A.Resize(size, size),\n A.HorizontalFlip(=0.5),\n A.VerticalFlip(=0.5),\n A.RandomBrightnessContrast(=0.75),\n A.GaussNoise(var_limit=[10, 50]),\n A.GaussianBlur(),\n A.RandomGamma(=0.5, gamma_limit=(50, 200)),\n A.OneOf([A.OpticalDistortion(=0.3), A.GridDistortion(=0.3), A.ElasticTransform(=0.1)], =0.5),\n A.GridDistortion(=5, =0.3, =0.5),\n A.CoarseDropout(=1, =int(size * 0.3), =int(size * 0.3),=0, \n =0.5),\n A.Normalize(\n mean= [0] * in_chanss,\n std= [1] * in_chanss\n ),\n ToTensorV2(=)\n *Height = *\n =.*BCE + .*Dice + .*Tversky\n - seres_xd\n - rest\n - resd\n Networks Seresnet + resnet34D : \n - Zaxis pooling\n - Mean pooling \n - Dialated CNNs \n - Dynamic Pooling\n #Apply post-processing techniques\n kernel = np.ones((, ), np.uint8)\n min_area_threshold = \n kernel_size = \n sigmaX = \n threshold = TH # Adjust the threshold value needed\n mask_pred = cv2.morphology\n labels, num_labels = scipy.ndimage.label(mask_pred)\n label_idx range(, num_labels + ):\n region_area = np.sum(labelslabel_idx)\n region_area < min_area_threshold:\n mask_pred = \n mask_pred = cv2., (kernel_size, kernel_size), sigmaX)\n mask_pred = (mask_pred > threshold).astype()\n - . with TTA rotate \n- : 0.63 \n- Public: \n - : \n - Public : 0.76\n A.Resize(size, size),\n A.HorizontalFlip(=0.5),\n A.VerticalFlip(=0.5),\n A.RandomBrightnessContrast(=0.75),\n A.GaussNoise(var_limit=[10, 50]),\n A.GaussianBlur(),\n A.RandomGamma(=0.5, gamma_limit=(50, 200)),\n A.OneOf([A.OpticalDistortion(=0.3), A.GridDistortion(=0.3), A.ElasticTransform(=0.1)], =0.5),\n A.GridDistortion(=5, =0.3, =0.5),\n A.CoarseDropout(=1, =int(size * 0.3), =int(size * 0.3),=0, \n =0.5),\n A.Normalize(\n mean= [0] * in_chanss,\n std= [1] * in_chanss\n ),\n ToTensorV2(=)\n *Height = *\n =.*BCE + .*Dice + .*Tversky\n - seres_xd\n - rest\n - resd\n Networks Seresnet + resnet34D : \n - Zaxis pooling\n - Mean pooling \n - Dialated CNNs \n - Dynamic Pooling\n #Apply post-processing techniques\n kernel = np.ones((, ), np.uint8)\n min_area_threshold = \n kernel_size = \n sigmaX = \n threshold = TH # Adjust the threshold value needed\n mask_pred = cv2.morphology\n labels, num_labels = scipy.ndimage.label(mask_pred)\n label_idx range(, num_labels + ):\n region_area = np.sum(labelslabel_idx)\n region_area < min_area_threshold:\n mask_pred = \n mask_pred = cv2., (kernel_size, kernel_size), sigmaX)\n mask_pred = (mask_pred > threshold).astype()\n - . with TTA rotate \n- : 0.63 \n- Public: \n - : \n - Public : 0.76\ncan you tell which model is the best single model?
", "rawMarkdown": "can you tell which model is the best single model?", "votes": null }, { "id": "2302947", "postDate": "06/15/2023 01:38:45", "content": "Seresnet stacked with dynamic pooling
", "rawMarkdown": "Seresnet stacked with dynamic pooling", "votes": null }, { "id": "2302959", "postDate": "06/15/2023 01:49:21", "content": "Congratulations.
", "rawMarkdown": "Congratulations.", "votes": null }, { "id": "2303630", "postDate": "06/15/2023 11:45:47", "content": "Do you share some code about these：
\nYes sure in a couple of days
", "rawMarkdown": "Yes sure in a couple of days", "votes": null }, { "id": "2303725", "postDate": "06/15/2023 12:39:12", "content": "Congratulations！Thank you！
", "rawMarkdown": "Congratulations！Thank you！", "votes": null }, { "id": "2304093", "postDate": "06/15/2023 17:11:23", "content": "Notebooks : https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/417432
", "rawMarkdown": "Notebooks : https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/417432", "votes": null } ], "comments": [ { "id": 2302943, "author_name": "cutebomb", "author_url": "", "post_date": "06/15/2023 01:34:53", "content": "can you tell which model is the best single model?
", "votes": null, "replies": [ { "id": 2302947, "author_name": "arunodhayan", "author_url": "", "post_date": "06/15/2023 01:38:45", "content": "Seresnet stacked with dynamic pooling
", "votes": null, "replies": [ { "id": 2303630, "author_name": "wangxuc", "author_url": "", "post_date": "06/15/2023 11:45:47", "content": "Do you share some code about these：
\nYes sure in a couple of days
", "votes": null, "replies": [ { "id": 2303725, "author_name": "wangxuc", "author_url": "", "post_date": "06/15/2023 12:39:12", "content": "Congratulations！Thank you！
", "votes": null, "replies": [] } ] } ] } ] } ] }, { "id": 2302959, "author_name": "robsonsan", "author_url": "", "post_date": "06/15/2023 01:49:21", "content": "Congratulations.
", "votes": null, "replies": [] }, { "id": 2304093, "author_name": "arunodhayan", "author_url": "", "post_date": "06/15/2023 17:11:23", "content": "Notebooks : https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/417432
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2302899": "#Data Augmentation\n A.Resize(size, size),\n A.HorizontalFlip(p=0.5),\n A.VerticalFlip(p=0.5),\n A.RandomBrightnessContrast(p=0.75),\n A.GaussNoise(var_limit=[10, 50]),\n A.GaussianBlur(),\n A.RandomGamma(p=0.5, gamma_limit=(50, 200)),\n A.OneOf([A.OpticalDistortion(p=0.3), A.GridDistortion(p=0.3), A.ElasticTransform(p=0.1)], p=0.5),\n A.GridDistortion(num_steps=5, distort_limit=0.3, p=0.5),\n A.CoarseDropout(max_holes=1, max_width=int(size * 0.3), max_height=int(size * 0.3),mask_fill_value=0, \n p=0.5),\n A.Normalize(\n mean= [0] * in_chanss,\n std= [1] * in_chanss\n ),\n ToTensorV2(transpose_mask=True)\n\n#Resolution\n Width*Height = 320*320\n#Loss\n loss =0.3*BCE + 0.4*Dice + 0.3*Tversky\n#Model Architectures \n - seresnext50_32x4d\n - resnet10t\n - resnet34d\n#Methods worked for me:\n Stacked Networks Seresnet + resnet34D : \n - Zaxis pooling\n - Mean pooling \n - Dialated CNNs \n - Dynamic Pooling\n\n#Post processing \n #Apply post-processing techniques\n kernel = np.ones((3, 3), np.uint8)\n min_area_threshold = 100\n kernel_size = 5\n sigmaX = 1.0\n threshold = TH # Adjust the threshold value as needed\n mask_pred = cv2.morphologyEx(mask_pred, cv2.MORPH_CLOSE, kernel)\n labels, num_labels = scipy.ndimage.label(mask_pred)\n for label_idx in range(1, num_labels + 1):\n region_area = np.sum(labels == label_idx)\n if region_area < min_area_threshold:\n mask_pred[labels == label_idx] = 0\n mask_pred = cv2.GaussianBlur(mask_pred.astype(np.float32), (kernel_size, kernel_size), sigmaX)\n mask_pred = (mask_pred > threshold).astype(int)\n \n#Threshold\n Thr - 0.6 with TTA rotate \n\n##Best Score single model \n - Private : 0.63 \n - Public: 0.74\n\n## 6 Ensemble model score\n - Private: 0.62\n - Public : 0.76", "2302943": "can you tell which model is the best single model?", "2302947": "Seresnet stacked with dynamic pooling", "2302959": "Congratulations.", "2303630": "Do you share some code about these：\n- Zaxis pooling\n- Mean pooling\n- Dialated CNNs \n- Dynamic Pooling", "2303685": "Yes sure in a couple of days", "2303725": "Congratulations！Thank you！", "2304093": "Notebooks : https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/417432" }, "source": "meta" }