DEV Community: wafaa Arbash

Time Series Data / Audio Labeling

wafaa Arbash — Thu, 05 Nov 2020 15:31:43 +0000

If you haven’t heard of the the Universal Data Tool, it’s an open-source web or desktop program to collaborate, build and edit text, image, video, and audio datasets with labels and annotations.

This is our ninth community update! By releasing these videos, we hope to engage the community and encourage new contributors.

You can watch the full video here

New Feature: Time Series Interface

We now support the annotation of time series data with the “time_series” interface! Using this interface, you can import audio, JSON time data or CSVs, then add durations or timestamps. Check it out!

Other Updates

-Fixed importing files from a directory in desktop application
-Fixed sample numbers issue: so now users can see samples number. (thanks, brian)
-Fixed multiple label classification bug, now users can click on multiple classifications.

That’s it for our ninth community update, be sure to follow us on Twitter, or join our Slack to hear more!

On-Premise Data Labeling

wafaa Arbash — Wed, 21 Oct 2020 19:36:49 +0000

If you haven’t heard of the the Universal Data Tool, it’s an open-source web or desktop program to collaborate, build and edit text, image, video, and audio datasets with labels and annotations.

This is our eight community update! By releasing these videos, we hope to engage the community and encourage new contributors.

You can watch the full video here

Join us for a Hacktoberfest Hackathon

We are super excited to announce our 8-hour Hacktoberfest Hackathon on October 29th. You can get all the information as well as register at this link
https://organize.mlh.io/participants/events/5164-universal-data-tool-livestream-mini-hackathon

New Feature: Asynchronous Datasets Loading

You can now add millions of samples in a single dataset when using an external server.

Easier On-Premise Usage

The Universal Data Tool along with a collaboration server can now be run as a single docker container. See the on-premise documentation for more details.

You can provide different options to the docker container to control various aspects of the Universal Data Tool. You'll know if a configuration value is used because it will appear in the starting logs as shown below. This allows you to configure things like S3 integrations without needing any user input.

Other Updates

-Fixed reload config state in Audio Transcription: Configuration in Audio Transcription is updating now (thanks @anaplian!)
-Fixed Portuguese translation: Now you can see the special characters in Portuguese translation (thanks, @miguelcarvalho13)
-Introducing Manage Plugins: You might have noticed the “Manage Plugins” option introduced in version 13 of UTD. It helps to add an ES6 module to add more functionalities. This is a work-in-progress as of now. However, there will be new features in this option soon.

That’s it for our eighth community update, be sure to follow us on Twitter, or join our Slack to hear more!

Build your dataset from COCO

wafaa Arbash — Mon, 05 Oct 2020 16:55:24 +0000

If you haven’t heard of the the Universal Data Tool, it’s an open-source web or desktop program to collaborate, build and edit text, image, video, and audio datasets with labels and annotations.

This is our seventh community update! We’re hoping by releasing these we can better engage the community and encourage new contributors. Let’s get started…

You can watch the full video here

New Feature: Import dataset from COCO(Common Objects in Context)

This new feature allows you to import and build your datasets using the “import from COCO” option. You just need to add the keywords of the object you want to import e.g., cats or dogs, then click “Add Samples.”

New Feature: Remove Invalid Image Samples

This feature allows you to automatically remove invalid image URLs that are not showing any images. Just go to transform data and click on “Remove Invalid Samples,” It will do the work.

New Feature: Universal Data Tool has a new URL

Sometimes it’s a bit annoying to type the long URL names in the web browsers, and with this new update, we have a new short URL for our tool, and that is udt.dev. The good news is that the old URL works fine too!

Other updates:

-More documents are on docs.universaldatatool.com: As promised, a vast quantity of documentation is available on using Universal Data Tool for your projects.
-Fixed Hotkeys: Clearing hotkeys now works (thanks @congdv!)

That’s it for our seventh community update, be sure to follow us on Twitter, or join our Slack to hear more!

New Skeletal/Pose/Landmark Annotation, Dutch, and Convert Options

wafaa Arbash — Thu, 24 Sep 2020 17:43:33 +0000

If you haven’t heard of the the Universal Data Tool, it’s an open-source web or desktop program to collaborate, build and edit text, image, video, and audio datasets with labels and annotations.

This is our sixth community update! By releasing these videos, we hope to engage the community and encourage new contributors.

You can watch the full video here

New Feature: Image landmark Annotation

This cool new feature allows people to annotate poses in the picture. We are looking for help in the configuration tab for this feature. To help us, you can click on the Github issue option from the configuration tab.

Other updates:

-Dutch translation: Now, you can also access the Dutch language (thanks @rickstaa!)
-Singularity support: Now you can use the singularity which is kind of a container registry for scientific applications thanks @rickstaa!)
-Icons on website options: Now, you can see our website’s icons with the option names.
-Our website is open source now: with this new update, our website is open-source. This will allow you to do any modifications in PR
-Previous & Next buttons fixed: Now the previous & next buttons will be disabled at the beginning or at the end sample pages, respectively (thanks @congdv! ).
-Allow multiple classifications per Image is fixed: This feature works perfectly now (thanks @mrdadah! )
-Convert Tool has multiple options now: Our previous new feature to convert the formats has many new options, and the option to add new formats as well.

That’s it for our sixth community update, be sure to follow us on Twitter, or join our Slack to hear more!

Universal Data Tool Weekly Update 5

wafaa Arbash — Tue, 15 Sep 2020 17:07:26 +0000

If you haven’t heard of the the Universal Data Tool, it’s an open-source web or desktop program to collaborate, build and edit text, image, video, and audio datasets with labels and annotations.

This is our fifth community update! We’re hoping by releasing these we can better engage the community and encourage new contributors. Let’s get started…

You can watch the full video here

New Feature: Documentation

We are super excited to announce the docs.universaldatatool.com site. It has all the how-to documents as well as ways to use UDT with machine learning frameworks. You are welcome to help us add more details in it using github.

New Feature: Convert tool

With this update, you can now convert JSON or CSV files into PNG or SVG files. To use it all you need to click on the “convert” option on the Universal Data Tool Homepage. This option supports conversion with image options output as of now, however, we are hoping to add more features in it soon. The good part is that you can use this option to convert fairly large datasets with ease. Also, this option can be used online as well as on local machines.

New update: Fixed Universal Data Tool Desktop Builds

Our desktop builds are back in action. Now you can download the Universal Data Tool for Mac, Linux, and Windows stable releases.

Other updates:

Portuguese translation: Now you can access the website in Portuguese translation as well (thanks @mrjunato!)
New responsive page: With this update, the Universal Data Tool landing page is responsive in any size (thanks @congdv!)

That’s it for our fifth community update, be sure to follow us on Twitter, or join our Slack to hear more!

Universal data tool weekly update 4

wafaa Arbash — Mon, 07 Sep 2020 15:27:57 +0000

If you haven’t heard of the Universal Data Tool, it’s an open-source web or desktop program to collaborate, build and edit text, image, video, and audio datasets with labels and annotations. You can get started with the Universal Data Tool

This is our fourth community update! We’re hoping by releasing these we can better engage the community and encourage new contributors. Let’s get started…

To watch the full video click here

New Feature: Universal Data Tool Courses

With this update, we are introducing the training courses for Universal DataTool. With this feature now you can ensure that your labeler team is working not only in a streamlined way but with the best quality as well. The people working on a dataset will have to complete it before they can begin working on the labels.
Creating a training course is super easy and it can be done in just a few minutes. You can insert a dataset of your choice to be put in the training. This training contains questions, exercises, tests, and different editable options.

New Feature: Text Entity Relations

This is a much-awaited feature that was missing in the NLP(Natural Language Processing). This feature allows you to draw between different texts to create the relation. You can do it just by using a one-line command and it works super-fast. Not only that but you can also click on read more options to get additional commands and ways to use them which can be useful in different scenarios.

New Feature: Download Mask Button

This is a small but very useful feature that we have added in the latest update. With this feature, you can download the colored masks just by clicking on the button which is in the download options list.

New update: Brand new layout

With this new update, you can see a totally new layout of the options. It’s less cluttered and different options can be seen in different tabs. This makes using the Universal Data Tool easier than ever.

Other updates:

Feedback button: Now you can share the feedback with us by clicking on this little button on the right bottom corner of the page. It allows you to add the screenshots as well. (thanks @moufette)
New Homepage: Now we have a redesigned and cool looking Homepage which can give you more information and insight of the Universal Data Tool

That’s it for our forth community update, be sure to follow us on Twitter, or join our Slack to hear more!

Universal Data Tool Weekly Update 3

wafaa Arbash — Fri, 21 Aug 2020 17:43:16 +0000

This is our third community update! We’re hoping by releasing these we can better engage the community and encourage new contributors. Let’s get started…

To watch full video click here

New Feature: Two different AWS authentication methods

Universal Data Tool now supports 2 methods of the authentication, you can either use Incognito mode or regular IAM mode.

To use the IAM mode, you just need to enter your Access Key ID, your Secret Access Key, and the region of your S3 bucket and you’re good to go.

The CORS proxy allows you to make requests of the things where the CORS origin security rules don’t allow you to use access to the resource. We introduced this week because it will allow us to use AWS S3 API. There are 2 new options to upload that’s IAM and S3.

New Feature: Fixes for Login with Cognito and import via S3 (Cognito)

We also have the fixes for Login With Cognito including the detection of the upload state, So when anyone logs in and refreshes the page, it remembers to keep logged in.

We also fixed import via S3 (Cognito), so it loads automatically from the bucket. Another fix is for the local directory import.

That’s it for our third community update, be sure to follow us on Twitter, or join our Slack to hear more!

Single-Label Image Classification with Keras

wafaa Arbash — Tue, 18 Aug 2020 17:59:22 +0000

Recently, generated.photos released a royalty-free dataset of images of human faces. But unlike most datasets, this dataset is completed generated by an AI. None of the faces are real!

I thought this would be a fun dataset to teach a machine learning algorithm to classify sex on. For this tutorial, I'll be using wao.ai to build my dataset and keras to run the algorithm. I downloaded about 2,000 faces from generated.photos. Believe it or not, none of the photos below are real!

Prepare the Data

To prepare the data, I resized each photo to 256x256 pixels, this makes it easier to fit all the photos in RAM. Resizing the photos can be done quickly with image magick with a command like this:

convert *.jpg -resize 256x256\! *.jpg

I repackaged all the images we'd use in this zip file so you can skip the resizing if you're following along.

Get Sex Labels

To train our algorithm, we'll need to a sex label for each face. We could manually create our labels using the universal data tool, but since we're dealing with more than 100 images I'm going use the wao.ai workforce.

Downloading a csv from wao.ai, we get a CSV full of the labels we'll learn (shown below). You can download the labels.csv here to continue following along.

wao.ai csv download screenshot

Train a Model

There are a lot of ways to choose a model to do training with. For a simple computer vision task, especially a classification task, I like to start with a simple model that performs well on the MNIST digit dataset (a common machine learning benchmark). I don't remember where I found this model, but it's good for small (1000-10000) image datasets. Note: To use it, we'll need to convert the images to grayscale and resize them to a resolution of 64x64.

from tensorflow import keras
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout, Flatten
from tensorflow.keras.layers import Conv2D, MaxPooling2D
num_output_classes = 2 # 0 = male, 1 = female
input_img_size = (64, 64, 1) # 64x64 image with 1 color channel
model = Sequential()
model.add(Conv2D(32, kernel_size=(3, 3), activation="relu", input_shape=input_img_size))
model.add(Conv2D(64, (3, 3), activation="relu"))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(64, activation="relu"))
model.add(Dropout(0.5))
model.add(Dense(num_output_classes, activation="softmax"))
model.compile(
loss=keras.losses.categorical_crossentropy,
optimizer=keras.optimizers.Adadelta(),
metrics=["accuracy"],
)

To see the surrounding code, including data transformations and training parameters, check out all my work in this notebook.

Results

In the end, this model was able to predict sex with about 80% accuracy. Not great, but a reasonable start considering we're working on a small 2,000 sample dataset.

Let's look at the data qualitatively. First, let's see the top 25 our model classified as male:

Now let's see what our model classified as female:

Interesting! I see a lot of similarities in the general outline of males and the general outline of females (look at the hair!). It would also appear that this dataset has many females wearing hats.

To see where our model is struggling, we can also see where the model was most unsure, i.e. where |P(male) - P(female)| was smallest.

Children, indirect face angles and neutral hair cuts seem to confuse the model.

To enhance the effectiveness of the model, we could try a different neural network architecture, create descriptive features (e.g. features that identify children or facial hair) or label additional data for testing.

Big thanks to generated.photos for providing the excellent AI-generated faces!

Be sure to follow us on Twitter, or join our Slack to hear more!

Universal Data Tool Weekly Update 2

wafaa Arbash — Sun, 16 Aug 2020 19:56:58 +0000

This is our second community update! We’re hoping by releasing these we can better engage the community and encourage new contributors. Let’s get started…

You can watch the video update here

(https://youtu.be/3bq9N08oc-U)

New Feature: Collapsible Sidebar

All Image Segmentation views now have a collapsible sidebar with expendable options. Now you can use the entire screen for segmentation if needed by just clicking on the arrow on the right side of the screen. It should help to clear the UI and be beneficial when used on smaller screens.

New Feature: Convert Image Samples into any number of segments

Now there is no limit to convert image samples into segments as opposed to earlier 5x5 limits. Also, there is an improvement to how we zoom in on segments when you have a sample split up in segments. It will now automatically zoom in on the segment that’s allowed to be annotated in every sample.

That’s it for our second community update, be sure to follow us on Twitter, or join our Slackto hear more!

Label Bounding Boxes with the Universal Data Tool

wafaa Arbash — Wed, 12 Aug 2020 19:08:00 +0000

Labeling your data with bounding boxes is an essential first step before training a machine learning algorithm. The Universal Data Tool is a free, open-source application that makes this process simple.

Before you start, you should have either...

A directory of images
A file where each line is an image URL (this is what I'll use)
Configure the Job

Configure the Job

To begin navigate to universaldatatool.com or download the latest release from github. I'll use the web version for now.

Click "Create from Template" and select Computer Vision, then configure it to only do bounding boxes. If you also want to label each bounding box with a classification, you can also configure that here.

Import the Data

Navigate to "Samples", then to the "Import" page. Click "Paste Image URLs", paste the image urls and click "Add Samples".

Label the Data

You can now begin labeling. Click the "Label" button at the top, then click on any sample to begin. Samples will turn blue as they're labeled.

Wrapping Up

When you're done, navigate to the "Settings" tab and click Download JSON. The resulting file can be loaded into the Universal Data Tool again, or parsed for usage with a machine learning algorithm.

That's it! You now have everything you need to add bounding boxes to an image dataset. Be sure to follow us on Twitter, or join our Slack to hear more!

Universal Data Tool Weekly Update

wafaa Arbash — Mon, 10 Aug 2020 20:16:28 +0000

This is our first community update! We’re hoping by releasing these we can better engage the community and encourage new contributors. Let’s get started…

(http://img.youtube.com/vi/YOUTUBE_VIDEO_ID_HERE/0.jpg)

New Feature: Automatic Pixel Segmentation (Autoseg)

It can take hours to accurately annotate every pixel of a large image. That’s why we created autoseg, autoseg is a new engine that takes a bunch of points or shapes and automatically creates a pixel mask super fast. What this means is you can now quickly do pixel segmentations directly in the Universal Data Tool!

Automatic pixel segmentation can be found in the “Pixel Segmentation” interface when you’re starting a new project. In the Setup page, just click on “autoseg” under “Automatic Segementation Engine”. Try it out!

For more information about the autoseg engine check out it’s github repo and our image annotation react component.

New Feature: Translation Support

The Universal Data Tool is now available in English, French and Chinese! Big thanks to @puskuruk for putting it together. You can change your language using the dropdown on the landing page.

Adding or fixing translations is also now very easy! All the files you need to edit are in the locales directory.

New Feature: Docker Container Support

For running on servers or quickly testing out the UDT, it can be much faster to

The Universal Data Tool can now run in a docker container! If you have docker installed, you can now run:

docker run -it -p 3000:3000 universaldatatool/universaldatatool

The Universal Data Tool is now running on http://localhost:3000

That’s it for our first community update, be sure to follow us on Twitter, or join our Slack to hear more!