https://github.com/nikhil-swamix/jnu-ml-final-project

https://github.com/nikhil-swamix/jnu-ml-final-project

Last synced: 15 days ago
JSON representation

• Host: GitHub
• URL: https://github.com/nikhil-swamix/jnu-ml-final-project
• Owner: nikhil-swamix
• Created: 2020-12-11T21:10:18.000Z (about 4 years ago)
• Default Branch: main
• Last Pushed: 2020-12-14T01:23:59.000Z (about 4 years ago)
• Last Synced: 2024-04-23T02:31:03.220Z (9 months ago)
• Language: Jupyter Notebook
• Size: 4.55 MB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
### Problem Statement : Covid-19 tweet sentiment Analysis

#### Best model Notebook:: 0.best_model.ipynb Probably first file in repo

[please see here](https://colab.research.google.com/drive/1Gq3_w9l2kPWKP_CQJjnnHB1mr7XiW9-M?usp=sharing) to evaluate in colab, and get realtime results.

#### Summary File :0z_bestmodel_summary.ipynb, probably second file in repo.

# How to use

- open in colab and run instance 

- download and extract this model ,only works for roberta [MODEL LINK outputs.zip](https://nikhil-colab-bucket.s3-us-west-2.amazonaws.com/outputs.zip) in ./outputs folder as it is.

- run as usual and comment out any ! commands using aws (amazon web services)

- do not train the model in! we are using trained weights. to avoid training comment out section 'START TRAINING'

![IMAGE](https://www.codemotion.com/magazine/wp-content/uploads/2020/05/bert-google-896x504.png)

# Summary

### Highlights

- Our model is perhaps the most versatile and sophisticated in our batch.

- Highest model Accuracy is 90% and may go up, original model was trained for 2hrs or 10 epochs. 

- Use of AWS services [S3] in Program to fully automate ML Pipeline and Save/load Of checkpoints, security increased by clearing output whenever we enter sensitive password or access keys.

- We can dray conclusions faster by parallely running models on cloud, and saving checkpoints in AWS instead of colab which refreshes and loses data.

# *Individual Model Summary* 

outputs and details each model tested has been put below in form of brief summary. 

All Model Links in the API available for use Can Be found Here : [PETRAINED MODELS](https://huggingface.co/transformers/pretrained_models.html)

## Model:  RoBERTa-base [A Robustly Optimized BERT Pretraining Approach]

  -Accuracy

  

    91%

  -Notebook

  

   https://github.com/BinarySwami-10/JNU-ML-Final-Project/blob/main/Mod_Roberta_Acc_85_Param_120M.ipynb

  -Details

  

    12-layer, 768-hidden, 12-heads, 125M parameters RoBERTa using the BERT-base architecture

## Model: DistilRoBERTa [Remove Insignificant Parameters from RoBERTa]

  -Accuracy

  

    82.91205897840969 %

  

  -Notebook

  

   https://github.com/BinarySwami-10/JNU-ML-Final-Project/blob/main/Mod_distilroberta_Acc_82_Param_80M.ipynb

    

  -Details

       	

    6-layer, 768-hidden, 12-heads, 82M parameters, The DistilRoBERTa model distilled from the RoBERTa model roberta-base checkpoint.

## Model: XLNet-base [Generalized Autoregressive Pretraining for Language Understanding]

  -Accuracy

  

    79%

  -Notebook

  

   https://github.com/BinarySwami-10/JNU-ML-Final-Project/blob/main/Mod_XLnet_Acc_79_Param_110M.ipynb

    

  -Details

    

    12-layer, 768-hidden, 12-heads, 110M parameters. XLNet English model

## Model: XLNet-Large [Generalized Autoregressive Pretraining for Language Understanding]

  -Accuracy

    

    75.92550265792721 %

  

  -Notebook

    

   https://github.com/BinarySwami-10/JNU-ML-Final-Project/blob/main/Mod_XLnet_large_Acc_75.ipynb

  

  -Details

    

    24-layer, 1024-hidden, 16-heads, 340M parameters. XLNet Large English model