The purpose of this repository is to train a binary image classifier, which can predict of a piece of clothing on an image is worn by a model, or if the piece is displayed by itself.
You can clone this repo by running
git clone https://github.com/MarcoCaglia/yt-content-analysis@[TAG]in your terminal.
The repo comes with a scraper for zalando data. To run the scraper, navigate to the sourcing directory and run the zalando_crawler.
For example like this:
cd sourcing/
scrapy crawl zalando_crawlerNote: For the scraper to run, you need to have a docker container running. For more details, see here: https://github.com/scrapy-plugins/scrapy-splash
By default, the scraper will output a log and write to a database, that are located in the WORKDIR that was specified in the .env file. By default this is set to workdir.
The resulting (sqlite3) project.db will contain two tables: pieces and images. Pieces will contain various information about the scraped pieces of clothing and images will contain a reference to each piece in the pieces table, alongside with the image_urls that have been sourced for any given piece.
By default, the scraper will source from zalando.nl. As of now, this can only be changed in the spider code itself.
NEED TO MAKE HERE A REMARK HOW THE IMAGE LABELER IS WORKING.
The scraper will not scrape the images themselves, but rather the URLs of those images, which will then be stored in the DB (table: images).
To load, label and store the data, the user can use the jupyter notebook scripts/label_images.ipynb. The code will load a random sample of images from the database, present them in the notebook for labeling and will the safe them in the appropriate folder, depending of the user's input.
The DVC pipeline is already created and stored in the dvc.yaml file. To reproduce the original model training, the user shoudl run dvc repro.
The trained and evaluated model will be saved to the workdir.
By default, DVC will store added files in a local minio bucket. Minio can be started by navigating into the workdir of this repo and running docker-compose up. Note that when running this command for the first time, the user would need to create a bucket called image-classification, which is the bucket, that DVC will store files in. By default the minio container runs on port 9000 and the credentials are loaded from the .env file in the workdir folder (not the main folder).
To apply the model, that the dvc pipeline created, run the following code in the console:
python3 fashion_image_classifier/label_images.py --model_path=workdir/model.joblib --db_path=workdir/project.db --brand=YOUR_BRANDIf the location of the model or the project DB was changed, that would need to be reflected in the function call.
At this moment it is also necessary to pass a specific brand for which the images should be labeled.
This repo includes a streamlit dashboard to visualize its findings. To run the dashboard, run streamlit run fashion_dashboard/fashion_dashboard.py in the console. The address of the dashboard will then be shown in the console
The Dashboard is split into sections, which can be selected from the sidebar. By default "All Brands" is selected and the App will show KPIs related to inter-brand comparison.
Alternatively the user can choose a specific brand from the sidebar to see brand-specific KPIs.
The repository contains a Dockerfile, which is ready for deployment.
To run the dashboard on a local container, the user can run:
docker build -t [IMAGE_TAG] .
docker run -p [EXPOSE_PORT]:[EXPOSE_PORTPORT] --env PORT=[EXPOSE_PORT ][IMAGE_TAG]Where [IMAGE_TAG] refers to the name the user wants the image to have and [EXPOSE_PORT] refers to the port the user wants the dashboard to be available on.
To run the container on Heroku (can be done within the free-tier limits), the user can refer to this documentation: https://devcenter.heroku.com/articles/container-registry-and-runtime