Thursday 8 August 2019

    Neural network inference pipeline for videos in Tensorflow

    Just as we saw a huge influx of images in the past decade or so, we are now seeing a lot of videos being produced on social media. The need to understand and moderate videos using machine learning has never been greater.

    In this post, I will show you how to build an efficient pipeline to processes videos in Tensorflow.  For simplicity, let us consider a Resnet50 model pre-trained on Imagenet. Pretty straightforward, using tf.keras.applications 

    Now, let us break it up to see what exactly is happening:
    1. We are loading the model with weights.
    2. We are reading an image and resizing it to 224x224.
    3. Do some preprocessing of the image.
    4. Run inference.
    5. Do some post processing. 

    If we want to do something similar for large videos, we need to have a pipeline that takes a stream of frames from the video, applies preprocess transformations, run inference of frames, unravel the inferences and apply post processing.

    We can see that doing all these in a sequence - frame by frame is clearly not the right thing as it is slow and inefficient. In order to tackle this, we will use and run inference in batch.

    First, lets create a generator that can produce frames from a video:

    We will use the method to create a dataset object out of this.

    Now let us define a function which does resizing, normalization and other preprocessing steps that are required on a batch of frames. Then, using the batch operation on the dataset created above, create a batch of size 64. Map the preprocess method that we defined onto the batch in parallel on CPU as it is a CPU intensive task.

    It is important to make sure that I/O is parallelized as much as possible. For best performance, instructions that are well suited for CPU should run on CPU and the ones suited for GPU should run on GPU. Also, If you observe the code above, we are prefetching. What this means is that, before consuming the dataset, a batch of 64 frames are preprocessed and is ready for consumption. By the time we run inference on a batch of frames, the next batch is ready for consumption. This is very important because, it ensures that we utilize CPU, GPU, I/O at its highest potential. Here, we are prefetching one unit at a time; and for your usecase, you may prefetch a different size. I always run nvidia-smi to tune the batch size, number of workers, prefetching etc so that both CPU and GPU are always in use in my job.

    Let's put all of this together with running the actual inference:

    This is all good, but what if you have some post processing that can be parallelized on CPU? That will keep the GPU idle till it is processed. So, let's make the inference step a generator that is part of the pipeline and feeds to a second dataset object.

    I have used post processing to do things like writing to disk, data visualization, generating content etc that can be parallelized. This was a simple guide to design data pipelines for inference in Tensorflow. Here is the code to a sightly complicated architecture that generates a video output (observe how I pass on the original frame through the pipeline).

    More resources: 

    Wednesday 13 June 2018

    Finding Where's Waldo using Mask R-CNN

    When I was a kid, I really loved solving Where's Waldo. There were few books (it used to be called Where's Wally) in our school library on which I spent hours finding Waldo. For people who do not know what it is, basically Waldo - a unique character is hidden among hundreds of other characters and you have to find him in all the chaos in the image.

    Now that I am too old to be solving it and too busy to spend hours on such things, I decided to build a system that uses deep learning to automatically solve it and spent weeks to build it. 

    I started off by treating this like a classification problem with two classes - Waldo and not Waldo, similar to Hot dog - not Hot dog . Once we can get the classification problem successfully solved, we can just apply a classification action mapping (CAM) layer to find Waldo's activations in the image and thus finding Waldo. However I couldn't find enough images of Waldo. I found this repo which has about 20 images. And as there are only 20 Waldo vs thousands of not-Waldo characters, there is very high imbalance in the classes. I still tried though. But the results weren't that great.

    When I looked if someone has already worked on it, I found a medium post which used Tensorflow's Faster R-CNN model to do this. But I didn't want to just find bounding boxes, I wanted to actually mask out Waldo in the image. But I got more images of Where's Waldo from it.

    Then I came across this paper on Mask R-CNN which sounded promising for this usecase. And it was indeed much better than my earlier approach:

    Waldo masked out in the image

    Original Image

    In this post I would like to share how I was able to get the data, tag it and train a model to be able to solve Where's Waldo. You can checkout my code on github here.

    Fork deepwaldo on Github

    Mask R-CNN 

    The main idea here is to:

    1. Take the input image and pass it into a set of convolutional layers that sort of generates a feature map for the given image.
    2. Now, you take this feature map and pass it into a region proposal network which generates rectangular regions that say that for the set of final classes, we might have an image in this region. This will have its own classification loss (rpn_class_loss) and bounding box loss (rpn_bbox_loss).
    3. Now you take these regions and pass it into a ROI pooling layer using something like non-max-suppression.
    4. The regions are then reshaped and passed on to set of convolution layers which predict if there is an object in them or not. This again will have its own classification and bounding box losses. 
    5. Now, you have a separate convolution layers which predicts, for every pixel in the bounding box predicted, is it the given class or not. This essentially gives the mask required. Here in addition to the bounding box and classification losses, we also have mask loss.
    6. You run all these networks together backpropogating all the losses.
    If you want a more clear explanation, checkout the lecture in CS231n

    Mask R-CNN arch from CS231n (In this case we have a 28x28 mask instead of 14x14)


    As I mentioned earlier, I got 20 images from this repo and few more images from the medium post that used Faster R-CNN.  So,  a total of 29 images. I split this into 26 for training and 3 for validation. Then I used the via-via tool (used to tag VGG) to manually draw masks over Waldo in every image.  You can find the images and annotations in my github repo here.


    I trained the model for 30 epochs with 100 steps per epoch. The losses on tensorboard:

    If you want to train on your own dataset, first set your configurations in the file or use the default.

    This will download the Mask-RCNN model trained on coco dataset to the MODEL_DIR folder and trains a model with the data in the DATA_DIR folder.
    For prediction, you can do the following which shows a popup with waldo detected in the image.

    # for example
    python models/logs/waldo20180612T1628/mask_rcnn_waldo_0030.h5 data/val/5.jpg

    In conclusion, the Mask R-CNN algorithm works fairly well to find Waldo for cases where it has already seen similar type of waldo image. Also, it looks like it works much better when the image quality is good and waldo is clearly visible. But I think it is still great since we only had a very tiny training data to train on.

    Tuesday 23 January 2018

    Higher level ops for building neural network layers with deeplearn.js

    I have been meddling with google's deeplearn.js lately for fun. It is surprisingly good given how new the project is and it seems to have a sold roadmap. However it still lacks something like tf.layers and tf.contrib.layers which have many higher level functions that has made using tensorflow so easy. It looks like they will be added to Graphlayers in future but their priorities as of now is to fix the lower level APIs first - which totally makes sense.

    So, I quickly built one for tf.layers.conv2d and tf.layers.flatten which I will share in this post. I have made them as close to function definitions in tensorflow as possible.

    1.  conv2d - Functional interface for the 2D convolution layer.

    • inputs Tensor input.
    • filters Integer, the dimensionality of the output space (i.e. the number of filters in the convolution).
    • kernel_size Number to specify the height and width of the 2D convolution window.
    • graph Graph opbject.
    • strides Number to specify the strides of convolution.
    • padding One of "valid" or "same" (case-insensitive).
    • data_format "channels_last" or "channel_first"
    • activation Optional. Activation function which is applied on the final layer of the function. Function should accept Tensor and graph as parameters
    • kernel_initializer An initializer object for the convolution kernel.
    • bias_initializer  An initializer object for bias.
    • name string which represents name of the layer.

    Tensor output.


    Add this to your code:

    2. flatten - Flattens an input tensor.

    I wrote these snippets while building a tool using deeplearnjs where I do things like loading datasets, batching, saving checkpoints along with visualization. I will share more on that in my future posts.

    Thursday 11 January 2018

    Hacking FaceNet using Adversarial examples

    With the rise in popularity of face recognition systems with deep learning and it's application in security/ authentication, it is important to make sure that it is not that easy to fool them. I recently finished the 4th course on where there is an assignment which asks us to build a face recognition system - FaceNet. While I was working on the assignment, I couldn't stop thinking about how easy it is to fool it with adversarial examples. In this post I will tell you how I managed to do it.

    First off, some basics about FaceNet. Unlike image recognition systems which map every image with a class, it is not possible to assign a class label to every face in face recognition. This is because one, there are way too many faces that a system should handle in the real world to assign class to each of them and two, if there are new people the system should handle, it can't do it. So, what we do is, we build a system that learns similarities and dissimilarities. Basically, there is a neural network similar to what we have in image recognition and instead of applying softmax in the end, we just take the logits as embedding for the given image input and then minimize something called the triplet loss.  Consider face A, we have a positive match P and negative match N. If f is the embedding function and L is the triplet loss, we have this:

    Triplet loss

    Basically, it is incentivizing small distance between A - P and large distance between A - N. Also, I really recommend watching Ian Goodfellow's lecture from Stanford's CS231n course if you want to know about adversarial examples.

    Like I said earlier, this thought came to me while doing an assignment from 4th course from which can be found here and I have built on top of it.  The main idea here is to find small noise that when added to someone's photo although causing virtually no visual changes, can make faceNet identify them as the target.

    Benoit (attacker)
    Add noise
    Kian Actual (Target)

    First lets load the images of the attacker Benoit and the target Kian.

    Now say that the attacker image is A` and the target image is T. We want to define triplet loss to achieve two things:

    1. Minimize distance between A` and T
    2. Maximize distance between A` and A` (original)
    In other words the triplet loss L is:

    L (A, P, N) = L (A`, T, A`)

    Now, let's compute the gradient of the logits with respect to the input image 

    These gradients are used to obtain the adversarial noise as follows :

    noise = noise - step_size * gradients

    According to the assignment, a l2 distance of the embeddings of less than 0.7 indicates that two faces have the same person. So lets do that.

    The distance decreases from 0.862257 to 0.485102 which is considered enough in this case.

    L2 distance between embeddings of attacker and target
    This is impressive because, all this is done while not altering the image visibly just by adding a little calculated noise!

    Also note that the l2 scores indicate that the generated image is more of Kian than Benoit in spite of looking practically identical to Benoit. So there you go, adversarial example generation for FaceNet.

    Sunday 17 December 2017

    Tensorflow and AEM

    It has been a while since google released Tensorflow support for java. Even though it is still in its infancy, I feel like it has everything we need. Build computation graphs - check, run session and compute stuff - check, GPU support - check. Now if you have all the time in the world to reinvent the wheel, you can pretty much build anything in java that we can build using python or c++.

    So, I have been working on Adobe Experience Manager since I joined Adobe and recently, I started experimenting with several use cases where machine learning can help in content creation and discovery. As I have zero knowledge in building any deep learning models in java, I decided to build everything in java. How hard can it be? Right? Right? Sarcasm aside, as I mentioned earlier, Tensorflow for java has everything we need and as it internally uses JNI we can have interoperability with python and c++ (that's why I preferred this over deeplearning4j).

    First off, I followed their official guide for the setup and had to face a lot of hurdles along the way. In this post I will show you how I managed to successfully setup Tensorflow on AEM (or any felix based systems).

    Step 1

    Add the dependency to your pom.xml file. Note that the scope set to compile.

    Step 2

    Add this configuration to your maven-bundle-plugin.

    Step 3

    Build and install to your AEM instance. Then, navigate to /system/console/bundles/ and look for the bundle which contains the dependency. See if the "Exported Packages" section has the following packages:

    Step 4

    Install JNI if necessary (this is mentioned in the link that I shared earlier).

    Then place the library file in the appropriate place.


    Lets write a simple sling servlet to check if everything is working as expected. Like I told earlier, Tensorflow for java is still in its infancy. So, I wrote a helper class a while back to manipulate the computation graph. Get and place it where it is accessible to the sling servlet.

    The following sling servlet includes things like:
    • Creating a computation graph
    • Creating placeholders, constants etc
    • Arithmetic operations, matrix multiplication.
    • Feeding data and computing values of placeholders.

    When you go to /services/tftest you should get something like this:

    4 -2 3 0 FLOAT tensor with shape [3, 3] 14 Testing done!
    Now you can start building any deep learning model on AEM. Also, I will be writing about some of the real life applications of deep learning in content creation and content discovery. So stay tuned!

    Wednesday 29 March 2017

    Most original prize at The 2017 Deep Learning Hackathon

    Although I have worked on several deep learning projects in the past, I still consider myself to be a newbie in deep learning because of all the new things that keep coming up and it is so hard to keep up with all that. So, I decided to take part in "The 2017 Deep Learning Hackathon" by Deepgram to work on something I have been wanting to do for a while now.

    I built something called Medivh - prophet from Warcraft who has seen the future.  The idea was to build a tool for web developers to predict how users are going to see / use the site even before deploying. Basically, it generates heat maps on websites which show where the user might look at. Example:

    I will write another post with all the technical details. Here is the sneak peak of how it was done.

    Apart from building that, We got an opportunity to interact with people like Bryan Catanzaro - maker of CUDNN and VP at Nvidia,  Jiaji Huangform from Baidu, Jonathan Hseu from Google Brain etc.

    We also got to interact with people from Deepgram and their caffe like framework called Kur which seems pretty good. I think I'll write a review about Kur after playing around with it for some more time.

    Also this:

    This is me presenting before the results.
    For Medivh, I won the "Most original prize" -  Nvidia Titan X pascal. What a beauty!

    Thursday 9 March 2017

    Introducing mailing in crontab-ui

    Now crontab-ui has option to send mails after execution of jobs along with output and errors attached as text files. This internally uses nodemailer and all the options available through nodemailer are available here.


    To change the default transporter and mail config you can modify config/mailconfig.js.
    var transporterStr = 'smtps://';
    var mailOptions = {
        from: '"Fred Foo 👥" <>', // sender address
        to: ',', // list of receivers
        subject: 'Job Test#21 Executed ✔', // Subject line
        text: 'Test#21 results attached 🐴', // plaintext body
        html: '<b>Test#21 🐴</b> results attached' // html body


    Make sure that you have node at /usr/local/bin/node else you need to create a softlink like this
    ln -s [location of node] /usr/local/bin/node

    Setting up crontab-ui on raspberry pi

    In this tutorial I will show you how to setup crontab-ui on raspberry pi.

    Step 1

    Find your architecture
    uname -a
    Linux raspberrypi 4.4.50-v7+ #970 SMP Mon Feb 20 19:18:29 GMT 2017 armv7l GNU/Linux
    Note that it is ARMv7. Download and extract latest node.
    tar xz node-v7.7.2-linux-armv7l.tar.xz
    sudo mv node-v7.7.2-linux-armv7l /opt/node

    Step 2

    Remove old nodejs if it is already installed and add the latest node to the $PATH
    sudo apt-get purge nodejs
    echo 'export PATH=$PATH:/opt/node/bin' > ~/.bashrc
    source ~/.bashrc

    Step 3

    Install crontab-ui and pm2. And start crontab-ui.
    npm install -g crontab-ui
    npm install -g pm2
    pm2 start crontab-ui
    Now your crontab-ui must be running. Visit http://localhost:8000 on your browser to see if it is working.

    Step 4 (Optional)

    In order to be able access crontab-ui from outside, you have to forward the port 8000. Install nginx and configure.
    sudo apt-get install nginx
    sudo vi /etc/nginx/sites-available/default
    Paste the following lines in the file:
    server {
        listen 8001;
        server_name localhost;
        location / {
            proxy_pass http://localhost:8000;
    Restart nginx
    sudo service nginx restart
    Now, crontab-ui must be accessible from outside through port 8001. So, to access crontab-ui, go to
    <ip address of pi>:8001
    You can also setup http authentication by following this.
    Fork me on Github

    Saturday 28 January 2017

    My solutions to cmdchallenge

    I recently stumbled upon which sort of tests your command line knowledge and comfortability. You have to basically solve all the challenges in a single line of bash. It is pretty simple and fun. You should give it a try before checking the solutions.


    # Print "hello world".
    # Hint: There are many ways to print text on
    # the command line, one way is with the 'echo'
    # command.
    # Try it below and good luck!
    echo "hello world"


    # Print the current working directory.


    # List all of the files in the current
    # directory, one file per line.
    ls -1


    # Print the last 5 lines of "access.log".
    tail -5 access.log


    # There is a file named "access.log" in the
    # current working directory. Print all lines
    # in this file that contains the string "GET".
    grep GET access.log


    # Print all files, one per line that contain
    # the string "500".
    grep -rl * -e 500


    # Print the relative file paths, one path
    # per line for all files that start with
    # "access.log" in the current directory.
    find . -name "access.log*"


    # Print all matching lines (without the filename
    # or the file path) in all files under the current
    # directory that start with "access.log" that
    # contain the string "500".
    find . -name "access.log*" | xargs grep -h 500


    # Extract all IP addreses from files that
    # that start with "access.log" printing one
    # IP address per line.
    find . -name "access.log*" | xargs grep -Eo '^[^ ]+'


    # Delete all of the files in this challenge
    # directory including all subdirectories and
    # their contents.
    find . -delete


    # Count the number of files in the current
    # working directory. Print the number of
    # files as a single integer.
    ls | wc -l


    # Print the contents of access.log
    # sorted.
    sort access.log


    # Print the number of lines
    # in access.log that contain the string
    # "GET".
    grep GET access.log | wc -l


    # The file split-me.txt contains a list of
    # numbers separated by a ';' character.
    # Split the numbers on the ';' character,
    # one number per line.
    cat split-me.txt | sed s/\;/\\n/g


    # Print the numbers 1 to 100 separated
    # by spaces.
    echo {1..100}


    # There are files in this challenge with
    # different file extensions.
    # Remove all files with the .doc extension
    # recursively in the current working directory.
    find . -name "*.doc" -delete


    # This challenge has text files that contain
    # the phrase "challenges are difficult". Delete
    # this phrase recursively from all text files.
    find . -name "*.txt" -exec sed -i 's/challenges are difficult//g' {} +


    # The file sum-me.txt have a list of numbers,
    # one per line. Print the sum of these numbers.
    cat sum-me.txt | xargs | sed -e 's/\ /+/g' | bc


    # Print all files in the current directory
    # recursively without the leading directory path.
    find . -type f -printf "%f\n"


    # Remove the extension from all files in
    # the current directory recursively.
    Solution: (note you cant use find .)
    find `pwd` -type f -exec bash -c 'mv "$1" "${1%.*}"' - '{}' \;


    # The files in this challenge contain spaces.
    # List all of the files in the current
    # directory but replace all spaces with a '.'
    # character.
    find . -type f -printf "%f\n" | xargs -0 -I {} echo {} | tr ' ' '.'


    # There are a mix of files in this directory
    # that start with letters and numbers. Print
    # the filenames (just the filenames) of all
    # files that start with a number recursively
    # in the current directory.
    find . -name '[0-9]*' -type f -printf "%f\n"


    # Print the 25th line of the file faces.txt
    sed '25q;d' faces.txt


    # Print the file faces.txt, but only print the first instance of each
    # duplicate line, even if the duplicates don't appear next to each other.
    awk '!seen[$0]++' faces.txt


    # You have a new challenge!
    # The following excerpt from War and Peace is saved to
    # the file 'war_and_peace.txt':
    # She is betraying us! Russia alone must save Europe.
    # Our gracious sovereign recognizes his high vocation
    # and will be true to it. That is the one thing I have
    # faith in! Our good and wonderful sovereign has to
    # perform the noblest role on earth, and he is so virtuous
    # and noble that God will not forsake him. He will fulfill
    # his vocation and crush the hydra of revolution, which
    # has become more terrible than ever in the person of this
    # murderer and villain!
    # The file however has been corrupted, there are random '!'
    # marks inserted throughout.  Print the original text.
    Solution: (Found this on hackernews)
    < war_and_peace.txt tr -s '!' | sed 's/!\([a-z]\)/\1/g' | sed 's/!\( [a-z]\)/\1/g' | sed 's/!\.!/./g' | sed 's/ !/ /g'

    Also, you can checkout the creator's solutions here.

    Thursday 19 January 2017

    Look before you paste from a website to terminal

    Most of the time when we see a code snippet online to do something, we often blindly copy paste it to the terminal. Even the tech savy ones just see it on the website before copy pasting. Here is why you shouldn't do this. Try pasting the following line to your terminal (SFW)

    ls ; clear; echo 'Haha! You gave me access to your computer with sudo!'; echo -ne 'h4cking ## (10%)\r'; sleep 0.3; echo -ne 'h4cking ### (20%)\r'; sleep 0.3; echo -ne 'h4cking ##### (33%)\r'; sleep 0.3; echo -ne 'h4cking ####### (40%)\r'; sleep 0.3; echo -ne 'h4cking ########## (50%)\r'; sleep 0.3; echo -ne 'h4cking ############# (66%)\r'; sleep 0.3; echo -ne 'h4cking ##################### (99%)\r'; sleep 0.3; echo -ne 'h4cking ####################### (100%)\r'; echo -ne '\n'; echo 'Hacking complete.'; echo 'Use GUI interface using visual basic to track my IP'

    It should look something like this once it is pasted onto your terminal.
    View post on
    You probably guessed it. There is some malicious code between ls and -lat that is hidden from the user

    Malicious code's color is set to that of the background, it's font size is set to 0, it is moved away from rest of the code and it is made un-selectable (that blue color thing doesn't reveal it); to make sure that it works in all possible OSes, browsers and screen sizes.

    This can be worse. If the code snippet had a command with sudo for instance, the malicious code will have sudo access too. Or, it can silently install a keylogger on your machine; possibilities are endless. So, the lesson here is, make sure that you paste code snippets from untrusted sources onto a text editor before executing it.

    Thanks for reading!