-
Notifications
You must be signed in to change notification settings - Fork 0
Environment and Execution Notes
This project was primarily developed and tested on a local laptop with the following specs:
- CPU: Intel(R) Core(TM) i7-1065G7 @ 1.30GHz (Quad Core)
- RAM: 32 GB
- OS: Windows 10 Home, 64-bit (Version 22H2)
Although this machine is not equipped with a dedicated GPU, it was sufficient for CPU-based training and running multiple lightweight simulations in parallel.
During the early stages of model development, frequent retraining and trial-and-error are often required due to:
- Model instability
- Hyperparameter sensitivity
- The rarity of successful training outcomes
- Random search being used for tuning
To avoid cloud computing costs during this exploratory phase, training was executed on a local laptop, using multi-core parallel processing where possible.
To make better use of CPU resources, I implemented multi-core parallel training with automatic monitoring of memory usage. The scripts restart themselves when memory usage exceeds a predefined threshold.
Key tips include:
- Limit the number of parallel processes to avoid thermal throttling.
- Consider inserting delays (e.g.,
time.sleep) between monitoring cycles. - Check CPU temperature periodically. Excessive heating can reduce performance or, in extreme cases, damage the device.
- Fan control software and thermal monitoring tools (like Core Temp or HWMonitor) are recommended.
For heavier models or longer training episodes, consider transitioning to:
- Google Colab Pro: Offers persistent sessions and access to GPU/TPU.
- AWS SageMaker: Recommended if you're moving toward a production-grade workflow, especially for reproducibility and deployment pipelines.
- Upload and document the parallel execution script used for random hyperparameter search.
- Measure and compare run-time efficiency on Colab vs. local machine.
- Possibly record CPU temperature during training and log it to detect overheating patterns.