Sentiment Classification on the IMDB Movie Reviews dataset

Implementation details:

• Base Model: cardiffnlp/twitter-roberta-base-sentiment-latest
• Dataset: IMDB Dataset of 50K Movie Reviews dataset from Kaggle
• Number of epochs trained: 4
• Evaluation Metrics:
  
• Training Notebook: Link
• Benchmarking Notebook: Link
• Trained Model Checkpoints:
  • LoRA: Link
  • LoRA merged to base model: Link
  • ORT Optimized model: Link

Optimization:

Used Optimum(Huggingface) to convert model to ONXX and used the ORTOptimizer to perform the following optimizations (Guide):

• Basic general optimizations
• Transformers-specific fusions
• GELU approximation
• Mixed precision

Benchmark Results:

• Total samples for benchmarking: 50000
• Dataloader batch size: 32
• GPU: NVIDIA GeForce RTX 2080 Ti (11GB)

Model	Latency(ms)
Original finetuned model	151.81 ± 5.58
Optimized model	144.87 ± 1.96