search

Priyanshiguptaaa / Flipkart_Product_Categorization

This repository illustrates the task of applying Machine Translation ( Seq2Seq Attention Network ) for Product Categorization of an E-Commerce Website data (Flipkart), classification of the description of products into the primary category of their category tree, and documenting the path to an optimal model pipeline
4 stars 1 forks source link

Predicting Entire path #2

Closed Bouncer51 closed 3 years ago

Bouncer51 commented 3 years ago

https://arxiv.org/pdf/1812.05774v1.pdf has shown that we could predict the entire product tree. Have you tried that instead of predicting the primary category. if we want to do entire tree what potential changes could be done in present implementation.

Could you also tell how much time did the training took and the configurations of the training system?

Thanks

Priyanshiguptaaa commented 3 years ago

Yes, we can predict the whole tree, but due to computational constraints, I first worked on classifying based on the primary category. For translating into the entire tree, I suggest, instead of extracting the primary category, you can use the category structure that is given initially, i.e. ["Clothing >> Women's Clothing >> Lingerie, Sleep & Swimwear >> Shorts >> Alisha Shorts >> Alisha Solid Women's Cycling Shorts"] as the second language instead of ["Clothing"], and clean up the dataset according to this norm. You'll have to make modifications in the Class Lang accordingly and make sure that it is able to identify >> as a seperater.

For training, without Colab GPU, it took 11 hours. With GPU it worked in 45 mins.

Hope this helps!

Bouncer51 commented 3 years ago

Thank you so much. It helps.