Skip to content

Latest commit

 

History

History
146 lines (93 loc) · 9.14 KB

README.md

File metadata and controls

146 lines (93 loc) · 9.14 KB

Portal

Build Tests Join Datature Slack

Portal is the fastest way to load and visualize your deep learning models. We are all sick of wrangling a bunch of cv2 or matplotlib codes to test your models - especially on videos. We created portal to help teams, engineers, and product managers interactively test their model on images and videos, inference thresholds, IoU values and much more.

Portal is an open-source browser-based app written in TypeScript, React and Flask.

Made with by Datature


Portal User Experience

Portal works on both images and videos - bounding boxes and masks - allowing you to use it as a sandbox for testing your model's performance. Additionally, Portal supports Datature Hub, TensorFlow and DarkNet models (PyTorch Support Incoming) and runs on either our electron app or your browser.

Portal User Experience Portal User Experience

Setting up Portal

Portal can be used as a Web Application or through downloading and installing our Electron package via Portal Releases.

Running Portal as a Web Application

This is the recommended way of running Portal

Portal is built using Python 3.7. Ensure that you have this version and up before beginning (Python 3.7 <= X < 3.9). Clone the repository and then navigate to the directory where requirements.txt is and install all necessary dependencies and setup using setup.sh:

git clone https://github.com/datature/portal
cd portal
pip3 install -r requirements.txt
./setup.sh

Running the following command will open the Portal application on the browser via http://localhost:9449.

If you wish to run the application on gpu, add a trailing --gpu flag (This only works for TensorFlow Models)

python3 portal.py

Using Virtual Environment

If you'd like to use virtual environments for this project - you can use a helpful script below to before activating the virtualenv -

./setup-virtualenv.sh

Running from Portal Executable

Portal comes with an installable version that runs on electron.js - this helps to provide a desktop application feel and ease of access of setting up. To install, please download the latest Portal Releases and run the Portal installer for your OS.

Navigating Portal

On starting Portal or navigating to http://localhost:9449 - The following steps details how you can load your YOLO or TensorFlow model on your image folders. To begin, let's assume we want to register a tf2.0 model. On Portal, a concept we use is that you can register multiple models but load one at each time.

Registering and Loading Portal

Start by clicking on the + sign and adding the relevant filepaths, e.g. /user/portal/downloads/MobileNet/ and a name. You will be prompted to load the model as seen below. Simply click on the model you'd like to load and the engine wil

Register Model Load Model

Loading Your Images / Videos

To load your dataset (images / videos), click on the Open Folders button in the menu and paste your folder path to your dataswr. Once you are done, press the enter button. The images should appear in the asset menu below. You can load and synchronize multiple folders at once on Portal.

Load Assets

Running Inferences

Click on any image or video, press Analyze, and Portal will make the inference and render the results. You can then adjust the confidence threshold or filter various classes as needed. Note that Portal run inferences on videos frame-by-frame, so that will take some time. You can change the inference settings, such as IoU or Frame Settings under Advanced Settings.

Image Prediction Image Prediction

Keymaps and Shortcuts

To view the various key maps and shortcuts, press ? on your keyboard whilst in Portal. There are various shortcuts such as showing labels of detections, going to the next photo, etc. If you have any suggestions or change recommendation, feel free to open a Feature Request

Portal works on both Mask and Bounding Box models. For detailed documentations about advanced features of Portal can be found here : Portal Documentation

Working with Datature Nexus

Portal works seemlessly with Nexus, our MLOps platform, that helps developers and teams build computer vision models - it comes fully featured with an advanced annotator, augmentation studio, 30+ models and ability to train on multi-GPU settings. Anyhoo, here's how to build a model and run it in Portal -

Image Prediction Image Prediction

To build a model on Nexus, simply create a project, upload the dataset, annotate the images, and create a training pipeline. You should be able to start a model training and this can take a few hours. As the model training progress, checkpoints are automatically generated and you should see them in the Artifacts page. For more details on how to use Nexus, consider watching our tutorial here.

Image Prediction Image Prediction

Once you are done with a training and a candidate checkpoint is selected, you can generate a TensorFlow model under the Artifacts page to obtain the model key required by Portal for the following setup. Use the register model interface to insert this key to the model under Datature Hub.

Image Prediction

With this, you can now run Analyze on your test images and you should be able to train and test new models between Nexus and Portal easily by repeating the steps above. If you'd like to give Datature Nexus a try, simply sign up for an account at https://datature.io - It comes with a free tier!

Image Prediction

Screencasts

Using Portal to Inspect Computer Vision Models

Building an Object Detection Model with Datature

Building an Instance Segmentation Model with Datature

Sample Weights

We have provided sample weights for you to test Portal:

Dataset Description Download Link
YOLO-v3 DarkNet Model based off pjreddie/darknet YOLOv3
SSD MobileNet V2 FPNLite 640x640 Tensorflow Model from tensorflow/models MobileNet