![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3830852%2F02777b71ba9b4886858ab554550f9a72%2Fkmeans.png?generation=1677580535294417&alt=media\")
Thank you to everyone involved in this competition.
\nThis is my first image competition, so I learned a lots of things.
\n(To be honest, I am so disappointed to have dropped so far from the Gold Zone. However, this is what I am capable of now.)
I want to write my brief summary of my solution.
\nI referred this windowing code. in this comp. I learned lots from @hengck23 . I really appreciate it.
\nSo windowing is just applied based on VOILUTFunction.
\nThen I put out images, which has 2048xXXXX size
I apply kmeans-clustering into each image using pixel values, then the pixel values of a cluster which has the lowest summation change into zero.
\nTo save time, I used the kmeans-pytorch
\nHere is my code.
This method gave me almost same CV score, but it boosted my LB (~0.58 --> ~0.63).
\nI cropped images using below code.
\nThis approach gave me slightly better CV and LB score than cropping using YOLO v5 (ref @remekkinas 's code, thank you very much).
\nthen it put out images (resize as 1536x960)
\n`
frame_org = copy.copy(frame)\n thres1 = np.min(frame)+68 #Adjustments were made while viewing the crop image.\n np.place(frame, frame < thres1, 0)\n thres2 = frame_org.sum() / (h*w)\n\n vertical_not_zero = [True if frame[:,idx].sum() > thres2 else False for idx in range(w)]\n horizontal_not_zero = [True if frame[idx,:].sum() > thres2 else False for idx in range(h)]\n\n crop = frame_org[horizontal_not_zero,:]\n crop = crop[:,vertical_not_zero]\n`
\nI used ExhaustiveWeightedRandomSampler (weight: Pos:Neg=7:1)
\nthis sampler also worked well for me.
I used the weighted binary cross entropy (weight: Pos:Neg = 2:1)
\nI used below code using Kornia. ( it also refered @hengck23 's code. Thanks)
\n`
self.flip = nn.Sequential(\n K.RandomHorizontalFlip(p=0.5),\n K.RandomVerticalFlip(p=0.5),\n )\n\n p=0.8\n self.transform_geometry = ImageSequential(\n K.RandomAffine(degrees=20, translate=0.1, scale=[0.8,1.2], shear=20, p=p),\n K.RandomThinPlateSpline(scale=0.25, p=p),\n random_apply=1, #choose 1\n )\n\n p=0.5\n self.transform_intensity = ImageSequential(\n K.RandomGamma(gamma=(0.5, 1.5), gain=(0.5, 1.2), p=p),\n K.RandomContrast(contrast=(0.8,1.2), p=p),\n K.RandomBrightness(brightness=(0.8,1.2), p=p),\n random_apply=1, #choose 1\n )\n`
\nThank you.
", "messages": [ { "id": 2162679, "postDate": "2023-02-28T11:10:17.667Z", "content": "Thank you to everyone involved in this competition.
\nThis is my first image competition, so I learned a lots of things.
\n(To be honest, I am so disappointed to have dropped so far from the Gold Zone. However, this is what I am capable of now.)
I want to write my brief summary of my solution.
\nI referred this windowing code. in this comp. I learned lots from @hengck23 . I really appreciate it.
\nSo windowing is just applied based on VOILUTFunction.
\nThen I put out images, which has 2048xXXXX size
I apply kmeans-clustering into each image using pixel values, then the pixel values of a cluster which has the lowest summation change into zero.
\nTo save time, I used the kmeans-pytorch
\nHere is my code.
This method gave me almost same CV score, but it boosted my LB (~0.58 --> ~0.63).
\nI cropped images using below code.
\nThis approach gave me slightly better CV and LB score than cropping using YOLO v5 (ref @remekkinas 's code, thank you very much).
\nthen it put out images (resize as 1536x960)
\n`
frame_org = copy.copy(frame)\n thres1 = np.min(frame)+68 #Adjustments were made while viewing the crop image.\n np.place(frame, frame < thres1, 0)\n thres2 = frame_org.sum() / (h*w)\n\n vertical_not_zero = [True if frame[:,idx].sum() > thres2 else False for idx in range(w)]\n horizontal_not_zero = [True if frame[idx,:].sum() > thres2 else False for idx in range(h)]\n\n crop = frame_org[horizontal_not_zero,:]\n crop = crop[:,vertical_not_zero]\n`
\nI used ExhaustiveWeightedRandomSampler (weight: Pos:Neg=7:1)
\nthis sampler also worked well for me.
I used the weighted binary cross entropy (weight: Pos:Neg = 2:1)
\nI used below code using Kornia. ( it also refered @hengck23 's code. Thanks)
\n`
self.flip = nn.Sequential(\n K.RandomHorizontalFlip(p=0.5),\n K.RandomVerticalFlip(p=0.5),\n )\n\n p=0.8\n self.transform_geometry = ImageSequential(\n K.RandomAffine(degrees=20, translate=0.1, scale=[0.8,1.2], shear=20, p=p),\n K.RandomThinPlateSpline(scale=0.25, p=p),\n random_apply=1, #choose 1\n )\n\n p=0.5\n self.transform_intensity = ImageSequential(\n K.RandomGamma(gamma=(0.5, 1.5), gain=(0.5, 1.2), p=p),\n K.RandomContrast(contrast=(0.8,1.2), p=p),\n K.RandomBrightness(brightness=(0.8,1.2), p=p),\n random_apply=1, #choose 1\n )\n`
\nThank you.
", "rawMarkdown": "Thank you to everyone involved in this competition.\nThis is my first image competition, so I learned a lots of things.\n(To be honest, I am so disappointed to have dropped so far from the Gold Zone. However, this is what I am capable of now.)\n\nI want to write my brief summary of my solution.\n\n\n# Overview\n1. Crop a breast area.\n2. Background noise reduction using k-means.\n3. Predict cancer for each image using efficientnet_v2\n4. Aggregate (mean aggregation)\n\n# Preprocess\n## Windowing\nI referred [this windowing code](https://www.kaggle.com/code/hengck23/3hr-tensorrt-nextvit-example). in this comp. I learned lots from @hengck23 . I really appreciate it.\nSo windowing is just applied based on VOILUTFunction.\nThen I put out images, which has 2048xXXXX size\n\n## Background noise reduction (kmeans)\nI apply kmeans-clustering into each image using pixel values, then the pixel values of a cluster which has the lowest summation change into zero.\nTo save time, I used the [kmeans-pytorch](https://github.com/subhadarship/kmeans_pytorch)\nHere is [my code](https://www.kaggle.com/code/taruto1215/66th-lb-10th-simple-image-crop-kmeans/notebook).\n\n\n\nThis method gave me almost same CV score, but it boosted my LB (~0.58 --> ~0.63).\n## Image crop\n I cropped images using below code.\nThis approach gave me slightly better CV and LB score than cropping using YOLO v5 (ref @remekkinas 's code, thank you very much).\nthen it put out images (resize as 1536x960)\n`\n\n frame_org = copy.copy(frame)\n thres1 = np.min(frame)+68 #Adjustments were made while viewing the crop image.\n np.place(frame, frame < thres1, 0)\n thres2 = frame_org.sum() / (h*w)\n \n vertical_not_zero = [True if frame[:,idx].sum() > thres2 else False for idx in range(w)]\n horizontal_not_zero = [True if frame[idx,:].sum() > thres2 else False for idx in range(h)]\n \n crop = frame_org[horizontal_not_zero,:]\n crop = crop[:,vertical_not_zero]\n`\n\n# Model\n- I used EffecientNet_V2_s (Final sub: 3 models ensemble which has different hyper-parameters, etc.) \n- I train the model for cancer and aux targets. loss weight is cancer:aux = 1:2 the multi-task learning gave me better CV and LB result.\n- Dropout_rate = 0.30, drop_path_rate=0.20\n- Multi-sampled dropout (drop_rate=0.2, num_drop=5). it gave me better CV and LB score.\n\n# Training\n- Optimzer: AdamW\n- Scheduler: CosineAnnealing (warmup=0.1)\n- epoch = 5\n- batch_size=36\n\n## Upsampling\nI used [ExhaustiveWeightedRandomSampler](https://github.com/louis-she/exhaustive-weighted-random-sampler) (weight: Pos:Neg=7:1)\nthis sampler also worked well for me.\n\n## Loss function \nI used the weighted binary cross entropy (weight: Pos:Neg = 2:1)\n\n## Data-augmentaion:\nI used below code using Kornia. ( it also refered @hengck23 's code. Thanks)\n`\n\n self.flip = nn.Sequential(\n K.RandomHorizontalFlip(p=0.5),\n K.RandomVerticalFlip(p=0.5),\n )\n\n p=0.8\n self.transform_geometry = ImageSequential(\n K.RandomAffine(degrees=20, translate=0.1, scale=[0.8,1.2], shear=20, p=p),\n K.RandomThinPlateSpline(scale=0.25, p=p),\n random_apply=1, #choose 1\n )\n\n p=0.5\n self.transform_intensity = ImageSequential(\n K.RandomGamma(gamma=(0.5, 1.5), gain=(0.5, 1.2), p=p),\n K.RandomContrast(contrast=(0.8,1.2), p=p),\n K.RandomBrightness(brightness=(0.8,1.2), p=p),\n random_apply=1, #choose 1\n )\n`\n\n## Post-processing.\n\n - I tried some aggregration method, but just mean aggregation is the best for my best submission.\n\nThank you.", "votes": 18 }, { "id": 2162717, "postDate": "2023-02-28T11:31:23.877Z", "content": "Good effort! There is no reason to feel disappointed. K-means clustering for background noice reduction is interesting. I have seen the methodology being used also for segmentation in some IEEE paper ages ago.
", "rawMarkdown": "Good effort! There is no reason to feel disappointed. K-means clustering for background noice reduction is interesting. I have seen the methodology being used also for segmentation in some IEEE paper ages ago.", "votes": 2, "replies": [ { "id": 2162732, "postDate": "2023-02-28T11:42:00.457Z", "content": "Thanks for the encouragement!
", "rawMarkdown": "Thanks for the encouragement!", "votes": 1 } ] }, { "id": 2165356, "postDate": "2023-03-02T06:22:50.960Z", "content": "Thank you for sharing your solution.
\nI didn't come up with kmeans noise reduction. How background noise reduction affected PB scores?
@ludditep
\nSorry for my late response.
\nThere were a little improve (less than 0.01) in the PB.
Good effort! There is no reason to feel disappointed. K-means clustering for background noice reduction is interesting. I have seen the methodology being used also for segmentation in some IEEE paper ages ago.
", "votes": 2, "replies": [ { "id": 2162732, "author_name": "taruto", "author_url": "", "post_date": "2023-02-28T11:42:00.457000", "content": "Thanks for the encouragement!
", "votes": 1, "replies": [] } ] }, { "id": 2165356, "author_name": "luddite^", "author_url": "", "post_date": "2023-03-02T06:22:50.960000", "content": "Thank you for sharing your solution.
\nI didn't come up with kmeans noise reduction. How background noise reduction affected PB scores?
@ludditep
\nSorry for my late response.
\nThere were a little improve (less than 0.01) in the PB.