Description
A course to train and evaluate neural networks with PyTorch and TensorFlow on your own infrastructure: your GPUs, Linux servers or IBM Power, not a notebook in the cloud. Every experiment is measured for quality, memory and time, so you know when an improvement is real and how much compute it costs.
Course details
| SIXE code: SXIA04 | Category: Artificial Intelligence / Deep learning |
| Delivery: Online and on-site | Duration: 3 days |
Who it is for
Data scientists, machine learning developers and R&D teams who need to train or adapt models with their own data and on their own machines.
Prerequisites
Confident Python and basic machine learning (what training and validation sets are). Basic linear algebra.
What you will learn
- Training a network from scratch: tensors, automatic differentiation, batches, loss function and optimiser, in PyTorch and its TensorFlow equivalent.
- Transfer instead of starting from zero: start from a pre-trained model, freeze layers and compare it with your own network on the same budget.
- Spotting overfitting: learning curves, regularisation and errors broken down by data group.
- Measuring the cost: GPU memory, batch size, mixed precision and time per epoch on your hardware.
- Reproducing the result: seeds, saved configuration and checkpoints someone else can repeat.
You work on a text or image classification case and finish with three comparable experiments, their curves and the compute budget for each.
What you take away
- A training and evaluation notebook with its configuration.
- Three comparable experiments with annotated curves.
- Checkpoint, model card and compute estimate.
If the model will end up being served on your own servers, the next step is the local LLM and GPU infrastructure course; and if data is the bottleneck, see how we approach storage for AI and HPC.
Frequently asked questions
How long is the course?
3 days.
Is it online or on-site?
Both: at your offices or remotely, with a live instructor.
What do I need to know beforehand?
Confident Python and basic machine learning (what training and validation sets are). Basic linear algebra.
Do we train a large language model from scratch?
No. You work with small networks and pre-trained models to understand training, errors and compute cost.
Is a GPU needed?
The first labs run on CPU. To measure memory and batch size a GPU is used: yours or a shared lab GPU.
Can it be tailored to my team?
Yes. We adapt the examples and labs to your team’s tools and processes, and it can be combined with other AI courses.
