🤖 AI & Networking Frameworks

Domain-specific research in applied AI and network engineering requires specialized tools beyond the standard development stack. This page catalogs the frameworks and platforms that define modern research in these areas.

1. Deep Learning Frameworks

PyTorch (Primary Recommendation)

PyTorch is the dominant framework in academic ML research. Its define-by-run (dynamic) computation graph is more intuitive for research than TensorFlow’s static graph model.

# Install with CUDA support (always get exact command from pytorch.org)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

Key PyTorch research libraries:

Library	Purpose
PyTorch Lightning	Training loop abstraction; removes boilerplate
Hugging Face Transformers	Pre-trained models for NLP, vision, audio
timm	1000+ pre-trained vision models
TorchMetrics	Standardized evaluation metrics
einops	Tensor manipulation with readable notation
torchinfo	Model summary and parameter counting

TensorFlow / JAX

Framework	Best For
TensorFlow 2	Production deployment, TensorFlow Serving, TFLite
JAX	High-performance research needing XLA compilation, functional transforms
Flax	Neural networks in JAX (Google Brain style)

2. Large Language Models

Local Inference

Running LLMs locally preserves data privacy and eliminates API costs:

Ollama — The easiest way to run LLMs locally. Supports Llama, Mistral, Gemma, Phi, and more.

# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# Pull and run a model
ollama pull llama3.2
ollama run llama3.2

# OpenAI-compatible API (localhost:11434)
curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "llama3.2", "messages": [{"role": "user", "content": "Hello"}]}'

llama.cpp — Efficient C++ inference for quantized LLMs. Runs 7B models on a MacBook; 70B models on a single consumer GPU.

git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp && make -j $(nproc)
./llama-cli -m models/llama-3.2-7b-q4_K_M.gguf -p "Explain transformers:"

High-Throughput Serving

vLLM — PagedAttention-based serving for GPU clusters. Best throughput for multi-user research API deployments.

pip install vllm
python -m vllm.entrypoints.openai.api_server \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --tensor-parallel-size 2

Fine-Tuning and PEFT

Tool	Purpose
PEFT	LoRA, QLoRA, prefix tuning, adapters
TRL	RLHF, DPO, SFT training
LLaMA-Factory	Unified fine-tuning for 100+ LLMs
Unsloth	2× faster LoRA fine-tuning, 50% less VRAM

3. Computer Vision

Object Detection and Segmentation

Ultralytics YOLO — State-of-the-art real-time detection, segmentation, pose estimation, and classification. The industry standard for applied CV research.

from ultralytics import YOLO

model = YOLO("yolo11n.pt")           # Load nano model
results = model("image.jpg")         # Inference
model.train(data="coco8.yaml", epochs=100)  # Training

Detectron2 — Meta’s research platform for object detection and segmentation. Reference implementation for Mask R-CNN, Panoptic FPN.
MMDetection — OpenMMLab’s modular detection framework with 50+ architectures.

Vision-Language Models

Model	Organization	Best For
LLaVA	UW Madison	Visual question answering, fine-tuning
CLIP	OpenAI	Zero-shot image classification, retrieval
Florence-2	Microsoft	Multi-task vision understanding

4. Network Research Tools

Traffic Analysis and Capture

# Capture traffic on an interface
tcpdump -i eth0 -w capture.pcap

# Analyze capture in Wireshark
wireshark capture.pcap

# Generate synthetic traffic for benchmarks
iperf3 -s                                    # Server
iperf3 -c server_ip -t 60 -P 4 -u -b 10G   # Client (UDP, 4 parallel streams)

Network Emulation

Mininet — Software-defined network emulator. Create arbitrary network topologies for SDN and networking research.

from mininet.net import Mininet
from mininet.topo import SingleSwitchTopo

net = Mininet(topo=SingleSwitchTopo(4))
net.start()
net.pingAll()
net.stop()

ns-3 — Discrete-event network simulator. The gold standard for academic network protocol simulation.
GNS3 — Graphical network simulator integrating real OS images (Cisco IOS, Linux).

Proxy and Tunnel Infrastructure

For research involving censorship, geo-restrictions, or university firewall bypass:

Tool	Protocol	Use Case
frp	TCP/UDP multiplexing	Expose local dev server behind NAT/firewall
mihomo (MetaCubeX)	Multi-protocol rule-based	Advanced routing with TUN mode, DNS hijacking
sing-box	Universal proxy platform	Lightweight Go-based tunnel for custom networks
WireGuard	Modern VPN	Low-overhead research VPN between lab nodes
OpenWrt	Linux router firmware	Network engineering research at gateway level

frp Setup Example

# frps.ini (server, needs public IP)
[common]
bind_port = 7000
token = your_secret_token

# frpc.ini (client, your local machine)
[common]
server_addr = your.server.com
server_port = 7000
token = your_secret_token

[jupyter]
type = tcp
local_ip = 127.0.0.1
local_port = 8888
remote_port = 18888

5. MLOps and Model Deployment

Experiment Tracking

See research/experiment-tracking.md for a detailed comparison.

Model Serving

Tool	Scale	Best For
Gradio	Prototype	Research demos, rapid UI for models
Streamlit	Prototype	Data dashboards, interactive demos
FastAPI	Production	High-performance REST API for model serving
BentoML	Production	ML model packaging and deployment
Triton Inference Server	GPU cluster	NVIDIA’s high-throughput GPU serving