Posts in category: Tensorflow

Posted by Jen Person, Senior Developer Relations Engineer, TensorFlow

Welcome to part 2 of my dual approach to content moderation! In this post, I show you how to implement content moderation using machine learning in a server-side environment. If you’d like to see how to implement this moderation client-side, check out part 1.

Remind me: what are we doing here again?

In short, anonymity can create some distance between people in a way that allows them to say things they wouldn’t say in person. That is to say, there are tons of trolls out there. And let’s be honest: we’ve all typed something online we wouldn’t actually say IRL at least once! Any website that takes public text input can benefit from some form of moderation. Client-side moderation has the benefit of instant feedback, but server-side moderation cannot be bypassed like client-side might, so I like to have both.

This project picks up where part 1 left off, but you can also start here with a fresh copy of the Firebase Text Moderation demo code. The website in the Firebase demo showcases content moderation through a basic guestbook using a server-side content moderation system implemented through a Realtime Database-triggered Cloud Function. This means that the guestbook data is stored in the Firebase Realtime Database, a NoSQL database. The Cloud Function is triggered whenever data is written to a certain area of the database. We can choose what code runs when that event is triggered. In our case, we will use the Text Toxicity Classifier model to determine if the text written to the database is inappropriate, and then remove it from the database if needed. With this model, you can evaluate text on different labels of unwanted content, including identity attacks, insults, and obscenity. You can try out the demo to see the classifier in action.

If you prefer to start at the end, you can follow along in a completed version of the project on GitHub.

cd text–moderation/functions

Open index.js and delete its contents. In index.js, add the following code:

const functions = require(‘firebase-functions’); const toxicity = require(‘@tensorflow-models/toxicity’); exports.moderator = functions.database.ref(‘/messages/{messageId}’).onCreate(async (snapshot, context) => {   const message = snapshot.val();   // Verify that the snapshot has a value   if (!message) {      return;   }   functions.logger.log(‘Retrieved message content: ‘, message);   // Run moderation checks on the message and delete if needed.   const moderateResult = await moderateMessage(message.text);   functions.logger.log(     ‘Message has been moderated. Does message violate rules? ‘,     moderateResult   ); });

This code runs any time a message is added to the database. It gets the text of the message, and then passes it to a function called `moderateResult`. If you’re interested in learning more about Cloud Functions and the Realtime Database, then check out the Firebase documentation.

Add the Text Toxicity Classifier model

Depending on your development environment, you probably have some sort of error now since we haven’t actually written a function called moderateMessage yet. Let’s fix that. Below your Cloud Function trigger function, add the following code:

exports.moderator = functions.database.ref(‘/messages/{messageId}’).onCreate(async (snapshot, context) => {         //…         // Your other function code is here. }); async function moderateMessage(message) {   const threshold = 0.9;   let model = await toxicity.load(threshold);   const messages = [message];   let predictions = await model.classify(messages);   for (let item of predictions) {     for (let i in item.results) {       if (item.results[i].match === true) {         return true;       }     }   }   return false; }

This function does the following:

1. Sets the threshold for the model to 0.9. The threshold of the model is the minimum prediction confidence you want to use to set the model’s predictions to true or false–that is, how confident the model is that the text does or does not contain the given type of toxic content. The scale for the threshold is 0-1.0. In this case, I set the threshold to .9, which means the model will predict true or false if it is 90% confident in its findings.
2. Loads the model, passing the threshold. Once loaded, it sets toxicity_model to the model` value.
3. Puts the message into an array called messages, as an array is the object type that the classify function accepts.
4. Calls classify on the messages array.
5. Iterates through the prediction results. predictions is an array of objects each representing a different language label. You may want to know about only specific labels rather than iterating through them all. For example, if your use case is a website for hosting the transcripts of rap battles, you probably don’t want to detect and remove insults.
6. Checks if the content is a match for that label. if the match value is true, then the model has detected the given type of unwanted language. If the unwanted language is detected, the function returns true. There’s no need to keep checking the rest of the results, since the content has already been deemed inappropriate.
7. If the function iterates through all the results and no label match is set to true, then the function returns false – meaning no undesirable language was found. The match label can also be null. In that case, its value isn’t true, so it’s considered acceptable language. I will talk more about the null option in a future post.

If you completed part 1 of this tutorial, then these steps probably sound familiar. The server-side code is very similar to the client-side code. This is one of the things that I like about TensorFlow.js: it’s often straightforward to transition code from the client to server and vice versa.

Complete the Cloud Functions code

Back in your Cloud Function, you now know that based on the code we wrote for moderateMessage, the value of moderateResult will be true or false: true if the message is considered toxic by the model, and false if it does not detect toxicity with certainty greater than 90%. Now add code to delete the message from the database if it is deemed toxic:

// Run moderation checks on the message and delete if needed.   const moderateResult = await moderateMessage(message.text);   functions.logger.log(     ‘Message has been moderated. Does message violate rules? ‘,     moderateResult   );   if (moderateResult === true) {     var modRef = snapshot.ref;     try {       await modRef.remove();     } catch (error) {       functions.logger.error(‘Remove failed: ‘ + error.message);     }   }

This code does the following:

1. Checks if moderateResult is true, meaning that the message written to the guestbook is inappropriate.
2. If the value is true, it removes the data from the database using the remove function from the Realtime Database SDK.
3. Logs an error if one occurs.

Deploy the code

To deploy the Cloud Function, you can use the Firebase CLI. If you don’t have it, you can install it using the following npm command:

npm install –g firebase–tools

Once installed, use the following command to log in:

firebase login

Run this command to connect the app to your Firebase project:

firebase use —add

From here, you can select your project in the list, connect Firebase to an existing Google Cloud project, or create a new Firebase project.
Once the project is configured, use the following command to deploy your Cloud Function:

firebase deploy

Once deployment is complete, the logs include the link to your hosted guestbook. Write some guestbook entries. If you followed part 1 of the blog, you will need to either delete the moderation code from the website and deploy again, or manually add guestbook entries to the Realtime Database in the Firebase console.

You can view your Cloud Functions logs in the Firebase console.

Building on the example

I have a bunch of ideas for ways to build on this example. Here are just a few. Let me know which ideas you would like to see me build, and share your suggestions as well! The best ideas come from collaboration.

Get a queue

I mentioned that the “match” value of a language label can be true, false, or null without going into detail on the significance of the null value. If the label is null, then the model cannot determine if the language is toxic within the given threshold. One way to limit the number of null values is to lower this threshold. For example, if you change the threshold value to 0.8, then the model will label the match value as true if it is at least 80% certain that the text contains language that fits the label. My website example assigns labels of value null the same as those labeled false, allowing that text through the filter. But since the model isn’t sure if that text is appropriate, it’s probably a good idea to get some eyes on it. You could add these posts to a queue for review, and then approve or deny them as needed. I said “you” here, but I guess I mean “me”. If you think this would be an interesting use case to explore, let me know! I’m happy to write about it if it would be useful.

What’s in ‘store

The Firebase moderation sample that I used as the foundation of my project uses Realtime Database. I prefer to use Firestore because of its structure, scalability, and security. Firestore’s structure is well suited for implementing a queue because I could have a collection of posts to review within the collection of posts. If you’d like to see the website using Firestore, let me know.

Don’t just eliminate – moderate!

One of the things I like about the original Firebase moderation sample is that it sanitizes the text rather than just deleting the post. You could run text through the sanitizer before checking for toxic language through the text toxicity model. If the sanitized text is deemed appropriate, then it could overwrite the original text. If it still doesn’t meet the standards of decent discourse, then you could still delete it. This might save some posts from otherwise being deleted.

What’s in a name?

You’ve probably noticed that my moderation functionality doesn’t extend to the name field. This means that even a halfway-clever troll could easily get around the filter by cramming all of their expletives into that name field. That’s a good point and I trust that you will use some type of moderation on all fields that users interact with. Perhaps you use an authentication method to identify users so they aren’t provided a field for their name. Anyway, you get it: I didn’t add moderation to the name field, but in a production environment, you definitely want moderation on all fields.

Build a better fit

When you test out real-world text samples on your website, you might find that the text toxicity classifier model doesn’t quite fit your needs. Since each social space is unique, there will be specific language that you are looking to include and exclude. You can address these needs by training the model on new data that you provide.

If you enjoyed this article and would like to learn more about TensorFlow.js, then there are a ton of things you can do:

• Check out the TensorFlow.js edX course by Jason Mayes. If you are even remotely interested in using TensorFlow.js, I cannot recommend this enough. It might look like a lot at first, but the course is broken up into easy-to-follow manageable pieces.
• View all the TensorFlow.js pretrained models in the TFJS repository on GitHub.
• Play around with TensorFlow.js projects on Glitch.

Read More

Posted by Rostam Dinyari, Nitin Srinivasan, Douglas Yarrington and Rishika Sinha of the TensorFlow team

Starting with TensorFlow 2.10, we are excited to announce our collaboration with Intel, AWS, ARM, and Linaro to develop official TensorFlow builds. This means that when you pip install TensorFlow on Windows Native and Linux Aarch64 hosts, you will receive a build of TensorFlow that has been reviewed and vetted by these platform experts. This happens transparently, and there are no changes to your workflow . We’ve updated the pip install scripts so it’s automatic for you.

Official builds are TensorFlow releases that follow rigorous functional and performance testing standards Google engineers and our collaborators publish with each release, which we align with our published support expectations under the SIG Build forum. Collaborators monitor the builds daily and publish artifacts to the community in coordination with the overall TensorFlow release schedule.

For the majority of use cases, there will be no changes to the behavior of pip install or pip uninstall TensorFlow. However, for Windows Native and Linux Aarch64 based systems an additional pip uninstall step may be needed. You can find details about install, uninstall and other best practices on tensorflow.org/install/pip.

Over time, we expect the number of collaborators to expand but for now we want to share with you the progress we have made together to release increasingly performant and robust builds for these important platforms. You can learn more about each of the collaborations below.

Read More

Posted by Bernhard Bauer and Terry Heo, Software Engineers, Google

Today we’re excited to announce that the Google Play services API for TensorFlow Lite is generally available on Android devices. We recommend this distribution as the path to adding custom machine learning to your apps. Last year, we launched a public beta of TensorFlow Lite in Google Play services at Google I/O. Since then, we’ve received lots of feedback and made improvements to the API. Most recently, we added the GPU delegate and Task Library support. Today we’re moving from beta to general availability on billions of Android devices globally.

TensorFlow Lite in Google Play services is already used by Google teams, including ML Kit, serving over a billion monthly active users and running more than 100 billion daily inferences.

TensorFlow Lite is an inference runtime optimized for mobile devices, and now that it’s part of Google Play services, it helps you deliver better ML experiences because it:

• Reduces your app size by up to 5 MB compared to statically bundling TensorFlow Lite with your app
• Uses the same API as available when bundling TF Lite into your app
• Receives regular performance updates in the background so it’s always getting better automatically

Get started by learning how to add TensorFlow Lite in Google Play Services to your Android app.Read More

Posted by the TensorFlow Team

TensorFlow 2.10 has been released! Highlights of this release include user-friendly features in Keras to help you develop transformers, deterministic and stateless initializers, updates to the optimizers API, and new tools to help you load audio data. We’ve also made performance enhancements with oneDNN, expanded GPU support on Windows, and more. This release also marks TensorFlow Decision Forests 1.0! Read on to learn more.

Keras

Expanded, unified mask support for Keras attention layers

Starting from TensorFlow 2.10, mask handling for Keras attention layers, such as tf.keras.layers.Attention, tf.keras.layers.AdditiveAttention, and tf.keras.layers.MultiHeadAttention have been expanded and unified. In particular, we’ve added two features:

Causal attention: All three layers now support a use_causal_mask argument to call (Attention and AdditiveAttention used to take a causal argument to __init__).

Implicit masking: Keras Attention, AdditiveAttention, and MultiHeadAttention layers now support implicit masking  (set mask_zero=True in tf.keras.layers.Embedding).

Combined, this simplifies the implementation of any Transformer-style model since getting the masking right is often a tricky part.

A basic Transformer self-attention block can now be written as:

import tensorflow as tf

embedding = tf.keras.layers.Embedding(
    input_dim=10,
    output_dim=3,
    mask_zero=True) # Infer a correct padding mask.

# Instantiate a Keras multi-head attention (MHA) layer,
# a layer normalization layer, and an `Add` layer object.
mha = tf.keras.layers.MultiHeadAttention(key_dim=4, num_heads=1)
layernorm = tf.keras.layers.LayerNormalization()
add = tf.keras.layers.Add()

# Test input.
x = tf.constant([[1, 2, 3, 4, 5, 0, 0, 0, 0],
                 [1, 2, 1, 0, 0, 0, 0, 0, 0]])
# The embedding layer sets the mask.
x = embedding(x)

# The MHA layer uses and propagates the mask.
a = mha(query=x, key=x, value=x, use_causal_mask=True)
x = add([x, a]) # The `Add` layer propagates the mask.
x = layernorm(x)

# The mask made it through all layers.
print(x._keras_mask)

And here’s the output: 

Try out the new Keras Optimizers API

In the previous release, Tensorflow 2.9, we published a new version of the Keras Optimizer API, in tf.keras.optimizers.experimental, which will replace the current tf.keras.optimizers namespace in TensorFlow 2.11. To prepare for the upcoming formal switch of the optimizer namespace to the new API, we’ve also exported all of the current Keras optimizers under tf.keras.optimizers.legacy in TensorFlow 2.10.

Most users won’t be affected by this change, but please check the API doc to see if any API used in your workflow has changed. If you decide to keep using the old optimizer, please explicitly change your optimizer to corresponding tf.keras.optimizers.legacy.Optimizer.

You can also find more details about new Keras Optimizers in this article.

Deterministic and Stateless Keras initializers

In TensorFlow 2.10, we’ve made Keras initializers (the tf.keras.initializers API) stateless and deterministic, built on top of stateless TF random ops. Starting in TensorFlow 2.10, both seeded and unseeded Keras initializers will always generate the same values every time they are called (for a given variable shape). The stateless initializer enables Keras to support new features such as multi-client model training with DTensor.

For unseeded initializers (seed=None), a random seed will be created and assigned at initializer creation (different initializer instances get different seeds). An unseeded initializer will raise a warning if it is reused (called) multiple times. This is because it would produce the same values each time, which may not be intended.

BackupAndRestore checkpoints with step level granularity

In the previous release, Tensorflow 2.9, the tf.keras.callbacks.BackupAndRestore Keras callback would backup the model and training state at epoch boundaries. In Tensorflow 2.10, the callback can also backup the model every N training steps. However, keep in mind that when BackupAndRestore is used with tf.distribute.MultiWorkerMirroredStrategy, the distributed dataset iterator state will be reinitialized and won’t be restored when restoring the model. More information and code examples can be found in the migrate the fault tolerance mechanism guide.

Easily generate an audio classification dataset from a directory of audio files

You can now use a new utility, tf.keras.utils.audio_dataset_from_directory, to easily generate audio classification datasets from directories of .wav files. Just sort your audio files into one different directory per file class, and a single line of code will get you a labeled tf.data.Dataset you can pass to a Keras model. You can find an example here.

The EinsumDense layer is no longer experimental

The einsum function is the swiss army knife of linear algebra. It can efficiently and explicitly describe a wide variety of operations. The tf.keras.layers.EinsumDense layer brings some of that power to Keras.

Operations like einsum, einops.rearrange, and the EinsumDense layer operate based on a string “equation” that describes the axes of the inputs and outputs. For EinsumDense the equation lists the axes of the input argument, the axes of the weights, and the axes of the output. A basic Dense layer can be written as: 

Notes:

• …i – This only works on the last axis of the input, that axis is called i.
• ij – The weights are a matrix with shape (ij).
• …j – The result sums out the i axis and leaves j.

For example, here is a stack of 5 Dense layers with 10 units each:

Here is a stack of Dense layers, where each one operates on a different input vector:

Here is a stack of Dense layers where each one operates on each input vector independently:

Performance and collaborations

Improved aarch64 CPU performance: ACL/oneDNN integration

We have worked with Arm, AWS, and Linaro to integrate Compute Library for the Arm® Architecture (ACL) with TensorFlow through oneDNN to accelerate performance on aarch64 CPUs. Starting with TensorFlow 2.10, you can try these experimental optimizations by setting the environment variable TF_ENABLE_ONEDNN_OPTS=1 before running your TensorFlow program.

There may be slightly different numerical results due to different computation and floating-point round-off approaches. If this causes issues for you, turn the optimizations off by setting TF_ENABLE_ONEDNN_OPTS=0 before running your program.

To verify that the optimizations are on, look for a message beginning with “oneDNN custom operations are on” in your program log. We welcome feedback on GitHub and the TensorFlow Forum.

Expanded GPU support on Windows

TensorFlow can now leverage a wider range of GPUs on Windows through the TensorFlow-DirectML plug-in. To enable model training on DirectX 12-capable GPUs from vendors such as AMD, Intel, NVIDIA, and Qualcomm, install the plug-in alongside standard TensorFlow CPU packages on native Windows or WSL2. The preview package currently supports a limited number of basic machine learning models, with a goal to increase model coverage in the future. You can view the open-source code and leave feedback at the TensorFlow-DirectML GitHub repository.

New features in tf.data

Create tf.data Dataset from lists of elements

Tensorflow 2.10 introduces a convenient new experimental API tf.data.experimental.from_list which creates a tf.data.Dataset comprising the given list of elements. The returned dataset will produce the items in the list one by one. The functionality is identical to tf.data.Dataset.from_tensor_slices when elements are scalars, but different when elements have structure.

Consider the following example:

In contrast, to get the same output with `from_tensor_slices`, the data needs to be reorganized:

Unlike the from_tensor_slices method, from_list supports non-rectangular input (achieving the same with from_tensor_slices requires the use of ragged tensors).

Sharing tf.data service with concurrent trainers

If you run multiple trainers concurrently using the same training data, it could save resources to cache the data in one tf.data service cluster and share the cluster with the trainers. For example, if you use Vizier to tune hyperparameters, the Vizier jobs can run concurrently and share one tf.data service cluster.

To enable this feature, each trainer needs to generate a unique trainer ID, and you pass the trainer ID to tf.data.experimental.service.distribute. Once a job has consumed the data, the data remains in the cache and is re-used by jobs with different trainer_ids. Requests with the same trainer_id do not re-use data. For example:

 tf.data service uses a sliding-window cache to store shared data. When one trainer consumes data, the data remains in the cache. When other trainers need data, they can get data from the cache instead of repeating the expensive computation. The cache has a bounded size, so some workers may not read the full dataset. To ensure all the trainers get sufficient training data, we require the input dataset to be infinite. This can be achieved, for example, by repeating the dataset and performing random augmentation on the training instances.

TensorFlow Decision Forests 1.0

In conjunction with the release of Tensorflow 2.10, Tensorflow Decision Forests (TF-DF) reaches version 1.0. With this milestone we want to communicate more broadly that Tensorflow Decision Forests has become a more stable and mature library. We’ve improved our documentation and established more comprehensive testing to make sure that TF-DF is ready for professional environments.

The new release of TF-DF also offers a first look at the Javascript and Go APIs for inference of TF-DF models. While these APIs are still in beta, we are actively looking for feedback for them. TF-DF 1.0 improves performance of oblique splits. Oblique splits allow decision trees to express more complex patterns by conditioning on multiple features at the same time – learn more in our Decision Forests class on developers.google.com. Benchmarks and real-world observations show that oblique splits outperform classical axis-aligned splits on the majority of datasets. Finally, the new release includes our latest bug fixes.

Next steps

Check out the release notes for more information. To stay up to date, you can read the TensorFlow blog, follow twitter.com/tensorflow, or subscribe to youtube.com/tensorflow. If you’ve built something you’d like to share, please submit it for our Community Spotlight at goo.gle/TFCS. For feedback, please file an issue on GitHub or post to the TensorFlow Forum. Thank you!

Read More

Posted by  Andreas Steiner and Marc van Zee, Google Research, Brain Team

Introduction

In this blog post we demonstrate how to convert and run Python-based JAX functions and Flax machine learning models in the browser using TensorFlow.js. We have produced three examples of JAX-to-TensorFlow.js conversion each with increasing complexity: 

1. A simple JAX function 
2. An image classification Flax model trained on the MNIST dataset 
3. A full image/text Vision Transformer (ViT) demo, which was used for the Google AI blog post Locked-Image Tuning: Adding Language Understanding to Image Models (a preview of the demo is shown in Figure 1 below)

For each example, there are Google Colab notebooks you can use to try the JAX-to-TensorFlow.js conversion yourself.

Figure 1. TensorFlow.js model matching user-provided text prompts to a precomputed image embedding (try it out yourself). See Part 3: LiT Demo below for implementation details.

Background: JAX and TensorFlow.js

JAX is a NumPy-like library developed by Google Research for high performance computing. It uses XLA to compile programs optimized for GPUs and TPUs. Flax is a popular neural network library built on top of JAX. Researchers have been using JAX/Flax to train very large models with billions of parameters (such as PaLM for language understanding and generation, or Imagen for image generation), making full use of modern hardware. If you’re new to JAX and Flax, start with this JAX 101 tutorial and this Flax Getting Started example.

TensorFlow started as a library for ML towards the end of 2015 and has since become a rich ecosystem that includes tools for productionizing ML pipelines (TFX), data visualization (TensorBoard), deploying ML models to edge devices (TensorFlow Lite), and devices running on a web browser or any device capable of executing JavaScript (TensorFlow.js). Models developed in JAX or Flax can tap into this rich ecosystem by first converting such a model to the TensorFlow SavedModel format, and then using the same tooling as if they had been developed in TensorFlow natively.

This is now made even easier for TensorFlow.js through the new Python API — tfjs.converters.convert_jax() — which allows users to convert a JAX model written in Python to a web format (.json) directly, so that the model can be used in the browser with Tensorflow.js.

To learn how to perform JAX-to-TensorFlow.js conversion, check out the three examples below.

Example 1: Converting a simple JAX function

In this introductory example, you’ll convert a few simple JAX functions using converters.convert_jax().

Internally, this function does the following:

1. It converts to the Tensorflow SavedModel format, which contains a complete TensorFlow program, including trained parameters (i.e., tf.Variables) and computation.
2. Then, it constructs a TensorFlow.js model from that SavedModel (refer to Figure 2 for more details).

Figure 2. High-level visualization of the conversion steps inside jax_conversion.from_jax, which converts a JAX function to a Tensorflow.js model.

To convert a Flax model to TensorFlow.js, you need a few things:

• A function that runs the forward pass of the model.
• The model parameters (this is usually a dict-like structure).
• A specification of the shapes and dtypes of the inputs to the function.

The following examples uses a single parameter weight and implements a function prod, which multiplies the input with the parameter (in a real example, params will contain the all weights of the modules used in the neural network):

Let’s call this function with some values and verify the output makes sense:

This yields a state.apply_fn that can be used to compute logits for input images. Note that the function expects the first argument to be the model weights state.params. Given a batch of input images shaped [batch_size, 28, 28, 1], this will produce the logits for the probability distribution over the ten labels for every model (shaped [batch_size, 10]).

On the JavaScript side, you load the model asynchronously, showing a simple progress update in the status text, making sure to give some feedback while the model weights are transferred:

A minimal UI is loaded from this snippet, and in the callback function you call the TensorFlow.js model and output the predictions. The function parameter img is a Uint8Array of length 28*28, which is first converted to a TensorFlow.js tf.tensor, before computing the model outputs, and converting them to probabilities via the tf.softmax() function. The output values from the computation are then waited for synchronously by calling .dataSync(), and converted to JavaScript arrays before they’re displayed.

The Colab then starts a webserver and tunnels the port so you can scan a QR code on a mobile phone and directly connect to the demo. Even though the training reports around 99.1% accuracy on the test set, you’ll see that the model can easily be fooled with digits that are easy to recognize for the human eye, but hard for a model that has only seen digits from the MNIST dataset (Figure 3).

Figure 3. Our model from the Colab with 99.1% accuracy on the MNIST test dataset is still surprisingly bad at recognizing hand-written digits. On the left, the model predicts all kinds of digits instead of “one”. On the right side, the “one” is drawn more like the data from the training set. Example 3: LiT Demo Writing a more realistic application with a TensorFlow.js model is a bit more involved. This section goes through the main steps that were used to create the demo app from the Google AI blog post Locked-Image Tuning: Adding Language Understanding to Image Models. Refer to that post for technical details on the implementation of the ML model. Also make sure to check out the final LiT Demo. Adapting the model Before starting to implement an ML demo, it’s a good moment to think carefully about the different options and their respective strengths and weaknesses. At a high level, you have two options: running the ML model on server-side infrastructure, or running the ML model on the edge (i.e. on the visiting user’s device). Running a model on a server has the advantage that it can use exactly the same framework / code that was used to develop the model. There are libraries like Streamlit or Gradio that make it very easy to quickly build interactive web apps around such centrally-hosted models. The servers running the model can be rather powerful, using lots of RAM and accelerators to run state-of-the-art ML models in near-real time, and such a website can be loaded even by the smallest mobile device. Running the demo on-device puts a limit on the size of the model that you can use, but comes with convincing advantages: No data is ever sent off the device, which is desirable both for privacy reasons and to bring down latency. Free scaling: For instance, a normal webserver (such as one running on GitHub Pages) can serve hundreds or thousands of users simultaneously free of charge. And running a powerful model on server-side infrastructure at this scale would be very expensive (massive compute is not cheap). The model you use for the demo consists of two parts: an image encoder, and a text encoder (see Figure 4). For computing image embeddings you use a large model, and for text embeddings—a small model. To make the demo run faster and produce better results, the expensive image embeddings are pre-computed, so the Tensorflow.js model only needs to compute the text embeddings and then compare the image and text embeddings to compute similarities. Figure 4. Image/text models like LiT (or CLIP) consist of two encoders that can be used separately to create vector representations of images and texts. Usually both image and text encoders are of similar size (LiT-B16B model, left image). For the demo, we precompute image embeddings using a large image encoder, and then run inference on the text on-device using a tiny text encoder (LiT-L16Ti model, right image). For the demo, we now get those powerful ViT-Large image representations for free, because we can precompute them for all demo images. This allows us to make for a compelling demo with a limited compute budget. In addition to the “tiny” text encoder, we have also prepared a “small” text encoder for the same image embeddings (LiT-L16S), which performs a bit better, but uses more bandwidth to download the model weights, and requires more GPU memory to run on-device. We have evaluated the different models with the code from this Colab: Image encoder Text encoder Zeroshot performance Model Params FLOPs Params FLOPs CIFAR-100 ImageNet LiT-B16B 86M (344 MB) 36B 109M (436 MB) 2.7B 79.2% 71.7% LiT-L16S  (“small” text encoder) 303M (1.2 GB) 123B 28M (111 MB) 0.7B 75.8% 60.7% LiT-L16Ti (“tiny” text encoder) 303M (1.2 GB) 123B 9M (36 MB) 0.2B 73.2% 53.4% Note though that the “zeroshot performance” should only be taken as a proxy. In the end, the model performance needs to be good enough for the demo, and in this case our manual testing showed that even the tiny text transformer was able to compute similarities good enough for the demo. Next, we tested the performance of the tiny and small text encoders using this TensorFlow.js benchmark tool on different platforms (using the “custom model” option, and benchmarking 5×16 tokens on the WebGL backend): LiT-L16T (“tiny” text encoder) – benchmark LiT-L16S (“small” text encoder) – benchmark Load time Warmup Average/10 Peak memory Load time Warmup Average/10 Peak memory MacBook Pro (Intel i7 2.6GHz / Radeon Pro 5300M) 1.1s 0.15s 0.12s 33.9 MB 3.9s 0.8s 0.8s 122 MB iPad Air (4th gen) 1.3s 0.6s 0.5s 33.9 MB 2.7s 2.4s 2.5s 141 MB Samsung S21 G5 (cell phone) 2.0s 1.3s 1.1s 33.9 MB – – – – Note that the results for the model with the “small” text encoder are missing for “Samsung S21 G5” in the above table because the model did not fit into memory. In terms of performance, the model with the “tiny” text encoder produces results within approximately 0.1-1 seconds, which still feels quite responsive, even on the smallest platform tested. The Lit-LiT web app  Preparing the model for this application is a bit more complicated, because we need not only convert the text transformer model weights, but also a matching tokenizer, and the precomputed image embeddings. The Colab loads a LiT model and showcases how to use it, and then prepares contents needed by the web app: The tiny/small text encoder converted to TensorFlow.js and the matching tokenizer vocabulary. Images in JPG format, as seen by the model (in particular, this means a fixed 224×224 pixel crop) Pre-computed image embeddings (since the converted model will only be able to compute embeddings for the texts). A selection of example prompts for every image. The embeddings of these prompts are also precomputed to allow to show precomputed answers if the prompts are not modified. These files are prepared inside the data/ directory and then downloaded as a ZIP file. This file can then be uploaded to a web hosting, from where it is loaded by the web app (for example on GitHub Pages: vision_transformer/lit/data). The code for the entire client-side application is available on Github: https://github.com/google-research/big_vision/tree/main/ui/lit_demo/ The application is built using Lit web components. The main index.html declares the demo application:

The parent directory of the data/ exported by the Colab above is referenced via the baseUrl in the file src/lit/constants.ts. By default it refers to the models from the official demo. When replacing the baseUrl with a different server, make sure to enable cross origin resource sharing.

In addition to the complete application, it’s also possible to export the functional parts without the UI as a single JavaScript file that can be linked statically. See the file playground.html as an example, and refer to the instructions in README.md for how to compile the entire application or the functional part before deploying the application.

Read More

Posted by Jen Person, Developer Advocate

Being kind: a perennial problem

I’ve often wondered why anonymity drives people to say things that they’d never dare say in person, and it’s unfortunate that comment sections for videos and articles are so often toxic! If you’re interested in content moderation, you can use machine learning to help detect toxic posts which you consider for removal.

ML for web developers

Machine learning is a powerful tool for all sorts of natural language-processing tasks, including translation, sentiment analysis, and predictive text. But perhaps it feels outside the scope of your work. After all, when you’re building a website in JavaScript, you don’t have time to collect and validate data, train a model using Python, and then implement some backend in Python on which to run said model. Not that there’s anything wrong with Python–it’s just that, if you’re a web developer, it’s probably not your language of choice.

Fortunately, TensorFlow.js allows you to run your machine learning model on your website in everybody’s favorite language: JavaScript. Furthermore, TensorFlow.js offers several pre-trained models for common use cases on the web. You can add the power of ML to your website in just a few lines of code! There is even a pre-trained model to help you moderate written content, which is what we’re looking at today.

The text toxicity classifier ML model

There is an existing pretrained model that works well for content moderation: the TensorFlow.js text toxicity classifier model. With this model, you can evaluate text on different labels of unwanted content, including identity attacks, insults, and obscenity. You can try out the demo to see the classifier in action. I admit that I had a bit of fun testing out what sort of content would be flagged as harmful. For example:

I recommend stopping here and playing around with the text toxicity classifier demo. It’s a good idea to see what categories of text the model checks for and determine which ones you would want to filter from your own website. Besides, if you want to know what categories the above quote got flagged for, you’ll have to go to the demo to read the headings.

Once you’ve hurled sufficient insults at the text toxicity classifier model, come back to this blog post to find out how to use it in your own code.

A dual approach

This started as a single tutorial with client and server-side code, but it got a bit lengthy so I decided to split it up. Separating the tutorials also makes it easier to target the part that interests you if you just want to implement one part. In this post, I cover the implementation steps for client-side moderation with TensorFlow.js using a basic website. In part 2, I show how to implement the same model server-side using Cloud Functions for Firebase.

Client-side moderation

Moderating content client-side provides a quicker feedback loop for your users, allowing you to stop harmful discourse before it starts. It can also potentially save on backend costs since inappropriate comments don’t have to be written to the database, evaluated, and then subsequently removed.

Starter code

I used the Firebase text moderation example as the foundation of my demo website. It looks like this:

Keep in mind TensorFlow.js doesn’t require Firebase. You can use whatever hosting, database, and backend solutions that work best for your app’s needs. I just tend to use Firebase because I’m pretty familiar with it already. And quite frankly, TensorFlow.js and Firebase work well together! The website in the Firebase demo showcases content moderation through a basic guestbook using a server-side content moderation system implemented through a Realtime Database-triggered Cloud Function. Don’t worry if this sounds like a lot of jargon. I’ll walk you through the specifics of what you need to know to use the TensorFlow.js model in your own code. That being said, if you want to build this specific example I made, it’s helpful to take a look at the Firebase example on GitHub.

If you’re building the example with me, clone the Cloud Functions samples repo. Then change to the directory of the text moderation app.

This project requires you to have the Firebase CLI installed. If you don’t have it, you can install it using the following npm command:

Once installed, use the following command to log in:

Run this command to connect the app to your Firebase project:

From here, you can select your project in the list, connect Firebase to an existing Google Cloud project, or create a new Firebase project. Once the project is configured, use the following command to deploy Realtime Database security rules and Firebase Hosting:

There is no need to deploy Cloud Functions at this time since we will be changing the sample code entirely.

Note that the Firebase text moderation sample as written uses the Blaze (pay as you go) plan for Firebase. If you choose to follow this demo including the server-side component, your project might need to be upgraded from Spark to Blaze. If you have a billing account set on your project through Google Cloud, you are already upgraded and good to go! Most importantly, if you’re not ready to upgrade your project, then do not deploy the Cloud Functions portion of the sample. You can still use the client-side moderation without Cloud Functions.

To implement client-side moderation in the sample, I added some code to the index.html and main.js files in the Firebase text moderation example. There are three main steps to implement when using a TensorFlow.js model: installing the required components, loading the model, and then running the prediction. Let’s add the code for each of these steps.

Install the scripts

Add the required TensorFlow.js dependencies. I added the dependencies as script tags in the HTML, but you can use Node.js if you use a bundler/transpiler for your web app.

Load the model

Add the following code to load the text toxicity model in the Guestbook() function. The Guestbook() function is part of the original Firebase sample. It initializes the Guestbook components and is called on page load.

The threshold of the model is the minimum prediction confidence you want to use to set the model’s predictions to true or false–that is, how confident the model is that the text does or does not contain the given type of toxic content. The scale for the threshold is 0-1.0. In this case, I set the threshold to .9, which means the model will predict true or false if it is 90% confident in its findings. It is up to you to decide what threshold works for your use case. You may even want to try out the text toxicity classifier demo with some phrases that could come up on your website to determine how the model handles them.

toxicity.load loads the model, passing the threshold. Once loaded, it sets toxicity_model to the model value.

Run the prediction

Add a checkContent function that runs the model predictions on messages upon clicking “Add message”:

This function does the following:

1. Verifies that the model load has completed. If toxicity_model has a value, then the load() function has finished loading the model.
2. Puts the message into an array called messages, as an array is the object type that the classify function accepts.
3. Calls classify on the messages array.
4. Iterates through the prediction results. predictions is an array of objects each representing a different language label. You may want to know about only specific labels rather than iterating through them all. For example, if your use case is a website for hosting the transcripts of rap battles, you probably don’t want to detect and remove insults.
5. Checks if the content is a match for that label. if the match value is true, then the model has detected the given type of unwanted language. If the unwanted language is detected, the function returns true. There’s no need to keep checking the rest of the results, since the content has already been deemed inappropriate.
6. If the function iterates through all the results and no label match is set to true, then the function returns false – meaning no undesirable language was found. The match label can also be null. In that case, its value isn’t true, so it’s considered acceptable language. I will talk more about the null option in a future post.

Add a call to the checkContent in the saveMessage function:

After a couple quick checks for input values, the contents of the message box is passed to the checkContent function.

If the content passes this check, the message is written to the Realtime Database. If not, a snack bar displays reminding the message author to be kind. The snack bar isn’t anything special, so I’m not going to include the code here. You can see it in the full example code, or implement a snack bar of your own.

Try it out

If you’ve been following along in your own code, run this terminal command in your project folder to deploy the website:

A message that’s not acceptable gets rejected An acceptable message gets published to the guestbook

Verifying that this code was working properly was really uncomfortable. I had to come up with an insult that the model would deem inappropriate, and then keep writing it on the website. From my work computer. I know nobody could actually see it, but still. That was one of the stranger parts of my job, to be sure! Next steps

Using client-side moderation like this could catch most issues before they occur. But a clever user might open developer tools and try to find a way to write obscenities directly to the database, circumventing the content check. That’s where server-side moderation comes in.

If you enjoyed this article and would like to learn more about TensorFlow.js, here are some things you can do:

• Check out the TensorFlow.js edX course by Jason Mayes. If you are even remotely interested in using TensorFlow.js, I cannot recommend this enough. It might look like a lot at first, but the course is broken up into easy-to-follow manageable pieces. 
• View all the TensorFlow.js pretrained models in the TFJS repository on GitHub. 
• Play around with TensorFlow.js projects on Glitch. 
• To see an example of ML image moderation on the web, try out Gant Laborde’s NSFW TFJS image checker.

Read More

A guest post by Dinko Franceschi, Broad Institute of MIT and Harvard

Kaggle has become the go-to place to practice data science skills and participate in machine learning model-building competitions. This tutorial will provide an easy-to-follow walkthrough of how to get started with a Kaggle notebook using TensorFlow Decision Forests. It’s a library that allows you to train tree-based models (like random forests and gradient-boosted trees) in TensorFlow.

Why should you be interested in decision forests? There are roughly two types of Kaggle competitions – and the winning solution (neural networks or decision forests) depends on the kind of data you’re working with.

If you’re working with a tabular data problem (these involve training a model to classify data in a spreadsheet which is an extremely common scenario) – the winning solution is often a decision forest. However, if you’re working with a perception problem that involves teaching a computer to see or hear (for example, image classification), the winning model is usually a neural network.

Here’s where the good news starts. You can implement a decision forest in TensorFlow with just a few lines of code. This relatively simple model often outperforms a neural network on many Kaggle problems.

We will explore the decision forests library with a simple dataset from Kaggle, and we will build our model with Kaggle Kernels which allow you to completely build and train your models online using free cloud compute power – similar to Colab. The dataset contains vehicle information such as cost, number of doors, occupancy, and maintenance costs which we will use to assign an evaluation on the car.

Kaggle Kernels can be accessed through your Kaggle account. If you do not have an account, please begin by signing up. On the home page, select the “Code” option on the left menu and select “New Notebook,” which will open a new Kaggle Kernel.

Once we have opened a new notebook from Kaggle Kernels, we download the car evaluation dataset to our environment. Click “Add data” near the top right corner of your notebook, search for “car evaluation,” and add the dataset.

Now we are ready to start writing code. Install the TensorFlow Decision Forests library and the necessary imports, as shown below. The code in this blog post has been obtained from the Build, train and evaluate models with the TensorFlow Decision Forests tutorial which contains additional examples to look at.

!pip install tensorflow_decision_forests
import numpy as np

import pandas 

import tensorflow_decision_forests as tfdf

We will now import the dataset. We should note that the dataset we downloaded did not contain headers, so we will add those first based on the information provided on the Kaggle page for the dataset. It is good practice to inspect your dataset before you start working with it by opening it up in your favorite text or spreadsheet editor.

df = pandas.read_csv("../input/car-evaluation-data-set/car_evaluation.csv")
col_names =['buying price', 'maintenance price', 'doors', 'persons', 'lug_boot', 'safety', 'class']

df.columns = col_names

df.head()

We must then split the dataset into train and test:

def split_dataset(dataset, test_ratio=0.30):
  test_indices = np.random.rand(len(dataset)) < test_ratio

  return dataset[~test_indices], dataset[test_indices]



train_ds_pd, test_ds_pd = split_dataset(df)
print("{} examples in training, {} examples for testing.".format(

    len(train_ds_pd), len(test_ds_pd)))

And finally we will convert the dataset into tf.data format. This is a high-performance format that is used by TensorFlow to train models more efficiently, and with TensorFlow Decision Forests, you can convert your dataset to this format with one line of code:

train_ds = tfdf.keras.pd_dataframe_to_tf_dataset(train_ds_pd, label="class")

test_ds = tfdf.keras.pd_dataframe_to_tf_dataset(test_ds_pd, label="class")

Now you can go ahead and train your model right away by executing the following:

model = tfdf.keras.RandomForestModel()
model.fit(train_ds)

The library has good defaults which are a fine place to start for most problems. For advanced users, there are lots of options to choose from in the API doc as random forests are configurable.

Once you have trained the model, you can see how it will perform on the test data.

model.compile(metrics=["accuracy"])
print(model.evaluate(test_ds))

In just a few lines of code, you reached an accuracy of >95% on this small dataset! This is a simple dataset, and one might argue that neural networks could also yield impressive results. And they absolutely can (and do), especially when you have very large datasets (think: hundreds of thousands of examples, or more). However, neural networks require more code and are resource intensive as they require significantly more compute power.

Easy preprocessing

Decision forests have another important advantage: there are fewer steps to preprocess the data. Notice in the code above that you were able to pass a dataset with both categorical and numeric values directly to the decision forests. You did not have to do any preprocessing like normalizing numeric values, converting strings to integers, and one-hot encoding them. This has major benefits. It makes decision forests simpler to work with (so you can train a model quickly), and there is less code that can go wrong.

Below, you will see some important differences between the two techniques.

Easy to interpret

A significant advantage of decision forests is that they are easy to interpret. While the pipeline for decision trees differs significantly from that of training neural networks, there are major advantages for selecting these models for a given task. This is because feature importance is particularly straightforward to determine with decision forests (ensemble of decision trees). Notably, the TensorFlow Decision Forests library makes it possible to visualize feature importance with its model plotter function. Let’s see below how this works!

tfdf.model_plotter.plot_model_in_colab(model, tree_idx=0)

We see in the root of the tree on the left the number of examples (1728) and the corresponding distribution indicated by the different colors. Here our model is looking at the number of persons that the car can fit. The largest section indicated by green stands for 2 persons and the red for 4 persons. Furthermore, as we go down the tree we continue to see how the tree splits and the corresponding number of examples. Based on the condition, examples are branched to one of two paths. Interestingly, from here we can also determine the importance of a feature by examining all of the splits of a given feature and then computing how much this feature lowered the variance.

Decision Trees vs. Neural Networks

Neural networks undoubtedly have incredible representation learning capabilities. While they are very powerful in this regard, it is important to consider whether they are the right tool for the problem at hand. When working with neural networks, one must think a lot about how they will construct the layers. In contrast, decision forests are ready to go out of the box (of course, advanced users can tune a variety of parameters).

Prior to even building a neural network layer by layer, in most cases one must perform feature pre-processing. For example, this could include normalizing the features to have mean around 0 and standard deviation of 1 and converting strings to numbers. This initial step can be skipped right away with Tree-based models which natively handle mixed data.

As seen in the code above, we were able to obtain results in just a few steps. Once we have our desired metrics, we have to interpret them within the context of our problem. Perhaps one of the most significant strengths of Decision Trees is their interpretability. We see in the code above the diagrams that were outputted. Starting at the root, we can follow the branches and quickly get a good idea of how the model made its decisions. In contrast, neural networks are a “black box” that can be difficult to interpret and to explain to a non-technical audience.

Learning more

If you’d like to learn more about TensorFlow Decision Forests, the best place to start is with the project homepage. You can also check out this previous article for more background. And if you have any questions or feedback, the best place to ask them is on https://discuss.tensorflow.org/ using the tag “tfdf”. Thanks for reading!

Read More

Posted by Chansung Park and Sayak Paul (ML-GDEs)

In this post, we’ll share the lessons and findings learned from conducting load tests for an image classification model across numerous deployment configurations. These configurations involve REST-based deployments with TensorFlow Serving. In this way, we aim to equip the readers with a holistic understanding of the differences between the configurations.

This post is less about code and more about the architectural decisions we had to make for performing the deployments. We’ll first provide an overview of our setup including the technical specifications. We’ll also share our commentaries on the design choices we made and their impact.

Technical Setup

TensorFlow Serving is feature-rich and has targeted specifications embedded in its designs (more on this later). For online prediction scenarios, the model is usually exposed as some kind of service.

To perform our testing we use a pre-trained ResNet50 model which can classify a variety of images into different categories. We then serve this model in the following way:

Docker to containerize the environment.

Kubernetes to orchestrate a cluster of container nodes for scalability. We use Kubernetes Engine (GKE) to manage this.

GitHub Actions to automatically roll out deployments on GKE.

Our deployment platform (nodes on the Kubernetes Cluster) is CPU-based. We don’t employ GPUs at any stage of our processes. For this purpose, we can build a CPU-optimized TensorFlow Serving image and take advantage of a few other options which can reduce the latency and boost the overall throughput of the system. We will discuss these later in the post.

You can find all the code and learn how the deployments were performed in this repository. Here, you’ll find example notebooks and detailed setup instructions for playing around with the code. As such, we won’t be discussing the code line by line but rather shed light on the most important parts when necessary.

Throughout the rest of this post, we’ll discuss the key considerations for the deployment experiments respective to TensorFlow Serving including its motivation, limitations, and our experimental results.

With the emergence of serverless offerings like Vertex AI, it has never been easier to deploy models and scale them securely and reliably. These services help reduce the time-to-market tremendously and increase overall developer productivity. That said, there might still be instances where you’d like more granular control over things. This is one of the reasons why we wanted to do these experiments in the first place.

Considerations

TensorFlow Serving has its own sets of constraints and design choices that can impact a deployment. In this section, we provide a concise overview of these considerations.

Deployment infrastructure: We chose GKE because Kubernetes is a standard deployment platform when using GCP, and GKE lets us focus on the ML parts without worrying about the infrastructure since it is a fully managed Google Cloud Platform service. Our main interest is in how to deploy models for CPU-based environments, so we have prepared a CPU-optimized TensorFlow Serving image.

Trade-off between more or fewer servers: We started experiments for TensorFlow Serving setups with the simplest possible VMs equipped with 2vCPU and 4GB RAM, then we gradually upgraded the specification up to 8vCPU and 64GB RAM. On the other hand, we decreased the number of nodes in the Kubernetes cluster from 8 to 2 because it is a trade-off between costs to deploy cheaper servers versus fewer expensive servers.

Options to benefit multi-core environments: We wanted to see if high-end VMs can outperform simple VMs with options to take advantage of the multi-core environment even though there are fewer nodes. To this end, we experimented with a different number inter_op_parallelism and intra_op_parallelism threads for TensorFlow Serving deployment set according to the number of CPU cores.

Dynamic batching and other considerations: Modern ML frameworks such as TensorFlow Serving usually support dynamic batching, initial model warm-up, multiple deployments of multiple versions of different models, and more out of the box. For our purpose of online prediction, we have not tested these features carefully. However, dynamic batching capability is also worth exploring to enhance the performance according to the official document. We have seen that the default batching configuration could reduce the latency a little even though the results of that are not included in this blog post.

Experiments

We have prepared the following environments. In TensorFlow Serving, the number of intra_op_parallelism_threads is set equal to the number of CPU cores while the number of inter_op_parallelism_threads is set from 2 to 8 for experimental purposes as it controls the number of threads to parallelize the execution of independent operations. Below we provide the details on the adjustments we performed on the number of vCPUs, RAM size, and the number of nodes for each Kubernetes cluster. Note that the number of vCPUs and the RAM size are applicable for the cluster nodes individually.

The load tests are conducted using Locust. We have run each load test for 5 minutes. The number of requests are controlled by the number of users, and it depends on the circumstances on the client side. We increased the number of users by one every second up to 150 which we found the handled number of requests reaches the plateau, and the requests are spawned every second to understand how TensorFlow Serving behaves. So you can assume that requests/second doesn’t reflect the real-world situation where clients try to send requests at any time.

We experimented with the following node configurations on a Kubernetes cluster. The configurations are read like so: {num_vcpus_per_node}-{ram}_{num_nodes}:

• 2vCPUs, 4GB RAM, 8 Nodes
• 4vCPUs, 8GB RAM, 4 Nodes
• 8vCPUs, 16GB RAM, 2 Nodes
• 8vCPUs, 64GB RAM, 2 Nodes

You can find code for experimenting with these different configurations in the above-mentioned repositories. The deployment for each experiment is provisioned through Kustomize to overlay the base configurations, and file-based configurations are injected through ConfigMap.

Results

This section presents the results for each of the above configurations and suggests which configuration is the best based on the environments we considered. As per Figure 1, the best configuration and the environmental setup is observed as 2 nodes, 8 intra_op_parallelism_threads, 8 inter_op_parallelism_threads, 8vCPUs, 16GB RAM based on the result.

Figure 1: Comparison between different configurations of TensorFlow Serving (original).

We have observed the following aspects by picking the best options.

• TensorFlow Serving is more efficient when deployed on fewer, larger (more CPU and RAM) machines, but the RAM capacity doesn’t have much impact on handling more requests. It is important to find the right number of inter_op_parallelism_threads with experimentation. With a higher number the better performance is not always guaranteed even when the nodes are equipped with high-capacity hardware.

TensorFlow Serving focuses more on reliability than throughput performance. We believe it sacrifices some throughput performance to achieve reliability, but this is the expected behavior of TensorFlow Serving, as stated in the official document. Even though handling as many requests as possible is important, keeping the server as reliable as possible is also substantially important when dealing with a production system.

There is a trade-off between performance and reliability, so you must be careful to choose the right one. However, it seems like the throughput performance of TensorFlow Serving is close enough to results from other frameworks such as FastAPI, and if you want to factor in richer features such as dynamic batching and sharing GPU resources efficiently between models, we believe TensorFlow Serving is the right one to choose.

Note on gRPC and TensorFlow Serving

We are dealing with an image classification model for the deployments, and the input to the model will include images. Hence the size of the request payload can spiral up depending on the image resolution and fidelity. Therefore it’s particularly important to ensure the message transmission is as lightweight as possible. Generally, message transmission is quite a bit faster in gRPC than REST. This post provides a good discussion on the main differences between REST and gRPC APIs.

TensorFlow Serving can serve a model with gRPC seamlessly, but comparing the performance of a gRPC API and REST API is non-trivial. This is why we did not include that in this post. The interested readers can check out this repository that follows a similar setup but uses a gRPC server instead.

Costs

We used the GCP cost estimator for this purpose. Pricing for each experiment configuration was assumed to be live for 24 hours per month (which was sufficient for our experiments).

Conclusion

In this post, we discussed some critical lessons we learned from our experience of load-testing a standard image classification model. We considered the industry-grade framework for exposing the model to the end-users – TensorFlow Serving. While our setup for performing the load tests may not fully resemble what happens in the wild, we hope that our findings will at least act as a good starting point for the community. Even though the post demonstrated our approaches with an image classification model, the approaches should be fairly task-agnostic.

In the interest of brevity, we didn’t do much to push further the efficiency aspects of the model in both the APIs. With modern CPUs, software stack, and OS-level optimizations, it’s possible to improve the latency and throughput of the model. We redirect the interested reader to the following resources that might be relevant:

• Scaling up BERT-like model Inference on modern CPU – Part 1
• Scaling up BERT-like model Inference on modern CPU – Part 2
• Load testing and monitoring AI Platform models
• Best practices for performance and cost optimization for machine learning

Acknowledgements

We are grateful to the ML Ecosystem team that provided GCP credits for supporting our experiments. We also thank Hannes Hapke and Robert Crowe for providing us with helpful feedback and guidance.

Read More

A guest post by Brad Dwyer, co-founder and CTO, Roboflow

Roboflow lets developers build their own computer vision applications, from data preparation and model training to deployment and active learning. Through building our own applications, we learned firsthand how tedious it can be to train and deploy a computer vision model. That’s why we launched Roboflow in January 2020 – we believe every developer should have computer vision available in their toolkit. Our mission is to remove any barriers that might prevent them from succeeding.

Our end-to-end computer vision platform simplifies the process of collecting images, creating datasets, training models, and deploying them to production. Over 100,000 developers build with Roboflow’s tools. TensorFlow.js makes up a core part of Roboflow’s deployment stack that has now powered over 10,000 projects created by developers around the world.

As an early design decision, we decided that, in order to provide the best user experience, we needed to be able to run users’ models directly in their web browser (along with our API, edge devices, and on-prem) instead of requiring a round-trip to our servers. The three primary concerns that motivated this decision were latency, bandwidth, and cost.

For example, Roboflow powers SpellTable‘s Codex feature which uses a computer vision model to identify Magic: The Gathering cards.

From Twitter

How Roboflow Uses TensorFlow.js

Whenever a user’s model finishes training on Roboflow’s backend, the model is converted and automatically converted to support sevel various deployment targets; one of those targets is TensorFlow.js. While TensorFlow.js is not the only way to deploy a computer vision model with Roboflow, some ways TensorFlow.js powers features within Roboflow include:

roboflow.js

roboflow.js is a JavaScript SDK developers can use to integrate their trained model into a web app or Node.js app. Check this video for a quick introduction:

Inference Server

The Roboflow Inference Server is a cross-platform microservice that enables developers to self-host and serve their model on-prem. (Note: while not all of Roboflow’s inference servers are TFjs-based, it is one supported means of model deployment.)

The tfjs-node container runs via Docker and is GPU-accelerated on any machine with CUDA and a compatible NVIDIA graphics card, or using a CPU on any Linux, Mac, or Windows device.

Preview

Preview is an in-browser widget that lets developers seamlessly test their models on images, video, and webcam streams.

Label Assist

Label Assist is a model-assisted image labeling tool that lets developers use their previous model’s predictions as the starting point for annotating additional images.

One way users leverage Label Assist is in-browser predictions:

Why We Chose TensorFlow.js

Once we had decided we needed to run in the browser, TensorFlow.js was a clear choice.

Because TFJS runs in our users’ browsers and on their own compute, we are able to provide ML-powered features to our full user base of over 100,000 developers, including those on our free Public plan. That simply wouldn’t be feasible if we had to spin up a fleet of cloud-hosted GPUs.

Behind the Scenes

To implement roboflow.js with TensorFlow.js was relatively straightforward.

We had to change a couple of layers in our neural network to ensure all of our ops were supported on the runtimes we wanted to use, integrate the tfjs-converter into our training pipeline, and port our pre-processing and post-processing code to JavaScript from Python. From there, it was smooth sailing.

Once we’d built roboflow.js for our customers, we utilized it internally to power features like Preview, Label Assist, and one implementation of the Inference Server.

Try it Out

The easiest way to try roboflow.js is by using Preview on Roboflow Universe, where we host over 7,000 pre-trained models that our users have shared. Any of these models can be readily built into your applications for things like seeing playing cards, counting surfers, reading license plates, and spotting bacteria under microscope, and more.

On the Deployment tab of any project with a trained model, you can drop a video or use your webcam to run inference right in your browser. To see a live in-browser example, give this community created mask detector a try by clicking the “Webcam” icon:

To train your own model for a custom use case, you can create a free Roboflow account to collect and label a dataset, then train and deploy it for use with roboflow.js in a single click. This enables you to use your model wherever you may need.

About Roboflow

Roboflow makes it easy for developers to use computer vision in their applications. Over 100,000 users have built with the company’s end-to-end platform for image and video collection, organization, annotation, preprocessing, model training, and model deployment. Roboflow provides the tools for companies to improve their datasets and build more accurate computer vision models faster so their teams can focus on their domain problems without reinventing the wheel on vision infrastructure.

Browse datasets on Roboflow Universe

Get started in the Roboflow documentation

View all available Roboflow features

Read More

Posted by Laurence Moroney and Josh Gordon for the TensorFlow team

With the release of the recent Stack Overflow Developer Survey, we’re delighted to see the growth of TensorFlow as the most-used ML tool, being adopted by 3 million software developers to enhance their products and solutions using Machine Learning. And we’re only getting started – the survey showed that TensorFlow was the most wanted framework amongst developers, with an estimated 4 million developers wanting to adopt it in the near future.

TensorFlow is now being downloaded over 18M times per month and has amassed 166k stars on GitHub – more than any other ML framework. Within Google, it powers virtually all AI production workflows, including Search, Ads, YouTube, GMail, Maps, Play, Photos, and many more. It also powers production systems at many of the largest companies in the world – Apple, Netflix, Stripe, Tencent, Uber, Roche, LinkedIn, Twitter, Baidu, Orange, LVMH, and countless others. And every month, over 3,000 new scientific publications that mention TensorFlow or Keras are being indexed by Google Scholar, including important applied science like the CANDLE research into understanding cancer.

We continue to grow the family of products and open source services that make up the Google AI/ML ecosystem. In recent years, we learned that a single universal framework could not work for all scenarios – in particular, the needs of production and cutting edge research are often in conflict. So we created JAX, a minimalistic API for distributed numerical computing to power the next era of scientific computing research. JAX is excellent for pushing new frontiers: reaching new scales of parallelism, advancing new algorithms and architectures, and developing new compilers and systems. The adoption of JAX by researchers has been exciting, and advances such as AlphaFold and Imagen underscore this.

In this new multi-framework world, TensorFlow is our answer to the needs of applied ML developers – engineers who need to build and deploy reliable, stable, performant ML systems, at any scale, and for any platform. Our vision is to create a cohesive ecosystem where researchers and engineers can leverage components that work together regardless of the framework where they originated. We’ve already made strides towards JAX and TensorFlow interoperability, in particular via jax2tf. Researchers who develop JAX models will be able to bring them to production via the tools of the TensorFlow platform.

Going forward, we intend to continue to develop TensorFlow as the best-in-class platform for applied ML, side-by-side with JAX to push the boundaries of ML research. We will continue to invest in both ML frameworks to drive forward research and applications for our millions of users.

There’s lots of great stuff baking that we can’t wait to share with you, so watch this blog for more details!

PS: Interested in working on any of our AI and ML frameworks? We’re hiring.

Read More