• Notebook

• In this project I trained Hindi Language Model with BBC Hindi News Dataset and then Built a Hindi News Classifier.

Dataset:

• BBC Hindi

Hindi Language Model:

• Architecture : AWD-LSTM
• Learn more about AWD-LSTM Here —> ASGD Weight-Dropped LSTM

• Model Summary :

• DataBlock :

• Used Mixed Precision training to decrease up the training time.

Achieved a final accuracy of 30% for the Hindi Language Model

Completing sentences using Hindi Language Model

Hindi News Classifier

• Hindi News Dataset :

• Classifier Data Block

• News categories
• ‘business’ , ‘china’ , ‘entertainment’ , ‘india’ , ‘institutional’ , ‘international’ , ‘learningenglish’

• ‘multimedia’ , ‘news’ , ‘pakistan’ , ‘science’ , ‘social’ , ‘southasia’ , ‘sport’

• Metrics I choose for the News dataset was a f1_score

  average = macro

• Because there was a class imbalance in the news dataset

Final f_beta score of the classifier was 0.789

• Top losses :

System Requirements

Python v3.6.x
fastai v1
Pytorch v1.2.x
Numpy
Pandas
tqdm (Progress bar)
Jupyter Notebook (Visualisations)

Shoutout to

• Practical Deep Learning for Coders MOOC by team fast.ai