freecad나 blender cad 확장으로는 이래저래 힘들고
python 통해서 제어하면 이상하게 나와서 제어가 유리한 무언가를 찾는 중
[링크 : https://github.com/fogleman/sdf]
| openSCAD + claude.ai = 파라메트릭 모델링 (0) | 2026.07.13 |
|---|---|
| openSCAD MCP (0) | 2026.06.20 |
| pythonSCAD (0) | 2026.06.20 |
| openSCAD cheatsheet (0) | 2026.06.20 |
| openSCAD 설치 (0) | 2026.06.20 |
-----
metric 함수. 얘도 함수였구나.
아무튼 metric은 격하게 오락가락 하고
[링크 : https://modulabs.co.kr/blog/loss-versus-accuracy]
loss 가 점진적 감소 -> 잘 학습
loss 급격히 감소 후 유지 -> 정상
loss 일정 시점 이후 증가 -> over fitting
[링크 : https://m.blog.naver.com/datapreprep/223748771141]
근데.. underfitting은 먼지 감이 안오네.
When to change the number of epochs (training cycles)Selecting the appropriate number of epochs is a balance between underfitting and overfitting.Underfitting: One of the most straightforward indicators of underfitting is if the model performs poorly on the training data. This can be observed in Edge Impulse Studio through metrics such as accuracy, or loss, depending on the type of problem (classification or regression). If these metrics indicate poor performance, it suggests that the model has not learned the patterns of the data well. In that case, increasing the number of epochs can improve your model performance. Please note that other solutions exist such as increasing your neural network architecture complexity, changing the preprocessing technique or reducing regularization. Overfitting: Detecting overfitting involves recognizing when the model has learned too much from the training data, including its noise and outliers, to the detriment of its performance on new, unseen data. Overfitting is characterized by the model performing exceptionally well on the training data but poorly on the validation or test data. Evaluating overfitting can be achieved by comparing the performance of the model between the training set and the validation set during training. When the performance on the validation set starts to degrade, it might indicate that the model is beginning to overfit the training data. In that case, decreasing the number of epochs can improve your model performance. As with underfitting, other solutions exist to reduce overfitting such as increasing the number of training data, adding regularization techniques to add penalties on large weights, adding dropout layers, simplifying the model architecture and even using early stopping.  |
[링크 : https://docs.edgeimpulse.com/knowledge/concepts/machine-learning/neural-networks/epochs]
overfitting 중인 loss 값 이라는데,
학습셋이 너무 적을 경우 overfitting이 빠르게 발생하는 것으로 보인다.
그나저나 저 130 miliion parameter는 ssd300 에서 어떻게 산출된걸까?
|
What you are experiencing is called overfitting and it happens because of your very small dataset.
All the model cares about is performance on the training dataset, so given the opportunity, it will simply attempt to memorize it. This is what happens in you case, you feed a model which contains over 130 Million parameters less than 319 images. So regarding your questions:
|
[링크 : https://datascience.stackexchange.com/questions/46456/issues-with-training-ssd-on-own-dataset]
| The SSD300 has 35.6 million parameters, while the SSDlite320 has about a tenth of them, i.e., 3.4 million. |
| ubuntu 26.04 + 3070 + tensorflow + python 3.14 + docker... (0) | 2026.06.17 |
|---|---|
| STFPM 실행 (0) | 2026.06.01 |
| NAS - Neural Architecture Search (0) | 2026.05.21 |
| MCUNet (0) | 2026.05.21 |
| STFPM - Student-Teacher Feature Pyramid Matching (0) | 2026.05.20 |
MCUNet 보다 보니 NAS 라는게 보여서 찾아봄
| Search NN model on an existing library e.g., ProxylessNAS, MnasNet |
[링크 : https://hanlab18.mit.edu/projects/tinyml/mcunet/assets/MCUNet-slides.pdf]
Neural Architecture Search
[링크 : https://doing-ai.tistory.com/entry/Neural-Architecture-Search-NAS-Hardware-Aware-NAS-ProxylessNAS]
| EfficientNAS 이후에 나온 더욱 효율적인 NAS를 위한 논문, ProxylessNAS다. TuNAS의 근간이 된 논문이기도 하며, 구체적으로 알아보자. |
[링크 : https://seokdonge.tistory.com/27] ProxylessNAS
[링크 : https://ech97.tistory.com/entry/MnasNet] MnasNet
TuNAS
[링크 : https://seokdonge.tistory.com/26]
[링크 : https://github.com/google-research/google-research/tree/master/tunas]
| STFPM 실행 (0) | 2026.06.01 |
|---|---|
| 딥러닝 학습 관련(epoch, loss) (0) | 2026.05.27 |
| MCUNet (0) | 2026.05.21 |
| STFPM - Student-Teacher Feature Pyramid Matching (0) | 2026.05.20 |
| EfficientAD (0) | 2026.05.20 |
STFPM -> PaSTe -> MCUNet 으로 찾아옴
[링크 : https://github.com/AMCO-UniPD/PaSTe]
MCU 에서 돌릴만큼 경량 네트워크
[링크 : https://cowkim-svd.tistory.com/7]
[링크 : https://github.com/mit-han-lab/mcunet]
| 딥러닝 학습 관련(epoch, loss) (0) | 2026.05.27 |
|---|---|
| NAS - Neural Architecture Search (0) | 2026.05.21 |
| STFPM - Student-Teacher Feature Pyramid Matching (0) | 2026.05.20 |
| EfficientAD (0) | 2026.05.20 |
| patchcore (0) | 2026.05.20 |
[링크 : https://github.com/gdwang08/STFPM]
[링크 : https://blog.naver.com/kingjykim/223495332696]
[링크 : https://claude.ai/share/6d307c41-77bc-411e-90a8-982fb5ce0467]
| STFPM | PaSTe |
 |
 |
Student-Teacher Feature Pyramid Matching
[링크 : https://www.alphaxiv.org/ko/overview/2503.02691v1]
Partially Shared Teacher-student
| Our results show that PaSTe decreases the inference time by 25%, while reducing the training time by 33% and peak RAM usage during training by 76%. |
| NAS - Neural Architecture Search (0) | 2026.05.21 |
|---|---|
| MCUNet (0) | 2026.05.21 |
| EfficientAD (0) | 2026.05.20 |
| patchcore (0) | 2026.05.20 |
| MVTec AD 데이터 셋 (0) | 2026.05.20 |
[링크 : https://github.com/nelson1425/EfficientAD]
[링크 : https://github.com/krec7748/EfficientAD]
[링크 : https://velog.io/@krec7748/논문-리뷰-및-구현-EfficientAD]
| MCUNet (0) | 2026.05.21 |
|---|---|
| STFPM - Student-Teacher Feature Pyramid Matching (0) | 2026.05.20 |
| patchcore (0) | 2026.05.20 |
| MVTec AD 데이터 셋 (0) | 2026.05.20 |
| VAD - PaDiM, Patchcore - 정상이 아님을 탐지 (0) | 2026.05.19 |
VAD 분야 SOTA(state of the art. 별걸 다 줄여 -_-) 모델이라는데 무거워서 모바일이나 엣지에서 돌리긴 좀 무리일 듯.
[링크 : https://coderecording.tistory.com/9]
[링크 : https://github.com/amazon-science/patchcore-inspection]
| STFPM - Student-Teacher Feature Pyramid Matching (0) | 2026.05.20 |
|---|---|
| EfficientAD (0) | 2026.05.20 |
| MVTec AD 데이터 셋 (0) | 2026.05.20 |
| VAD - PaDiM, Patchcore - 정상이 아님을 탐지 (0) | 2026.05.19 |
| U-net, segmentation (0) | 2026.05.14 |
가장 만만한(?) 데이터 셋인 것 같은데
| DAGM 데이터셋은 총 10가지 도메인의 데이터로 구성되어 있으며, 모델링을 통해 가상으로 결함을 합성하여 만든 데이터셋입니다. NanoTWICE 데이터셋은 nanofibrous material 데이터이며 5장의 정상 데이터와 40장의 결함 데이터로 구성이 되어있습니다 자, 이제 오늘의 본론인 MVTec-AD 데이터셋에 대해 설명드리겠습니다. 앞서 설명드렸던 DAGM, NanoTWICE의 아쉬웠던 부분들을 개선하며 총 15종류의 도메인의 데이터셋을 구축하였습니다. 크게는 Texture와 Object로 구분을 하였고, 각각 5가지, 10가지 종류의 도메인 데이터로 구성이 되어있습니다 |
[링크 : https://hoya012.github.io/blog/MVTec-AD/]
5.3 GB(!!)
[링크 : https://www.kaggle.com/datasets/ipythonx/mvtec-ad]
[링크 : https://www.mvtec.com/research-teaching/datasets/mvtec-ad]
[링크 : https://huggingface.co/datasets/Voxel51/mvtec-ad]
5.9 GB(!!!)
[링크 : https://www.kaggle.com/datasets/mhskjelvareid/dagm-2007-competition-dataset-optical-inspection]
[링크 : https://github.com/M-3LAB/awesome-industrial-anomaly-detection]
| EfficientAD (0) | 2026.05.20 |
|---|---|
| patchcore (0) | 2026.05.20 |
| VAD - PaDiM, Patchcore - 정상이 아님을 탐지 (0) | 2026.05.19 |
| U-net, segmentation (0) | 2026.05.14 |
| QAT, PTQ .. 모델 경량화 (0) | 2026.05.14 |
비정상을 학습시키는게 아니라
정상만 학습시키고 정상에서 벗어나는 것을 탐지하는 기법
PaDiM
Ganomaly (GAN + Anomaly Detection)
PaDiM-Lite / PatchCore-Lite
[링크 : https://www.alphaxiv.org/ko/overview/2603.20288v1]
PaDiM (Patch Ditrubution Modeling)
[링크 : https://huggingface.co/papers/2011.08785]
[링크 : https://data-analysis-expertise.tistory.com/128]
[링크 : https://dippingtodeepening.tistory.com/110]
[링크 : https://cumulu-s.tistory.com/44]
VAD - Visual Anomaly Detection
MVTec-AD, MVTec-3D, MVTec-LOCO, VisA 데이터 셋
PatchCore
[링크 : https://velog.io/@ljwljy51/PatchCore-Paper-Review]
PatchCore
SPADE , PaDiM
[링크 : https://ffighting.net/deep-learning-paper-review/anomaly-detection/patchcore/]
| patchcore (0) | 2026.05.20 |
|---|---|
| MVTec AD 데이터 셋 (0) | 2026.05.20 |
| U-net, segmentation (0) | 2026.05.14 |
| QAT, PTQ .. 모델 경량화 (0) | 2026.05.14 |
| unet (0) | 2026.05.06 |
귀찮으니(!) gpt를 통해서 딸깍~ 시전
| % 이미지 읽기 img = imread("test.jpg"); % RGB -> grayscale 직접 변환 if ndims(img) == 3 r = double(img(:,:,1)); g = double(img(:,:,2)); b = double(img(:,:,3)); gray = 0.299*r + 0.587*g + 0.114*b; else gray = double(img); end % 블럭 크기 blockSize = 32; [h, w] = size(gray); % 결과 저장 scoreMap = zeros(floor(h/blockSize), floor(w/blockSize)); % 블럭별 FFT 처리 for by = 1:blockSize:(h - blockSize + 1) for bx = 1:blockSize:(w - blockSize + 1) % 블럭 추출 block = gray(by:by+blockSize-1, bx:bx+blockSize-1); % 평균 제거 block = block - mean(block(:)); % FFT F = fft2(block); % magnitude mag = abs(F); % 중앙 이동 mag = fftshift(mag); center = floor(blockSize/2) + 1; radius = 4; [X, Y] = meshgrid(1:blockSize, 1:blockSize); dist = sqrt((X-center).^2 + (Y-center).^2); % 고주파만 사용 mask = dist > radius; % 고주파 에너지 energy = sum(mag(mask).^2); % diversity score score = log(1 + energy); iy = (by-1)/blockSize + 1; ix = (bx-1)/blockSize + 1; scoreMap(iy, ix) = score; end end % 정규화 scoreMap = scoreMap - min(scoreMap(:)); scoreMap = scoreMap / max(scoreMap(:)); % 결과 표시 figure; imshow(uint8(gray)); title("Original"); figure; imagesc(scoreMap); axis image; colorbar; title("FFT Diversity Map"); |
원본
그레이스케일
원인은 모르겠으나 blocksize 5 이하로는 검은색으로 나온다.
| blocksize = 6 | blocksize = 8 | blocksize = 10 |
 |
 |
 |
| blocksize = 12 | blocksize = 16 | blocksize = 32 |
 |
 |
 |
원본
그레이스케일
| blocksize = 6 | blocksize = 10 | blocksize = 12 |
 |
 |
 |
| blocksize = 16 | blocksize = 18 | blocksize = 20 |
 |
 |
 |
| blocksize = 24 | blocksize = 28 | blocksize = 32 |
 |
 |
 |
이미지 폭이 큰 차이가 없어서 그런진 모르겠지만..
blocksize 10~12 정도가 가장 무난한 것 같다.
rgb에 대해서 각각 처리 후 합산인데 눈에 보이는건 크게 달라 보이진 않았는데
| img = imread("test.jpg"); img = double(img); blockSize = 32; [h, w, c] = size(img); scoreMap = zeros(floor(h/blockSize), floor(w/blockSize)); for by = 1:blockSize:(h - blockSize + 1) for bx = 1:blockSize:(w - blockSize + 1) totalEnergy = 0; % RGB 각각 처리 for ch = 1:3 block = img(by:by+blockSize-1, ... bx:bx+blockSize-1, ch); block = block - mean(block(:)); F = fft2(block); mag = abs(fftshift(F)); center = floor(blockSize/2) + 1; radius = 4; [X, Y] = meshgrid(1:blockSize, 1:blockSize); dist = sqrt((X-center).^2 + (Y-center).^2); mask = dist > radius; energy = sum(mag(mask).^2); totalEnergy = totalEnergy + energy; end score = log(1 + totalEnergy); iy = (by-1)/blockSize + 1; ix = (bx-1)/blockSize + 1; scoreMap(iy, ix) = score; end end scoreMap = scoreMap - min(scoreMap(:)); scoreMap = scoreMap / max(scoreMap(:)); figure; imshow(uint8(img)); figure; imagesc(scoreMap); axis image; colorbar; title("RGB FFT Diversity"); |
blocksize = 10
두개를 비교해보면 하단에 파란색 마트 정도는 좀 더 잘 보이는 차이가 있는데
처리 시간대비로는 그리 큰 효과를 얻긴 힘들지도 모르겠다.
| grayscale | rgb 각각 처리후 합산 |
 |
|
LAB 으로 전환해서 색상에 대해서 처리, blocksize = 10
| % ========================================== % LAB 기반 Fabric Diversity / Anomaly Map % ========================================== img = imread("test.jpg"); img = double(img) / 255.0; %------------------------------------------ % RGB -> LAB 변환 %------------------------------------------ % sRGB gamma correction 제거 mask = img <= 0.04045; img_linear = zeros(size(img)); img_linear(mask) = img(mask) / 12.92; img_linear(~mask) = ((img(~mask)+0.055)/1.055).^2.4; % RGB -> XYZ M = [ ... 0.4124564 0.3575761 0.1804375; 0.2126729 0.7151522 0.0721750; 0.0193339 0.1191920 0.9503041]; [h, w, ~] = size(img); rgb_reshaped = reshape(img_linear, [], 3); xyz = rgb_reshaped * M'; X = xyz(:,1) / 0.95047; Y = xyz(:,2); Z = xyz(:,3) / 1.08883; % XYZ -> LAB helper f = @(t) ((t > 0.008856).*t.^(1/3) + ... (t <= 0.008856).*(7.787*t + 16/116)); fx = f(X); fy = f(Y); fz = f(Z); L = 116*fy - 16; A = 500*(fx - fy); B = 200*(fy - fz); lab = zeros(h, w, 3); lab(:,:,1) = reshape(L, h, w); lab(:,:,2) = reshape(A, h, w); lab(:,:,3) = reshape(B, h, w); %------------------------------------------ % FFT Block Analysis %------------------------------------------ blockSize = 32; step = 16; % overlap sliding outH = floor((h - blockSize)/step) + 1; outW = floor((w - blockSize)/step) + 1; scoreMap = zeros(outH, outW); % FFT mask 생성 center = floor(blockSize/2) + 1; [Xm, Ym] = meshgrid(1:blockSize, 1:blockSize); dist = sqrt((Xm-center).^2 + (Ym-center).^2); radius = 4; highMask = dist > radius; %------------------------------------------ % Sliding Window %------------------------------------------ oy = 1; for by = 1:step:(h - blockSize + 1) ox = 1; for bx = 1:step:(w - blockSize + 1) totalEnergy = 0; % LAB 각 채널 FFT for ch = 1:3 block = lab( ... by:by+blockSize-1, ... bx:bx+blockSize-1, ... ch); % 평균 제거 block = block - mean(block(:)); % FFT F = fft2(block); mag = abs(fftshift(F)); % 고주파 에너지 energy = sum(mag(highMask).^2); totalEnergy = totalEnergy + energy; end % 로그 스케일 score = log(1 + totalEnergy); scoreMap(oy, ox) = score; ox = ox + 1; end oy = oy + 1; end %------------------------------------------ % Normalize %------------------------------------------ scoreMap = scoreMap - min(scoreMap(:)); scoreMap = scoreMap / max(scoreMap(:)); %------------------------------------------ % Display %------------------------------------------ figure; imagesc(img); axis image; title("Original"); figure; imagesc(scoreMap); axis image; colorbar; title("LAB FFT Diversity / Anomaly Map"); |
block 10, step 16
block 10, step 10
| rgb 각각 처리후 합산 | RGB->LAB, 색상만 처리 |
 |
|
| octave - pkg load image (0) | 2026.05.19 |
|---|---|
| octave csv에서 읽어들이기 (0) | 2024.11.26 |
| octave audioread wav (0) | 2023.07.12 |
| 공짜 matlab? octave (0) | 2015.11.05 |
|
