable tracing.
And what those do is they record the graphs of executing TF
functions.
So functions that execute well, the tracing
is enabled to the summary writer.
And this better reflects the TF 2.0 understanding
of graphs as something that are associated with functions
as they execute.
Then this is what I was alluding to.
It's still a little bit tricky to use default writers
with graph mode since it's not always the case
that you know which writer you want to use
as you're assembling the graph.
So we're working on making that a little bit
more user friendly.
And setting the step for each op is also definitely boilerplate.
So that's another area where we're
working to make it possible to set the step, maybe in one
place, or somehow in the context to avoid the need to pass it
into ops individually.
And then, the event file binary representation has changed.
This only affects you.
This doesn't affect TensorBoard in that TensorBoard already
supports this format.
But if you were parsing event files in any manual way,
you might notice this change.
And I'll talk a little bit more about that change
in the next section.
And finally, as mentioned, the writers now have a one to one
mapping to the underlying resource and event file.
So there's no more sharing of writer resources.
OK.
And then the last section will be about the summary data
format.
So this is log directories, event files,
how your data is actually persisted.
So first off, what is a log directory?
The TensorBoard command expects a required dash dash logdir
flag.
In fact, your first introduction to TensorBoard
may have been trying to run it.
And then it spits out an error that you
didn't pass the logdir flag.
So the log directory flag is the location
that TensorBoard expects to read data from.
And this is often the primary output directory
for a TensorFlow program.
Frameworks, again, like Estimator and Keras
have different knobs for where output goes.
But often, people will put it all under one root directory.
And that's often what people use is the log directory.
But TensorBoard has this flexible interpretation
where really all it cares about is that it's a directory
tree containing summary data.
And when I say directory tree, I really do mean a tree.
Because the data can be arbitrarily deep.
TensorBoard will traverse the entire tree
looking for summary data.
And you might think, that could sort of be a problem sometimes,
especially if there's hundreds, thousands of event files.
And it's true.
Yeah, log directories can be pretty large.
And so TensorBoard tries to take advantage
of structure in the log directory
by mapping sub directories of the logdir
to this notion of runs, which we talked
about a little bit in the early section
about the TensorBoard UI.
So again, these are runs.
Like, a run of a program, they're
not individual session.run calls.
And when TensorBaord loads a run, the definition it uses
is that it's any directory in the logdir that has
at least one event file in it.
And in this case, we mean only direct children.
So the directory has to contain an actual event file.
And an event file is just defined
as a file that has the name, has the string
tfevents in the name.
Which is just the standard naming convention
used by summary writers.
So as an example of this, we have this log directory
structure which has a root directory logs.
It has two experiment sub directories in it.
The first experiment contains an event file.
So that makes that itself a run.
It also contains two sub directories, train and eval,
with event files.
So those two also become runs.
Visually, they look like sub runs.
But they're all considered independent runs
for TensorBoard, at least in the current interpretation.
And then, in experiment two, that
doesn't contain an event file directly.
So it's not a run.
But it has a train sub directory under it.
So TensorBoard looks at this log directory
and traverses it and finds four different runs.
And this traversal step happens continuously.
TensorBoard will pull a log directory for new data.
And this is to facilitate using TensorBoard as a way
to monitor the progress of a running job,
or even potentially a job that hasn't started yet.
You might start TensorBoard and your job at the same time.
So this directory may not even exist yet.
And we may expect that different runs
will be created as it proceeds.
So we need to continuously check for new directories being
created, new data being appended.
In the case where you know that your data is not changing,
like you're just viewing old data,
you can disable this using the reload interval flag.
And you can also adjust the interval at which it pulls.
So when it's traversing the log directory,
it does this in two passes.
The first pass is finding new runs.
So it searches the directory tree for new directories
with TF event files in them.
This can be very expensive if your tree is deeply nested,
and especially if it's on a remote file system.
And especially if the remote file system
is on a different continent, which I've seen sometimes.
So a key here is that walking the whole directory tree
can be pretty slow.
We have some optimizations for this.
So for example, on Google Cloud Storage,
rather than walking each directory individually,
we have this iterative globbing approach.
Which we basically use to find all directories at a given
depth at the same time, which takes advantage of the fact
that GCS doesn't actually really have directories.
They're sort of an illusion.
And there's other file system optimizations like this
that we would like to make as well.
But that's just one example.
And then the second pass, after it's found all the new runs,
is that it reads new event file data from each run.
And it goes through the runs essentially in series.
There is a limiting factor in Python itself
for paralyzing this.
But again, something that we are interested in working
on improving.
And then, when you actually have the set
of event files for a run, TensorBoard
iterates over them, basically, in directory listing order.
You might have noticed on the previous slide with the example
logdir that the event files all have a fix, prefix,
and then a time stamp.
And so what this means is that the directory
order is essentially creation order of the event files.
And so, in each event file, we read records from it
sequentially until we get to the end of the file.
And then at that point, TensorBoard
checks to see if there's a subsequent file already
created.
If so, it continues to that one.
Otherwise, it says, OK, I'm done with this run.
And then it goes to the next run.
And after it finishes all the runs,
it waits for the reload interval.
And then it starts the new reload cycle.
And this reload resumes the read from the same
offset in every file per run that it stopped in.
And an important thing to point out here
is that TensorBoard won't ever revisit an earlier
file within a run.
So if it finishes reading a file and continues to a later one,
it won't ever go back to check if the previous file contains
new data.
And this is based on the assumption
that the last file is the only active one, the only one being
actively written to.
And it's important to avoid checking all event files, which
can be-- sometimes there's thousands
of event files in a single run directory.
And so that's a mechanism for avoiding wasted rechecking.
But this assumption definitely doesn't always hold.
There are cases when a single program
is using multiple active writers within a run.
And in that case, it can seem like the data is being skipped.
Because you proceed to a new file.
And then data added to the original file no longer appears
in TensorBoard.
And luckily, it's fairly straightforward
to work around this.
You just restart TensorBoard.
And it will always pick up all the data that existed
at the time that it started.
But we're working on a better fix for this.
So that we can still detect and read when there is
data added to files other than the last one.
But this is something that has bitten people before.
So just a heads up.
And then, so the actual event file format,
this is based on TFRecord, which is the standard TensorFlow
format.
It's the same as tf.io TFRecordWriter.
And it's a pretty simple format, enough that it fits
on the left side of this slide.
It's basically just a bunch of binary strings prefixed
by their length with CRCs for data integrity.
And one particular thing that I'll note
is that because there's no specific length
for each string, there's no real way to seek ahead in the file.
You basically have to read it sequentially.
And there's also no built in compression of any kind.
And TensorBoard, it's possible in theory
to have the whole file be compressed.
TensorBoard doesn't support this yet.
But it's something that could help save space
when there's a lot of redundant strings within the event file.
And then each individual record--
so TFRecord is the framing structure for the event file.
Each individual record is a serialized event protocol
buffer.
And this simplified schema for the protocol buffer
is shown on the left.
We have a wall time and a step, which
are used to construct the time series.
And then we have a few different ways to store data.
But the primary one is a summary sub-message.
The main exception is graph data get stored in the GraphDef
separately.
And then we can look at the summary sub-message,
which is itself basically a list of value sub-messages.
That's where the actual interesting part is.
And each one of these contains a tag.
Again, from our overview of the UI,
that's the name or idea of the summary
as shown in TensorBoard.
We have metadata, which is used to describe the more
generic tensor formats.
And then specific type fields, including
ones for the original TF 1.x, specific fields for each type.
And then the tensor field, which can be used with the new tensor
style instrumentation ops to hold general forms of data.
And then, in terms of loading the summary data
into memory-- so I mentioned this briefly
in the architecture stage.
But TensorBoard has to load the summary data into memory.
Because, like I said, there's no real indexing or random access
in the event file.
You can think of them like they're just like raw logs.
And so TensorBoard loads it into memory
and creates its own indexes of data by run, plugin, and tag.
Which support the different kinds of visualization queries
that plugins need.
And TensorBoard also does downsampling in order
to avoid running out of memory since the log directory may
contain far more data than could reasonably
fit in TensorBoard's RAM.
And to do the downsampling, it uses
a reservoir sampling algorithm.
It's essentially just an algorithm for uniform
sampling when you don't know the size of your sequence
in advance, which is the case when we're consuming data
from an active job.
And because of this, it has a random aspect which
can be surprising to users.
Where you might not understand-- like,
why is this step being taken and not this one.
And there's, like, a gap between steps.
This can be tuned with a samples per plugin flag.
So that tunes the size of the reservoir.
Basically, if you make the reservoir larger
than your total number of steps, you'll
always see all of your data.
So that gives you a certain amount of control
over how about sampling works.
And just to review some of the section,
there's some best practices for-- at least
in the current TensorBoard, how to get the best performers.
One of the basic ones is just to avoid having enormous log
directories, in terms of the number
of files, subdirectories, quantity of data.
This doesn't actually mean that the overall log
directory has to be small.
It just means that TensorBoard itself
will run better if you can launch it
at a directory that just contains relevant data for what
you want to examine.
So, a hierarchical structure where
you can pick sort of a experiment sub-directory,
or group a few experiments together
works really well here.
I mentioned the reload interval.
You can set it to 0 to disable reload.
So for unchanging data, this helps avoid extra overhead.
And that's especially useful on a remote file system case.
In that case, it's also useful if you
can run TensorBoard close to your data,
or download a subset of it that you need so that it doesn't all
have to be fetched over the network.
And for now, due to the way that superseded event files aren't
re-read, it's best to avoid multiple active writers pointed
at the same directory.
This is something, again, that we're actively
working on improving this.
But at least for now, that can lead to this appearance
that some data gets skipped.
And then, in general, stay tuned for logdir performance
improvements.
We're hoping to make a number of these improvements soon.
And that pretty much rounds out the content
that I have for today.
So I'd like to thank everybody for attending.
And if you want to find out more about TensorBoard,
we have a new sub site on tensorflow.org/tensorboard.
You can also find us on GitHub.
And people interested in contributing
are welcome to join the TensorBoard special interest
group.
Thank you all very much.
AUDIENCE: What are the potential uses for the multiple summary
writers?
NICK: The potential use cases for the multiple summary
writers, sometimes this is just a matter of code structure.
Or if you're using a library that itself creates a summary
writer, it's not always straightforward to ensure
that you can use the same writer.
So TF 1.x had this File Writer cache, which was--
the best practice then was to use this shared cache
since we only had one writer per directory.
And it was to work around this problem.
I think it's better to work around it
on the TensorBoard side and have some ideas for how to do that.
So hopefully, that part will be out of date soon.
Like, within a month or two.
AUDIENCE: Are there plans to change the event file format?
NICK: Yeah.
So I think a lot of this depends on--
I think that the event file format
itself could be a lot better tailored to what TensorBoard
actually needs.
And some of the things I mentioned would just be--
even if we had an index into the event file,
that could potentially help--
we could potentially paralyze reads.
Or we could sort of scan ahead and do smarter sampling.
Like, rather than reading all the data
and then down sampling it, we could just
pick different offsets and sample from there.
We're constrained right now mostly by this being the legacy
format.
But I think it would be pretty interesting to explore
new formats.
Particularly when you have different data types mixed in,
something sort of columnar could be kind of useful,
where you can read only images if you
need to read images, or otherwise
avoid the phenomenon where-- so, this happens sometimes
where one particular event file contains
lots and lots of large images or large graph defs.
And this blocks the reading of a lot of small scaler data.
And that's, obviously, not really--
it doesn't really make sense.
But again, it's a limitation of having the data only be
accessible sequentially.
AUDIENCE: Can you talk a little bit more
about the graph dashboard?
Is graph just another summary?
Or how does--
NICK: Yeah.
So the graph visualization, which was actually
originally created by Big Picture,
It's a visualization of the actual TensorFlow graph.
So like, the computation graph with ops and edges
connecting them and sort of the data flow.
It's pretty cool.
It's really a good way if you want to visually understand
what's going on.
It works best when the code uses scoping to delineate
different parts of the model.
If it's just a giant soup of ops,
it's a little hard to understand the higher order structure.
And there's actually some cool work
done for TF 2.0, which isn't in the presentation,
about for Keras showing the Keras conceptual graph using
Keras layers to give you a better view into the high level
structure of the model like you'd
expect to see in a diagram written by a human.
But the graph dashboard can still
be useful for understanding exactly what ops are happening.
And sometimes it's useful for debugging cases.
If some part of the graph is behaving weirdly,
maybe you didn't realize that you actually
have an edge between two different ops that
was unexpected.
AUDIENCE: Are there any plans to add support
for basically that sort of higher order structure
annotation?
So I'm imagining, for instance, like,
if you have [INAUDIBLE] having the whole model, and then
a block, and then a sub block, and if there's
like five layers of structural depth, that would
be nice to be able to imitate.
NICK: Yeah.
I think this is an interesting question.
So right now, the main tool you have for that structure
is just name scoping.
But it only really works if the part of the graph
is all been defined in the same place anyway.
I think it would be really nice to have the graph visualization
support more strategies for human friendly annotation
and organization.
I think the recent work that we've done on this
has been the Keras conceptual graph, which
[? Stephan ?] did over there.
But I think having that work for not just Keras layers, but more
general model decomposition approaches would
be really nice.
AUDIENCE: Often, I find that the problem isn't
that it's not capturing enough.
It's actually that it's capturing too much.
So for instance, you'll have a convolution.
But then there's a bunch of loss and regularization.
There winds up being a bunch of tensors that sort of clutter.
And so actually, even the ability to filter stuff
out of the conceptual graph.
NICK: So there id a feature in the graph dashboard
where you can remove nodes from the main graph.
But I believe it has to be done by hand.
It does a certain amount of automatic extraction
of things that are sort of less important out of the graph.
But maybe that's a place we could look into having either
a smarter procedure for doing that,
or a way to sort of tag like, hey, this section of--
I don't actually want to see any of these.
Or this should be factored out in some way.
Yeah.
Thanks.
[APPLAUSE]