Many at times i would see my sisters take significant effort into finding the perfect match of clothes when it was time for us to go out. They would ask me with each and every selection how it looked on them. I realized that this was an opportunity for my next app

What it does

Glamagenie uses AI techniques to find the perfect outfit for a user and helps them manage their clothing wardrobe (closet). The App relies on an image classifier built using a Habana DL1 instance to identify clothing images that a user would have uploaded and sorts them into the virtual wardrobe.

How we built it

First i had to find out the different types of clothing available to women, next i catalogued the items into 33 classes with 10 sample images for each class. I set up the model training environment on a habana DL1 instance that comes with Tensorflow, Habana Synaspe AI and Docker preinstalled. I proceeded to install Jupyter Server to run my notebooks on then setup an ssh tunnel in Putty to enable browser based access from my local machine to the Jupyter notebooks server. The catalogued items with 33 folders were then placed in a "glamagenie_image_data" folder which i then first "tar"d with 7zip Utility and the "gzipped" into one single file to produce a "tgz" file. I then uploaded the tgz file, named "glamagenie_image_data.tar.gz", onto an S3 bucket with public access. I then proceeded to write and run the model training code in a jupyter notebook.

Labels available in the dataset and model

['ballerina_flats', 'belt', 'blazer', 'blouse', 'bodysuit', 'boot', 'bracelet', 'cap', 'coat', 'dress', 'earrings', 'gloves', 'handbag', 'hoodie', 'jacket', 'jeans', 'necklace', 'oxfords', 'pants', 'pumps', 'romper_jumpsuit', 'sandals', 'scarf', 'shirt', 'short', 'skirt', 'sneaker', 'socks', 'suit_set', 'sunglasess', 'sunhat', 'sweater', 'top']

Training Environment

The model was trained in a jupyter notebook: training environment

Training observations

Training for 20 Epochs:

The model achieved a validation accuracy of around 32 % plot graph 20 epochs

Training for 30 Epochs:

The model achieved a validation accuracy of around 32 % plot graph 30 epochs

Training for 50 Epochs:

The model achieved a validation accuracy of around 30 % plot graph 50 epochs

Model Deployment Strategy

The model is going to be exported in saved model format then a rest-api is created on a server with tensorflow-serving. This API will be used by the Glamagenie App to make inference on the type of uploaded clothing items.

Potential Value

Besides being used as one of the classification engines for the Glamagenie App, this model can be used as a service by online retail stores to quickly label and categorize clothing items into sections on their digital stores. This saves time as online stores owners no longer have to manually update product items in their catalog. They simply feed images into the model and the model identifies the item then sends the correct label and supporting business logic is added to, for example, identify the label category(Footwear, Formal wear, Casual Wear) and update the item type in the database.

Potential Customers

When this model is launched as a web service, monetization is possible. Potential customers that may consume the service include the following:

  1. Shopify stores that focus on clothing, particulary women's wear.
  2. Web developers with ecommerce interests in Women's fashion.
  3. Fashion houses that wish to efficiently categorize their clothing product lines in the shortest possible time.
  4. Data Scientists

The Glamagenie App.

This newly created model is set to improve image recognition of wardrobe items uploaded by the user and identify them. Follow these steps to gain access into glamagenie:

  1. Navigate to Glamagenie
  2. Click on "Sign In"
  3. Enter email as ""
  4. Enter password as "chimtron34"
  5. Navigate to "My Wardrobe"
  6. Add images of clothing Items and wait for them to be classified and identified.

What's next for Glamour Genie (glamagenie)

Glamagenie launches as a web app for women, then next step is to add support for men then build a mobile app for an inclusive experience. The models will be further improved and trained using multiple gaudi accelerators and a richer data set.

Built With

Share this project: