ai-labs

If you want to run ML tools and their prerequisites/dependencies in your experimentation environment, what would be the shortest path? Here we collected examples of experiments that you can relatively simply execute even without experience in even installing Python.

Inside an Artificial Neural Network

lab-contents/001_inside_an_artificial_neural_network.

First Machine Learning Experiments

Problem Class	Training/Inference	Environement	ML Toolset	Experiment
LLM	inference	cloud	Gemini 1.5	Section
LLM	inference	cloud	Gemini 2.0	Section
LLM	prompt with image	cloud	Gemini 2.0	Section
LLM	inference	cloud VM	Llama 4 Scout	Page
Tabular	training and inference	docker	PyTorch, fastai	Section
Tabular	training and inference	docker	PyTorch, fastai, Jupiter	Section
visual	training and inference	docker	PyTorch Lightning, Jupiter	Page
visual	training and inference	cloud VM	PyTorch Lightning, Jupiter	Page
visual	training and inference	cloud VM	PyTorch Lightning, CLI	Page
visual	training and inference	Macbook	PyTorch Lightning, Jupiter	Page

Ways to Execute ML Training and Inference

First, let's go through some methods of executing the ML processes without preliminary installing prerequisites in your physical environment, like Laptop.

Using ML Cloud Providers

Machine learning cloud providers allow you using the most powerful models that might be quite impossible for you to run otherwise.

flowchart TB
  subgraph provider[ML Cloud Provider]
    service[ML Service]
  end
  laptop-->|Calling API|service

Loading

Using Docker

Use container images that already have such preinstalled software as Python, PyTorch, fastai, Pandas, Jupiter, etc.

flowchart TB
    container_image-->|saved to|docker_hub
    subgraph laptop[Your Laptop]
      persistend_files[Persistent Files]
      container[Disposable Container]-->|mounts|persistend_files
    end
    docker_hub[Docker Hub]
    container-->|pulled from|docker_hub
    subgraph container_image[Container Image]
      Python
      PyTorch
      fastai
      Pandas
      Jupiter
    end

Loading

Prerequisite for using this approach is Docker installed in Mac, Linux or WSL (Windows) environment.

Using Lab VMs

When running on a local docker takes too much resources or too much time, an option might be running the load in the cloud.

flowchart TB
  subgraph provider[Cloud Provider]
    VM[GPU-Accelerated VM]
  end
  laptop-->|remotely control|VM

Loading

Prerequisite for using this approach is having installed tools for remote control of Cloud provider such as Azure.

Examples of LLM models and their requirements:

LLM Model	VRAM required
DeepSeek-R1-Distill-Qwen-1.5B	1 GB
Llama 3.2 3b-instruct-fp16	6.4 GB
Llama 3.3 70b-instruct-fp16	160 GB
Llama 4 Scout	210 GB
Mistral-Large-Instruct-2407	250 GB
Llama 4 Maverick	790 GB
Llama 3.1 405b	930GB
DeepSeek-R1	1500 GB

Examples of Azure VM sizes that can be used for ML training and inference:

VM Size	GPU Type	GPU Memory	GPUs	Price per Hour
Standard_NC4as_T4_v3	Nvidia T4	16 GB	1	$0.6
Standard_NC64as_T4_v3	Nvidia T4	64 GB	4	$5
Standard_NC24ads_A100_v4	Nvidia A100 80GB	80 GB	1	$4.6
Standard_NC40ads_H100_v5	Nvidia H100 94GB	94 GB	1	$9
Standard_NC48ads_A100_v4	Nvidia A100 80GB	160 GB	2	$9
Standard_NC80adis_H100_v5	Nvidia H100 94GB	188 GB	2	$14
Standard_NC96ads_A100_v4	Nvidia A100 80GB	320 GB	4	$20
Standard_ND96isr_H100_v5	Nvidia H100 80GB	640 GB	8	$127
Standard_ND96isr_H200_v5	Nvidia H200 141GB	1128 GB	8	$110
Standard_ND96isr_MI300X_v5	AMD MI300X 192GB	1535 GB	8	$67