| 일 | 월 | 화 | 수 | 목 | 금 | 토 |
|---|---|---|---|---|---|---|
| 1 | 2 | |||||
| 3 | 4 | 5 | 6 | 7 | 8 | 9 |
| 10 | 11 | 12 | 13 | 14 | 15 | 16 |
| 17 | 18 | 19 | 20 | 21 | 22 | 23 |
| 24 | 25 | 26 | 27 | 28 | 29 | 30 |
| 31 |
- 컴퓨터시스템응용기술사
- chrome
- iF
- 정보관리기술사 #컴퓨터시스템응용기술사
- 브루트포스
- codyseey
- 하버드구조
- 컴퓨터시스테응용기술사
- 정보관리기술사
- 컴퓨터시슴테응용기술사
- CPU
- 정보관리시술사
- 코디세이
- 백준
- 크롬확장프로그램
- 알고리즘
- 정보관리기술사 #seongyle
- Cam
- onnx
- 폰노이만구조
- Today
- Total
yeonseong
ONNX 본문
ONNX (Open Neural Network Exchange)
1. 개요
ONNX(Open Neural Network Exchange)는 머신러닝 모델의 상호운용성을 실현하기 위해 개발된 개방형 표준 포맷이다[1][2]. 2017년 Microsoft와 Facebook이 주도하여 개발한 이 기술은[1][3], 서로 다른 AI 프레임워크 간의 모델 이식성을 보장하는 공통 표현 언어 역할을 수행한다[1][2].
현대 AI 개발 환경에서는 PyTorch, TensorFlow, Keras 등 다양한 프레임워크가 존재하며, 각 프레임워크는 고유한 모델 형식과 연산자를 사용한다[4]. 이러한 이질적인 환경에서 ONNX는 학습 프레임워크와 배포 환경 사이의 격차를 해소하고, 모델의 재학습 없이 다양한 플랫폼에서 실행 가능하게 하는 핵심 인프라로 자리 잡았다[1][5].
AI 소프트웨어 및 도구 시장이 2024년 1,271억 달러에서 2029년 4,399억 달러로 연평균 28.2% 성장할 것으로 전망되는 가운데[6], ONNX는 AI 상호운용성의 기술적 초석으로서 그 중요성이 더욱 부각되고 있다[3].
2. ONNX의 핵심 구성요소 및 기술 원리
2.1 중간 표현(Intermediate Representation, IR)
ONNX의 핵심은 계산 그래프 기반의 중간 표현이다[7]. 이는 신경망 모델을 방향성 비순환 그래프(DAG, Directed Acyclic Graph) 형태로 추상화하여 표현하며[3], 노드는 연산자(operator)를, 엣지는 텐서 데이터 흐름을 나타낸다[4].
IR은 세 가지 주요 요소로 구성된다[1]:
- 계산 그래프 모델: 신경망의 구조와 데이터 흐름을 그래프로 표현
- 표준 연산자 정의: Conv, ReLU, MatMul, Reshape 등 표준화된 연산자 집합[8]
- 데이터 타입 명세: float, int, 다차원 텐서 등 표준 데이터 타입[4]
2.2 연산자(Operator) 및 Opset 버전 관리
ONNX는 표준 연산자 집합을 정의하고, 이를 opset(operator set) 버전으로 관리한다[9][8]. 각 연산자는 명확한 수학적 의미와 입출력 스펙을 가지며, opset 버전은 하위 호환성을 유지하면서 점진적으로 확장된다[9][7].
Opset 관리 메커니즘의 특징은 다음과 같다[9][8]:
- 도메인별 버전 관리: 기본 도메인(ai.onnx)과 전통적 ML 도메인(ai.onnx.ml) 등 도메인별로 독립적인 버전 체계 운영
- 단조 증가 버전: 연산자 추가, 제거, 의미 변경 시 버전 증가
- 역호환성 보장: 이전 버전 모델의 지속적 실행 가능성 확보
2.3 Protobuf 기반 직렬화
ONNX 모델은 Protocol Buffers(Protobuf) 형식으로 직렬화되어 .onnx 파일로 저장된다[7]. Protobuf는 언어 중립적이고 플랫폼 독립적인 데이터 직렬화 방식으로, 효율적인 저장과 전송을 가능하게 한다[7].
3. ONNX Runtime 및 실행 메커니즘
3.1 ONNX Runtime 아키텍처
ONNX Runtime은 Microsoft가 개발한 고성능 추론 엔진으로[1][10], ONNX 모델을 다양한 하드웨어에서 효율적으로 실행하는 역할을 담당한다[11]. 주요 처리 단계는 다음과 같다[5]:
- 모델 로딩: ONNX 파일을 메모리 내 그래프 표현으로 변환
- 그래프 최적화: 노드 융합(node fusion), 상수 폴딩(constant folding), 불필요 노드 제거 등 플랫폼 독립적 최적화 수행[11][12]
- 그래프 파티셔닝: 사용 가능한 Execution Provider 기반으로 서브그래프 분할[11][13]
- 하드웨어별 실행: 각 Execution Provider가 할당된 서브그래프를 최적화된 방식으로 실행[13]
3.2 Execution Provider(EP) 프레임워크
Execution Provider는 ONNX Runtime의 핵심 확장 메커니즘으로, 다양한 하드웨어 가속을 가능하게 한다[14][13]. 주요 EP로는 다음이 있다:
| Execution Provider | 대상 하드웨어 | 주요 특징 |
|---|---|---|
| CPU (MLAS) | 범용 CPU | 기본 EP, OpenMP 기반 멀티스레딩[12] |
| CUDA | NVIDIA GPU | CUDA 라이브러리 활용, 고성능 병렬 연산[12] |
| TensorRT | NVIDIA GPU | 추론 최적화, INT8 양자화 지원[15] |
| OpenVINO | Intel CPU/GPU/VPU | Intel 하드웨어 최적화[14] |
| CoreML | Apple Silicon | iOS/macOS 최적화[5] |
| NNAPI | Android | Android 하드웨어 가속[16] |
| DirectML | Windows GPU | DirectX 12 기반 범용 GPU 가속[1] |
EP 프레임워크는 GetCapability() 인터페이스를 통해 각 EP가 지원 가능한 노드를 파악하고 할당하는 방식으로 작동한다[13]. 지원되지 않는 연산은 자동으로 CPU로 폴백되어 실행 안정성을 보장한다[17].
3.3 그래프 최적화
ONNX Runtime은 세 가지 수준의 그래프 최적화를 제공한다[12][18]:
- ORT_ENABLE_BASIC: 상수 폴딩 등 기본 최적화
- ORT_ENABLE_EXTENDED: 노드 융합 등 확장 최적화
- ORT_ENABLE_ALL: 레이아웃 최적화를 포함한 모든 최적화
최적화는 온라인(세션 시작 시) 또는 오프라인(사전 최적화 후 저장) 방식으로 적용 가능하며[19], 오프라인 방식은 모델 시작 시간을 대폭 단축시킨다[20][19].
4. 프레임워크 간 모델 변환 및 상호운용성
4.1 주요 프레임워크 지원
ONNX는 광범위한 AI 프레임워크와의 통합을 지원한다[21][5]:
학습 프레임워크:
- PyTorch:
torch.onnx.export()함수로 직접 내보내기 지원[2][22] - TensorFlow:
tf2onnx변환 도구 활용[4][23] - Keras: TensorFlow 백엔드를 통한 변환[4]
- Scikit-Learn:
skl2onnx변환 도구[4][5]
배포 런타임:
- ONNX Runtime[1]
- TensorRT (NVIDIA)[5]
- OpenVINO (Intel)[21]
- CoreML (Apple)[5]
- TensorFlow Lite[24]
4.2 PyTorch에서 ONNX로 변환
PyTorch는 ONNX 내보내기를 위한 내장 지원을 제공한다[2]. 변환 과정은 추적(tracing) 방식으로 작동하며, 샘플 입력으로 모델을 실행하여 연산 그래프를 기록한다[2]:
import torch
import torch.onnx
model = MyModel()
dummy_input = torch.randn(1, 3, 224, 224)
torch.onnx.export(
model,
dummy_input,
"model.onnx",
input_names=['input'],
output_names=['output'],
opset_version=14,
dynamic_axes={'input': {0: 'batch_size'}}
)
dynamic_axes 매개변수는 배치 크기 등 가변 차원을 지정하여 유연한 추론을 가능하게 한다[2].
4.3 ONNX 모델 실행
ONNX Runtime을 사용한 추론은 간결한 API로 수행된다[2]:
import onnxruntime as ort
session = ort.InferenceSession(
"model.onnx",
providers=['CUDAExecutionProvider', 'CPUExecutionProvider']
)
outputs = session.run(
['output'],
{'input': input_data.numpy()}
)
5. ONNX의 특징 및 장단점
5.1 주요 장점
1. 프레임워크 독립성
ONNX는 특정 프레임워크에 종속되지 않는 중립적 포맷을 제공하여, 연구 단계와 프로덕션 환경 간의 원활한 전환을 가능하게 한다[1][4].
2. 하드웨어 최적화 및 가속
다양한 하드웨어 벤더의 최적화 라이브러리를 활용하여, CPU, GPU, NPU, FPGA 등 다양한 하드웨어에서 최적 성능을 달성할 수 있다[1][25].
3. 배포 간소화
모델 변환 과정을 제거하고 단일 포맷으로 다양한 플랫폼에 배포 가능하여, 클라우드에서 엣지 디바이스까지 일관된 배포 파이프라인을 구축할 수 있다[25].
4. 성능 향상
ONNX Runtime은 PyTorch 대비 최대 2-5배의 추론 속도 향상을 달성할 수 있으며[1][10], 메모리 사용량도 감소시킨다[18].
5. 커뮤니티 및 생태계
Microsoft, NVIDIA, Intel, AMD, Qualcomm 등 주요 기업들의 지원과 오픈소스 커뮤니티의 활발한 기여로 지속적으로 발전하고 있다[1][3].
5.2 한계점 및 과제
1. 연산자 지원 제한
모든 프레임워크의 모든 연산을 완벽히 지원하지는 못하며, 특히 최신 또는 커스텀 연산자의 경우 변환 실패가 발생할 수 있다[2][26].
2. 제어 흐름의 복잡성
동적 제어 흐름을 포함하는 모델의 경우 추적(tracing) 방식의 한계로 인해 모든 실행 경로를 캡처하지 못할 수 있다[2].
3. 프레임워크별 불일치
변환된 모델이 원본 모델과 수치적으로 완전히 동일하지 않을 수 있으며, 철저한 검증이 필요하다[2][26].
4. 버전 호환성 문제
ONNX 버전, opset 버전, ONNX Runtime 버전, 프레임워크 버전 간의 호환성 관리가 복잡하며, 버전 불일치 시 오류가 발생할 수 있다[27][28].
5. LLM 및 Transformer 모델 지원
대규모 언어 모델(LLM)의 경우 동적 토큰 생성, 복잡한 양자화(4비트, 8비트), KV 캐싱, Flash Attention 등의 최적화 기법 지원이 제한적이었으나[29], 최근 개선이 진행되고 있다[29].
6. 최적화 기법 및 실무 적용
6.1 모델 양자화(Quantization)
양자화는 모델의 정밀도를 낮춰 크기와 연산량을 줄이는 기법이다[30][31]. ONNX는 두 가지 양자화 방식을 지원한다:
- 정적 양자화(Static Quantization): 보정 데이터를 사용하여 가중치와 활성화를 사전 양자화. CNN 모델에 효과적[30]
- 동적 양자화(Dynamic Quantization): 가중치만 사전 양자화하고 활성화는 런타임에 양자화. Transformer 모델에 적합[30]
선택적 양자화(Selective Quantization)는 정확도 손실을 최소화하면서 모델 크기를 줄이는 고급 기법으로, TuneQn 같은 도구를 활용하여 레이어별로 양자화 적용 여부를 최적화할 수 있다[30]. 실험 결과 완전 양자화 모델 대비 정확도 손실을 최대 54.14% 감소시키면서 모델 크기를 72.9%까지 줄일 수 있다[30].
6.2 모델 프루닝(Pruning)
프루닝은 중요도가 낮은 파라미터나 구조를 제거하여 모델을 경량화하는 기법이다[32]. 구조적 프루닝(Structured Pruning)은 전체 레이어나 채널을 제거하여 추론 속도를 실질적으로 개선한다[32].
6.3 성능 튜닝
ONNX Runtime의 성능 최적화를 위한 주요 설정은 다음과 같다[12]:
CPU EP 최적화:
intra_op_num_threads: 모델 실행을 위한 스레드 수 조절execution_mode: 순차(ORT_SEQUENTIAL) 또는 병렬(ORT_PARALLEL) 실행 선택inter_op_num_threads: 병렬 실행 시 그래프 노드 간 병렬화 스레드 수
GPU EP 최적화:
- TensorRT EP 사용 시 엔진 캐싱으로 초기화 시간 단축: 빌드 없음(384초) → 타이밍 캐시(42초) → 엔진 캐시(9초) → 임베디드 엔진(1.9초)[15]
- CUDA EP에서
cudnn_conv_algo_search옵션으로 최적 컨볼루션 알고리즘 자동 탐색[12]
7. 활용 분야 및 적용 사례
7.1 산업별 활용 사례
1. Facebook AI Research
이미지 인식 모델을 ONNX로 변환하여 모바일 디바이스와 웹 브라우저에 배포, 접근성을 크게 향상[25].
2. Microsoft
음성 인식 모델을 ONNX Runtime으로 최적화하여 다양한 플랫폼에서 실시간 음성-텍스트 변환 구현[25].
3. Intel CPU 최적화
BERT 모델을 ONNX Runtime과 Intel CPU 코어에서 실행하여 처리량과 성능 향상 달성[1].
4. Hugging Face Optimum
ONNX Runtime과 통합하여 NLP 모델 추론 속도를 PyTorch 대비 5배 향상[1], Transformer 모델 레이턴시를 2.03배 감소[1].
5. Scikit-learn 가속화
전통적인 머신러닝 모델을 ONNX로 변환하여 5배의 성능 향상 달성[1].
7.2 엣지 AI 및 IoT 배포
ONNX는 엣지 디바이스와 모바일 환경에서 특히 강점을 발휘한다[33][6]. 제한된 연산 자원에서도 효율적인 추론이 가능하며, 프레임워크 비종속적 특성으로 다양한 하드웨어 아키텍처를 지원한다[33][6].
7.3 MLOps 통합
ONNX는 MLOps 파이프라인에서 중요한 역할을 수행한다[1]:
- 모델 서빙 플랫폼(Azure ML, TensorFlow Serving, TorchServe)과의 원활한 통합[1]
- 여러 프레임워크의 모델을 단일 포맷으로 관리하여 버전 관리 복잡도 감소[1]
- CI/CD 파이프라인에서 프레임워크 독립적인 모델 검증 및 배포 가능[3]
8. 최신 기술 동향 및 발전 방향
8.1 대규모 언어 모델(LLM) 지원 강화
ONNX는 LLM 시대의 요구사항에 대응하기 위해 진화하고 있다[29]:
- 동적 형상(Dynamic Shape) 지원: 가변 길이 시퀀스 처리 개선[29]
- 고급 양자화: 4비트, 8비트 양자화 지원 강화[29]
- 최적화 기법: KV 캐싱, Flash Attention 등 LLM 특화 최적화[29]
Liquid AI의 LFM2 언어 모델 시리즈가 ONNX 포맷으로 Hugging Face에 공개되는 등[29], LLM 생태계에서의 입지가 확대되고 있다.
8.2 멀티모달 모델 지원
2025년 현재 AI 발전 방향은 텍스트, 이미지, 음성 등 다양한 모달리티를 통합하는 멀티모달 모델로 향하고 있다[34][35]. ONNX도 이러한 추세에 맞춰 복잡한 멀티모달 아키텍처 지원을 강화하고 있다.
8.3 모델 압축 및 효율성
모델 크기와 연산량을 줄이는 기술이 지속적으로 발전하고 있다[35]:
- 양자화 및 프루닝 기법의 고도화[30][32]
- Sparse 모델 및 MoE(Mixture of Experts) 아키텍처 지원[36]
- 지식 증류(Knowledge Distillation)를 통한 경량화[37]
8.4 ONNX Model Zoo 전환
ONNX Model Zoo는 2025년 7월부터 Hugging Face로 전환되고 있다[38]. 이는 ML 커뮤니티의 모델 공유 중심지가 Hugging Face로 이동했음을 반영하며, ONNX 모델의 접근성과 공유를 더욱 향상시킬 것으로 기대된다[38].
8.5 하드웨어 가속 생태계 확장
NPU(Neural Processing Unit), VPU(Vision Processing Unit) 등 AI 전용 하드웨어에 대한 Execution Provider 지원이 지속적으로 확대되고 있다[16][17]. STM32MP2 시리즈의 VSInpu EP[17], Android NNAPI[16] 등이 그 예시다.
8.6 PyTorch 2.0의 영향
PyTorch 2.0은 torch.compile()과 Torch-TensorRT 같은 직접 변환 파이프라인을 도입하여, 일부 사용 사례에서 ONNX의 필요성을 감소시켰다[33]. 그러나 ONNX는 엣지 배포, 프레임워크 간 호환성, 표준화가 중요한 영역에서 여전히 필수적인 도구로 자리하고 있다[33][16].
9. 결론 및 제언
9.1 기술적 의의
ONNX는 AI 모델의 상호운용성과 이식성을 혁신적으로 개선한 표준으로, 다음과 같은 핵심 가치를 제공한다:
- 프레임워크 독립성: 특정 벤더나 도구에 종속되지 않는 중립적 생태계 구축
- 배포 효율성: 학습과 추론 환경의 분리로 최적의 도구 조합 가능
- 하드웨어 최적화: 다양한 하드웨어에서 최고 성능 달성
- 산업 표준화: 주요 기업과 커뮤니티의 지원으로 사실상의 표준으로 자리매김
9.2 실무 적용 시 고려사항
ONNX를 프로덕션 환경에 도입할 때 다음 사항을 고려해야 한다:
- 철저한 검증: 변환된 모델의 수치적 정확성을 원본과 비교 검증 필수[2][26]
- 버전 관리: ONNX, opset, 프레임워크, ONNX Runtime 버전 간 호환성 명확히 파악[27]
- 연산자 지원 확인: 모델의 모든 연산이 타겟 opset에서 지원되는지 사전 검증[2]
- 성능 프로파일링: 다양한 EP와 최적화 설정을 실험하여 최적 구성 도출[12]
- 점진적 도입: 파일럿 프로젝트로 시작하여 점진적으로 확대
9.3 미래 전망
AI 기술이 클라우드에서 엣지로, 단일 모달에서 멀티모달로, 거대 모델에서 효율적 모델로 진화하는 가운데, ONNX의 역할은 더욱 중요해질 것이다. LLM과 생성형 AI 지원 강화, 새로운 하드웨어 가속기 통합, 모델 최적화 기법 발전 등을 통해 ONNX 생태계는 지속적으로 성장할 것으로 전망된다[29][34].
특히 엣지 AI와 IoT 영역에서 ONNX는 프레임워크 비종속적 배포를 가능하게 하는 핵심 인프라로서, 산업 AI의 대중화와 민주화에 기여할 것이다[6][3]. 조직의 AI 전략 수립 시 ONNX 기반의 모델 관리 및 배포 체계 구축을 적극 검토할 것을 제안한다.
Sources
[1] Open Neural Network Exchange (ONNX) Explained https://www.splunk.com/en_us/blog/learn/open-neural-network-exchange-onnx.html
[2] ONNX for PyTorch & TensorFlow Model Conversion https://apxml.com/courses/pytorch-for-tensorflow-developers/chapter-5-pytorch-model-persistence-deployment/onnx-pytorch-tensorflow-interoperability
[3] ONNX: Towards standardized interoperability of artificial ... https://datascientest.com/en/all-bout-onnx
[4] Understanding ONNX: Enhancing AI Model Interoperability ... https://encord.com/blog/onnx-open-neural-network-exchange-format/
[5] ONNX Explained: A New Paradigm in AI Interoperability https://viso.ai/computer-vision/onnx-explained-a-new-paradigm-in-ai-interoperability/
[6] ONNX Embedded AI: Scalable Solutions for Edge Devices https://vietsol.com.vn/onnx-embedded-ai-scalable-runtime-vietsol-edge/
[7] ONNX | Jimmy Song https://jimmysong.io/en/ai/onnx/
[8] ONNX Concepts - ONNX 1.21.0 documentation https://onnx.ai/onnx/intro/concepts.html
[9] ONNX Versioning - ONNX 1.21.0 documentation https://onnx.ai/onnx/repo-docs/Versioning.html
[10] Optimum+ONNX Runtime - Easier, Faster training for your ... https://huggingface.co/blog/optimum-onnxruntime-training
[11] ONNX Runtime | onnxruntime https://onnxruntime.ai/docs/
[12] ONNX Runtime Performance Tuning https://iot-robotics.github.io/ONNXRuntime/docs/performance/tune-performance.html
[13] ONNX Runtime Execution Providers https://onnxruntime.ai/docs/execution-providers/
[14] Now available: ONNX Runtime 0.5 with support for edge ... https://opensource.microsoft.com/blog/2019/08/26/announcing-onnx-runtime-0-5-edge-hardware-acceleration-support
[15] TensorRT Execution Provider - NVIDIA https://onnxruntime.ai/docs/execution-providers/TensorRT-ExecutionProvider.html
[16] “Top 10 Edge AI Frameworks for 2025: Best Tools for Real- ... https://blog.huebits.in/top-10-edge-ai-frameworks-for-2025-best-tools-for-real-time-on-device-machine-learning/
[17] How to use hardware acceleration with TensorFlow Lite and ... https://wiki.st.com/stm32mpu/wiki/How_to_use_hardware_acceleration_with_TensorFlow_Lite_and_ONNX_Runtime_frameworks
[18] ONNX vs PyTorch Speed: In-Depth Performance Comparison https://dev-kit.io/blog/machine-learning/onnx-vs-pytorch-speed-comparison
[19] End-to-End AI for NVIDIA-Based PCs: ONNX Runtime and ... https://developer.nvidia.com/blog/end-to-end-ai-for-pcs-onnx-runtime-and-optimization/
[20] ONNX Runtime: performant on-device inferencing https://opensource.microsoft.com/blog/2023/02/08/performant-on-device-inferencing-with-onnx-runtime
[21] Open Neural Network Exchange https://itwiki.kr/w/Open_Neural_Network_Exchange
[22] Interoperability Between Deep Learning Toolbox, ... https://www.mathworks.com/help/deeplearning/ug/interoperability-between-deep-learning-toolbox-tensorflow-pytorch-and-onnx.html
[23] What is ONNX, and why is it used? https://milvus.io/ai-quick-reference/what-is-onnx-and-why-is-it-used
[24] Optimize Production with PyTorch/TF, ONNX, TensorRT & ... https://www.digitalocean.com/community/tutorials/ai-model-deployment-optimization
[25] Boosting Model Interoperability and Efficiency with the ... https://www.design-reuse.com/article/61447-boosting-model-interoperability-and-efficiency-with-the-onnx-framework/
[26] A Case Study in the ONNX Ecosystem https://arxiv.org/html/2303.17708v3
[27] A conflict doc abount compatibility between Onnx Version ... https://github.com/onnx/onnx/issues/6134
[28] What are the limitations of ONNX models compared ... https://massedcompute.com/faq-answers/?question=What+are+the+limitations+of+ONNX+models+compared+to+PyTorch+models%3F
[29] ONNX makes a comeback for LLMs: dynamic shapes and ... https://www.linkedin.com/posts/mlech26l_onnx-communitylfm2-12b-onnx-hugging-face-activity-7351983839264534529-m-7S
[30] Selective Quantization Tuning for ONNX Models https://arxiv.org/html/2507.12196v1
[31] Quantize ONNX models | onnxruntime https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html
[32] A Comprehensive Guide to Neural Network Model Pruning https://datature.com/blog/a-comprehensive-guide-to-neural-network-model-pruning
[33] The Rise and Fall of ONNX (feat. PyTorch 2.0) https://blog.squeezebits.com/the-rise-and-fall-of-onnx-feat-pytorch-20-42184
[34] The Future of Large Language Models in 2025 | LLM ... https://www.prodshell.com/blog/llm-2025
[35] Large Language Models In 2025: Your Guide To Next-Gen AI https://acecloud.ai/blog/large-language-models/
[36] Best Open Source LLMs for 2025: Your Complete Guide https://collabnix.com/the-10-best-open-source-llms-for-2025-your-complete-guide-to-free-language-models/
[37] [AI]Optimization Technique 01 https://brunch.co.kr/@leadbreak/1
[38] onnx/models: A collection of pre-trained, state-of-the-art ... https://github.com/onnx/models
[39] onnx/onnx: Open standard for machine learning ... https://github.com/onnx/onnx
[40] Leveraging ONNX: Seamless Integration Across AI Frameworks https://www.cohorte.co/blog/leveraging-onnx-seamless-integration-across-ai-frameworks
[41] ONNX | Home https://onnx.ai
[42] News https://onnx.ai/news.html
[43] Edge AI: TensorFlow Lite vs. ONNX Runtime vs. PyTorch ... https://dzone.com/articles/edge-ai-tensorflow-lite-vs-onnx-runtime-vs-pytorch
[44] Profile of Open Neural Network Exchange https://pypi.org/org/onnx/
[45] Tune performance | onnxruntime - GitHub Pages https://fs-eire.github.io/onnxruntime/docs/performance/tune-performance/
[46] Can I use ONNX models with multiple frameworks like ... https://massedcompute.com/faq-answers/?question=Can+I+use+ONNX+models+with+multiple+frameworks+like+TensorFlow+and+PyTorch%3F
[47] ONNX — Open Neural Network Exchange https://blog.stackademic.com/onnx-open-neural-network-exchange-29f39a84c5f2
[48] exportONNXNetwork - Export network to ONNX model format https://www.mathworks.com/help/deeplearning/ref/exportonnxnetwork.html
[49] Understanding the ONNX Format - Metric Coders https://www.metriccoders.com/post/understanding-the-onnx-format
[50] Introducing the Open AI Accelerator Exchange https://www.networkoptix.com/blog/2024/03/05/introducing-the-open-ai-accelerator-standard
[51] Industrial AI market: 10 insights on how AI is transforming ... https://iot-analytics.com/industrial-ai-market-insights-how-ai-is-transforming-manufacturing/
[52] ONNX with Python - ONNX 1.21.0 documentation https://onnx.ai/onnx/intro/python.html
[53] What are the key differences between ONNX and ... https://massedcompute.com/faq-answers/?question=What+are+the+key+differences+between+ONNX+and+TensorFlow+Lite%3F
[54] Pruning and quantization in computer vision: A quick guide https://www.ultralytics.com/blog/pruning-and-quantization-in-computer-vision-a-quick-guide
[55] What are the limitations of ONNX model compatibility with ... https://massedcompute.com/faq-answers/?question=What+are+the+limitations+of+ONNX+model+compatibility+with+large+language+models%3F
[56] ONNX related – General — Olive documentation https://microsoft.github.io/Olive/0.2.0/tutorials/passes/onnx.html
[57] NVIDIA/TensorRT-Model-Optimizer https://github.com/NVIDIA/TensorRT-Model-Optimizer
[58] ONNX - Model Zoo https://www.tutorialspoint.com/onnx/onnx-model-zoo.htm
[59] The ONNX Model Zoo is a collection of pre-trained ... https://github.com/ankkhedia/model-zoo
[60] Execution providers https://ort.pyke.io/perf/execution-providers
[61] Contributing to ONNX: How you can improve Machine ... https://programme.europython.eu/europython-2024/talk/KLXQAM/
[62] ONNX Model Hub - ONNX 1.21.0 documentation https://onnx.ai/onnx/repo-docs/Hub.html
[63] Standardization and Development of ONNX (Open Neural ... https://events.it4i.cz/event/302/attachments/721/2422/20250227_SPACE_ONNX_Fehlner_final.pdf
[64] ONNX models https://learn.microsoft.com/en-us/windows/ai/windows-ml/get-onnx-model
[65] ONNX Model Zoo https://onnx.ai/models/
[66] A User Survey and Failure Analysis of ONNX Model ... https://dl.acm.org/doi/10.1145/3650212.3680374
[67] ONNX Models https://onnxruntime.ai/models
'정보관리기술사' 카테고리의 다른 글
| CAM(Content Addressable Memory) (0) | 2025.11.23 |
|---|---|
| 스테이블 디퓨전 (0) | 2025.11.16 |
| PIM (Processing In Memory) (0) | 2025.10.26 |
| HBM (0) | 2025.10.12 |
| MMU (0) | 2025.10.05 |