Pvillamanario / Panerai_or_not_Ironhack_final_project

Show us your Panerai Watch!

Welcome to this MVP created as final project at Ironhack Data Analysis course 2020.

Upon a image provided by you, it will try to identify if a watch belongs to Panerai or not and will provide some useful information if you are a true Paneristi.

Project structure

This projects is intended to cover almost all the issues and tech we have studied along the course.

 

Stage 1.- Identify is a watch image belongs to a certain watch brand (Panerai).

• Get all the Panerai watches, images and info on sale at Watchfinder.

  • Around 280 items.
  • All the info (price, year, id ...) stored as dataframe.
  • Images saved and renamed with the id and model.

• Scraps Chrono24 by model, brand, etc to get enough amount of labeled images.

  • 8184 images not Panerai vs 7731 Panerai images for further trainning.

• Using a transfer learning, starting with InceptionV3 convolutional neural network architecture, achieving good scores in image classification after trainning: loss: 0.0937 - accuracy: 0.9785

• You can find all info about this model on the notebook attached or in Kaggle.

 

Stage 2.- If the watch image is confirmed as a Panerai watch, make some suggestions based on aesthetic aspects.

• Using as dataset the images fetched from Watchfinder (this is, watches that are on sale).

• Taking a VGG19 cnn and freezing the last layer, we can perform a feature extraction from each image on sale. This results in a unique vector than represents the image and that can be simplified applying PCA.

• Passing a new image through our feature extractor and PCA reduction provides a new vector, which most similars ones can be found through cosine distance, resulting in a group with most similar preprocessed images of watches on sale.

• You can find all info about this model on the notebook attached or in Kaggle.

 

Stage 3.- Getting some more info about our choosen watch.

• Starting from the watches selection offered and using the PAM (model reference in Panerai) of the choosen one, we fetch some info using the PAM as #hashtag.

• Without using Instagram's API, we can obtain the last post, images and hashtags about the selected watch.

• Performs an analysis about the used hashtags, comments sentiment through TextBlob and display the las images about the selected watch.

 

Stage 4.- Getting all the pieces together.

• Using Streamlit, we run a web app were it's possible to upload an image, check if it is a Panerai watch and if the prediction is positive, get the 6 most similar watches on sale. Selecting any of it from this selection activates Instagram scraping and displays all the info.

Setting up the app

Optional

• You can use C24_scraper to create a new dataset (this trained model is included in the repository, so itsn't necesary)

• If you want to update the watches on sale in Watchfinder, feel free to use WF_scraper included.

• Since VGG19 model is a big file to be uploaded to GitHub, it will be checked when executing the app. If it doesn't exists, it'll be downloaded.

 

Running the app

• streamlit run app.py will launch the app on the browser.

 

Have fun!!

• Try with differents pictures, angles, brands, etc!

• Please, don't hesitate to contact me with any issue, suggestion, bugs and so! It will be much appreciated ^_^

• Keep in mind that this was made to learn and above all, to have fun.

 

Tech-stack:

• requests
• pandas
• beautifulsoup4
• matplotlib
• html5lib
• numpy
• tensorflow
• scikit-learn
• scipy
• streamlit
• pillow
• plotly
• wordcloud
• textblob

Don't forget to visit Panerai ;)

To be done:

• Guess what to do when there isn't any content in Instagram.
• Deploy the app (maybe Heroku or similar).