{ "id": 203413, "title": "PyTorch Baseline", "url": "/competitions/ranzcr-clip-catheter-line-classification/discussion/203413", "author_name": "", "post_date": "2020-12-15T06:49:07.386407400Z", "votes": 40, "comment_count": 9, "views": 0, "content": "

I prepared PyTorch Baseline.

\n

Version1

\n\n

Version2

\n\n

Version3

\n\n

Hope this helps, happy kaggling!

", "messages": [ { "id": "1113069", "postDate": "12/15/2020 06:49:07", "content": "

I prepared PyTorch Baseline.

\n

Version1

\n\n

Version2

\n\n

Version3

\n\n

Hope this helps, happy kaggling!

", "rawMarkdown": "I prepared PyTorch Baseline.\n\nVersion1\n- Model: resnext50_32x4d\n- Split: GroupKFold 5 folds\n- Size: 320x320\n- CV: 0.959, LB: 0.971 (old metric)\n- CV: 0.893, LB: 0.923 (new metric)\n- training: https://www.kaggle.com/yasufuminakama/ranzcr-resnext50-32x4d-starter-training?scriptVersionId=49348116\n- inference: https://www.kaggle.com/yasufuminakama/ranzcr-resnext50-32x4d-starter-inference?scriptVersionId=49367987\n\nVersion2\n- Model: resnext50_32x4d\n- Split: GroupKFold 5 folds\n- Size: 448x448\n- CV: 0.9281, LB: 0.943\n- training: https://www.kaggle.com/yasufuminakama/ranzcr-resnext50-32x4d-starter-training?scriptVersionId=49697928\n- inference: https://www.kaggle.com/yasufuminakama/ranzcr-resnext50-32x4d-starter-inference?scriptVersionId=49707639\n\nVersion3\n- Model: resnext50_32x4d\n- Split: GroupKFold 4 folds\n- Size: 600x600\n- CV: 0.9337, LB: 0.948\n- training: https://www.kaggle.com/yasufuminakama/ranzcr-resnext50-32x4d-starter-training?scriptVersionId=49722999\n- inference: https://www.kaggle.com/yasufuminakama/ranzcr-resnext50-32x4d-starter-inference?scriptVersionId=49757551\n\nHope this helps, happy kaggling!", "votes": null }, { "id": "1113072", "postDate": "12/15/2020 06:51:49", "content": "

I'm not sure we can share resized dataset (refer to https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/203342)
\nHere I put the code of preparing 384x384 resized dataset used in this notebook.

\n
import os\nimport cv2\nimport zipfile\nimport numpy as np\nimport pandas as pd\nfrom tqdm.auto import tqdm\nfrom matplotlib import pyplot as plt\n\ntrain = pd.read_csv('../input/ranzcr-clip-catheter-line-classification/train.csv')\n\nTRAIN_PATH = '../input/ranzcr-clip-catheter-line-classification/train/'\n\nwith zipfile.ZipFile(f'train.zip', 'w') as img_out:\n    for uid in tqdm(train['StudyInstanceUID'].values):\n        image = cv2.imread(TRAIN_PATH + f'{uid}.jpg')\n        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n        image = cv2.resize(image, (384, 384))\n        image = cv2.imencode('.png', image)[1]\n        img_out.writestr(f\"{uid}.png\", image)\n
", "rawMarkdown": "I'm not sure we can share resized dataset (refer to https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/203342)\nHere I put the code of preparing 384x384 resized dataset used in this notebook.\n\n```\nimport os\nimport cv2\nimport zipfile\nimport numpy as np\nimport pandas as pd\nfrom tqdm.auto import tqdm\nfrom matplotlib import pyplot as plt\n\ntrain = pd.read_csv('../input/ranzcr-clip-catheter-line-classification/train.csv')\n\nTRAIN_PATH = '../input/ranzcr-clip-catheter-line-classification/train/'\n\nwith zipfile.ZipFile(f'train.zip', 'w') as img_out:\n for uid in tqdm(train['StudyInstanceUID'].values):\n image = cv2.imread(TRAIN_PATH + f'{uid}.jpg')\n image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n image = cv2.resize(image, (384, 384))\n image = cv2.imencode('.png', image)[1]\n img_out.writestr(f\"{uid}.png\", image)\n```", "votes": null }, { "id": "1113346", "postDate": "12/15/2020 11:37:36", "content": "

Wow this seems very high for a baseline model :( Not much room to improve on that I guess.

", "rawMarkdown": "Wow this seems very high for a baseline model :( Not much room to improve on that I guess.", "votes": null }, { "id": "1113365", "postDate": "12/15/2020 11:54:58", "content": "

As a baseline, I haven't used any augmentations & segmentation annotations during training.
\nAnd as discussed in this thread, AUC averaging column wise is better for competition metric…

", "rawMarkdown": "As a baseline, I haven't used any augmentations & segmentation annotations during training.\nAnd as discussed in [this thread](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/203375), AUC averaging column wise is better for competition metric...", "votes": null }, { "id": "1113372", "postDate": "12/15/2020 12:03:45", "content": "

very good work! kagglers are fast.

\n

i am stuck at other competitions … i will come back later.
\nAll the best to you!

", "rawMarkdown": "very good work! kagglers are fast.\n\ni am stuck at other competitions ... i will come back later.\nAll the best to you!", "votes": null }, { "id": "1113393", "postDate": "12/15/2020 12:14:15", "content": "

Thanks!
\nWe have many competitions these days…

", "rawMarkdown": "Thanks!\nWe have many competitions these days...", "votes": null }, { "id": "1113808", "postDate": "12/15/2020 17:48:41", "content": "

Maggie confirmed that it's fine to have preprocessing notebooks/datasets as long as it's clearly specified and attached to your model notebooks.

", "rawMarkdown": "Maggie confirmed that it's fine to have preprocessing notebooks/datasets as long as it's clearly specified and attached to your model notebooks.", "votes": null }, { "id": "1115762", "postDate": "12/16/2020 14:58:25", "content": "

Hi @yasufuminakama

\n

Your notebook has a score of 0.923. Another public notebook which is identical to yours (Kaggle shows no lines added or removed) has a score of 0.971

\n

Can you please help me understand why there should be a difference in scores with same code and same data? Am i missing something… Thanks for explaining!

", "rawMarkdown": "Hi @yasufuminakama \n\nYour notebook has a score of 0.923. Another public notebook which is identical to yours (Kaggle shows no lines added or removed) has a score of 0.971\n\nCan you please help me understand why there should be a difference in scores with same code and same data? Am i missing something... Thanks for explaining!", "votes": null }, { "id": "1115872", "postDate": "12/16/2020 16:42:53", "content": "

I'm not sure but it seems my notebook shows new metric score, forked notebook shows old metric score.

", "rawMarkdown": "I'm not sure but it seems my notebook shows new metric score, forked notebook shows old metric score.", "votes": null }, { "id": "1115900", "postDate": "12/16/2020 17:00:54", "content": "

Thanks @yasufuminakama

", "rawMarkdown": "Thanks @yasufuminakama", "votes": null } ], "comments": [ { "id": 1113072, "author_name": "yasufuminakama", "author_url": "", "post_date": "12/15/2020 06:51:49", "content": "

I'm not sure we can share resized dataset (refer to https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/203342)
\nHere I put the code of preparing 384x384 resized dataset used in this notebook.

\n
import os\nimport cv2\nimport zipfile\nimport numpy as np\nimport pandas as pd\nfrom tqdm.auto import tqdm\nfrom matplotlib import pyplot as plt\n\ntrain = pd.read_csv('../input/ranzcr-clip-catheter-line-classification/train.csv')\n\nTRAIN_PATH = '../input/ranzcr-clip-catheter-line-classification/train/'\n\nwith zipfile.ZipFile(f'train.zip', 'w') as img_out:\n    for uid in tqdm(train['StudyInstanceUID'].values):\n        image = cv2.imread(TRAIN_PATH + f'{uid}.jpg')\n        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n        image = cv2.resize(image, (384, 384))\n        image = cv2.imencode('.png', image)[1]\n        img_out.writestr(f\"{uid}.png\", image)\n
", "votes": null, "replies": [ { "id": 1113808, "author_name": "xhlulu", "author_url": "", "post_date": "12/15/2020 17:48:41", "content": "

Maggie confirmed that it's fine to have preprocessing notebooks/datasets as long as it's clearly specified and attached to your model notebooks.

", "votes": null, "replies": [] } ] }, { "id": 1113346, "author_name": "philippsinger", "author_url": "", "post_date": "12/15/2020 11:37:36", "content": "

Wow this seems very high for a baseline model :( Not much room to improve on that I guess.

", "votes": null, "replies": [ { "id": 1113365, "author_name": "yasufuminakama", "author_url": "", "post_date": "12/15/2020 11:54:58", "content": "

As a baseline, I haven't used any augmentations & segmentation annotations during training.
\nAnd as discussed in this thread, AUC averaging column wise is better for competition metric…

", "votes": null, "replies": [] } ] }, { "id": 1113372, "author_name": "hengck23", "author_url": "", "post_date": "12/15/2020 12:03:45", "content": "

very good work! kagglers are fast.

\n

i am stuck at other competitions … i will come back later.
\nAll the best to you!

", "votes": null, "replies": [ { "id": 1113393, "author_name": "yasufuminakama", "author_url": "", "post_date": "12/15/2020 12:14:15", "content": "

Thanks!
\nWe have many competitions these days…

", "votes": null, "replies": [] } ] }, { "id": 1115762, "author_name": "kmldas", "author_url": "", "post_date": "12/16/2020 14:58:25", "content": "

Hi @yasufuminakama

\n

Your notebook has a score of 0.923. Another public notebook which is identical to yours (Kaggle shows no lines added or removed) has a score of 0.971

\n

Can you please help me understand why there should be a difference in scores with same code and same data? Am i missing something… Thanks for explaining!

", "votes": null, "replies": [ { "id": 1115872, "author_name": "yasufuminakama", "author_url": "", "post_date": "12/16/2020 16:42:53", "content": "

I'm not sure but it seems my notebook shows new metric score, forked notebook shows old metric score.

", "votes": null, "replies": [] }, { "id": 1115900, "author_name": "kmldas", "author_url": "", "post_date": "12/16/2020 17:00:54", "content": "

Thanks @yasufuminakama

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1113069": "I prepared PyTorch Baseline.\n\nVersion1\n- Model: resnext50_32x4d\n- Split: GroupKFold 5 folds\n- Size: 320x320\n- CV: 0.959, LB: 0.971 (old metric)\n- CV: 0.893, LB: 0.923 (new metric)\n- training: https://www.kaggle.com/yasufuminakama/ranzcr-resnext50-32x4d-starter-training?scriptVersionId=49348116\n- inference: https://www.kaggle.com/yasufuminakama/ranzcr-resnext50-32x4d-starter-inference?scriptVersionId=49367987\n\nVersion2\n- Model: resnext50_32x4d\n- Split: GroupKFold 5 folds\n- Size: 448x448\n- CV: 0.9281, LB: 0.943\n- training: https://www.kaggle.com/yasufuminakama/ranzcr-resnext50-32x4d-starter-training?scriptVersionId=49697928\n- inference: https://www.kaggle.com/yasufuminakama/ranzcr-resnext50-32x4d-starter-inference?scriptVersionId=49707639\n\nVersion3\n- Model: resnext50_32x4d\n- Split: GroupKFold 4 folds\n- Size: 600x600\n- CV: 0.9337, LB: 0.948\n- training: https://www.kaggle.com/yasufuminakama/ranzcr-resnext50-32x4d-starter-training?scriptVersionId=49722999\n- inference: https://www.kaggle.com/yasufuminakama/ranzcr-resnext50-32x4d-starter-inference?scriptVersionId=49757551\n\nHope this helps, happy kaggling!", "1113072": "I'm not sure we can share resized dataset (refer to https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/203342)\nHere I put the code of preparing 384x384 resized dataset used in this notebook.\n\n```\nimport os\nimport cv2\nimport zipfile\nimport numpy as np\nimport pandas as pd\nfrom tqdm.auto import tqdm\nfrom matplotlib import pyplot as plt\n\ntrain = pd.read_csv('../input/ranzcr-clip-catheter-line-classification/train.csv')\n\nTRAIN_PATH = '../input/ranzcr-clip-catheter-line-classification/train/'\n\nwith zipfile.ZipFile(f'train.zip', 'w') as img_out:\n for uid in tqdm(train['StudyInstanceUID'].values):\n image = cv2.imread(TRAIN_PATH + f'{uid}.jpg')\n image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n image = cv2.resize(image, (384, 384))\n image = cv2.imencode('.png', image)[1]\n img_out.writestr(f\"{uid}.png\", image)\n```", "1113346": "Wow this seems very high for a baseline model :( Not much room to improve on that I guess.", "1113365": "As a baseline, I haven't used any augmentations & segmentation annotations during training.\nAnd as discussed in [this thread](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/203375), AUC averaging column wise is better for competition metric...", "1113372": "very good work! kagglers are fast.\n\ni am stuck at other competitions ... i will come back later.\nAll the best to you!", "1113393": "Thanks!\nWe have many competitions these days...", "1113808": "Maggie confirmed that it's fine to have preprocessing notebooks/datasets as long as it's clearly specified and attached to your model notebooks.", "1115762": "Hi @yasufuminakama \n\nYour notebook has a score of 0.923. Another public notebook which is identical to yours (Kaggle shows no lines added or removed) has a score of 0.971\n\nCan you please help me understand why there should be a difference in scores with same code and same data? Am i missing something... Thanks for explaining!", "1115872": "I'm not sure but it seems my notebook shows new metric score, forked notebook shows old metric score.", "1115900": "Thanks @yasufuminakama" }, "source": "meta" }