[ASK] Training and retraining speed

[SnowHawkeye]

Problem Statement

Does the machine learning model (MIPHA kernel) converge faster with re-used / pre-trained feature extractors?

Context: #6

Deliverables

• Graph plotting training time / epochs / batches against model performance for different models:
  • ML model (e.g. simple CNN, ResNet, LSTM)
  • Same ML model with pre-trained feature extractors (on a different (part of the) dataset)

We expect the pre-trained features to make the convergence faster.

Approaches

Experiments needed to address the question

• [ ]

Understanding and Exploration

[anshuchowdaryalapati]

If you need a quick solution with limited computational resources, retraining the classifier might be the way to go. However, if you aim for the best possible performance and have the necessary resources, fine-tuning the feature extractor is likely the better choice.

[SnowHawkeye]

If you need a quick solution with limited computational resources, retraining the classifier might be the way to go. However, if you aim for the best possible performance and have the necessary resources, fine-tuning the feature extractor is likely the better choice.

@anshuchowdaryalapati Thank you for your input! I also realized my question was inaccurate (I'm still in the process of defining the issues). Your comment helped me find a better formulation