Github repository\nFastai forum post
\n\nHi Everyone!
\n\nPlease allow me the pleasure of sharing with you a starter pack using FastAi
\n\nThe starter pack begins with downloading and processing the data. I extract images from the DICOM files and resize them to 112x112. I then randomly split the data assigning 20% of examples to the validation set. I use pretrained resnet-18 and a custom classification head that comes with fastai out of the box with default values.
\n\nThe training is done with progressive unfreezing and the last phase of training with a completely unfrozen model leverages discriminative learning rates.
\n\nThe model is trained for a total of 3 epochs without data augmentation. On a 1080ti training completes under 15 minutes.
\n\nThe techniques I use are covered in the ULMFIT paper. From what I remember I believe they are also introduced in the first three lectures of the FastAi Deep Learning for Coders v2 course (it's free).
\n\nBest of luck in the competition and have fun!
\n\nEDIT: Please note, window_and_normalize should divide window_width by 2 in order to align with the 'brain window' traditionally used by radiologists for visualization. As is right now, the window is twice as wide. Good discussion on relevance of windowing to visualization / modelling in this thread
Github repository\nFastai forum post
\n\nHi Everyone!
\n\nPlease allow me the pleasure of sharing with you a starter pack using FastAi
\n\nThe starter pack begins with downloading and processing the data. I extract images from the DICOM files and resize them to 112x112. I then randomly split the data assigning 20% of examples to the validation set. I use pretrained resnet-18 and a custom classification head that comes with fastai out of the box with default values.
\n\nThe training is done with progressive unfreezing and the last phase of training with a completely unfrozen model leverages discriminative learning rates.
\n\nThe model is trained for a total of 3 epochs without data augmentation. On a 1080ti training completes under 15 minutes.
\n\nThe techniques I use are covered in the ULMFIT paper. From what I remember I believe they are also introduced in the first three lectures of the FastAi Deep Learning for Coders v2 course (it's free).
\n\nBest of luck in the competition and have fun!
\n\nEDIT: Please note, window_and_normalize should divide window_width by 2 in order to align with the 'brain window' traditionally used by radiologists for visualization. As is right now, the window is twice as wide. Good discussion on relevance of windowing to visualization / modelling in this thread
thanks for the starter pack. learned a lot
", "rawMarkdown": "thanks for the starter pack. learned a lot", "votes": 1 }, { "id": 651858, "postDate": "2019-10-18T02:52:31.007Z", "content": "Thanks for an amazing starter kit using fastai!
", "rawMarkdown": "Thanks for an amazing starter kit using fastai!", "votes": 1 }, { "id": 642513, "postDate": "2019-10-06T07:52:38.200Z", "content": "Thank you very much for contributing so much to the community, Radek!\nYour notebooks helped me to get started faster with the data preprocessing & Co.\nKeep up the great work! :-D
", "rawMarkdown": "Thank you very much for contributing so much to the community, Radek!\nYour notebooks helped me to get started faster with the data preprocessing & Co.\nKeep up the great work! :-D", "votes": 1 }, { "id": 634830, "postDate": "2019-09-26T20:26:37.850Z", "content": "Good to see you in kaggle Radek, thanks for the great insight. I hope I can learn lots from you again just as following you in the whale competition :)
", "rawMarkdown": "Good to see you in kaggle Radek, thanks for the great insight. I hope I can learn lots from you again just as following you in the whale competition :) ", "votes": 1 }, { "id": 634065, "postDate": "2019-09-25T19:38:47.293Z", "content": "Great will take a look at that
", "rawMarkdown": "Great will take a look at that", "votes": 1 }, { "id": 632255, "postDate": "2019-09-23T11:49:26.267Z", "content": "Awesome!!
", "rawMarkdown": "Awesome!!", "votes": 1 }, { "id": 631394, "postDate": "2019-09-22T00:04:50.693Z", "content": "This is a great help, thanks for sharing your skills!
", "rawMarkdown": "This is a great help, thanks for sharing your skills!", "votes": 1 }, { "id": 631229, "postDate": "2019-09-21T16:22:22.553Z", "content": "Much appreciated!
", "rawMarkdown": "Much appreciated!", "votes": 1 }, { "id": 631089, "postDate": "2019-09-21T10:27:44.277Z", "content": "Thanks @radek1 for the amazing starter pack.
", "rawMarkdown": "Thanks @radek1 for the amazing starter pack.", "votes": 1 }, { "id": 630959, "postDate": "2019-09-21T04:42:09.233Z", "content": "Wow.... FastAI...!!
\n\nThanks for the Starter Kit...!! @radek1
", "rawMarkdown": "Wow.... FastAI...!!\n\nThanks for the Starter Kit...!! @radek1 ", "votes": 1 }, { "id": 630879, "postDate": "2019-09-20T22:55:45.830Z", "content": "Thanks for an amazing starter kit using fastai!
", "rawMarkdown": "Thanks for an amazing starter kit using fastai!", "votes": 1 }, { "id": 631076, "postDate": "2019-09-21T09:51:05.160Z", "content": "Thanks for creating and sharing this @radek1!\nDoes this mean, we’ll be seeing you on Kaggle LB(s) again? 😄
", "rawMarkdown": "Thanks for creating and sharing this @radek1!\nDoes this mean, we’ll be seeing you on Kaggle LB(s) again? 😄 ", "votes": 2, "replies": [ { "id": 631095, "postDate": "2019-09-21T10:48:13.333Z", "content": "I'm not sure 🙂 This competition seems like it might be a lot of fun plus I feel I can create something useful to others.
\n\nOn the other hand, these things take a lot of time which is hard to justify in the context of kids / work / life in general 🙂 Let's see 🙂
", "rawMarkdown": "I'm not sure 🙂 This competition seems like it might be a lot of fun plus I feel I can create something useful to others.\n\nOn the other hand, these things take a lot of time which is hard to justify in the context of kids / work / life in general 🙂 Let's see 🙂", "votes": 3 } ] }, { "id": 1019683, "postDate": "2020-09-20T15:41:27.310Z", "content": "Dear Sir I have some questions:
\nIn your '08_CAM_binary_classifier', may I know where does 'train_labels_as_strings_with_label_class.csv' come from? I do not see its definition in this or others files. Is it the same as 'train_labels_as_strings.csv'?
I have installed the fastai_dev via 'python setup.py install' from https://github.com/fastai/fastai_dev
\nand then set: path_to_fastai_dev_local = 'C:/Users/Administrator/Desktop/fastai_dev-master/dev'
\n sys.path.append(path_to_fastai_dev_local)
Then I have error:
\nIn [19]: from local.torch_basics import *
\nTraceback (most recent call last):
File \"\", line 1, in
\n from local.torch_basics import *
ModuleNotFoundError: No module named 'local'
\nCould you give some suggestions?
", "rawMarkdown": "Dear Sir I have some questions:\n\n1. In your '08_CAM_binary_classifier', may I know where does 'train_labels_as_strings_with_label_class.csv' come from? I do not see its definition in this or others files. Is it the same as 'train_labels_as_strings.csv'?\n\n2. I have installed the fastai_dev via 'python setup.py install' from https://github.com/fastai/fastai_dev\nand then set: path_to_fastai_dev_local = 'C:/Users/Administrator/Desktop/fastai_dev-master/dev'\n sys.path.append(path_to_fastai_dev_local)\n\nThen I have error: \n\nIn [19]: from local.torch_basics import *\nTraceback (most recent call last):\n\n File \"Hi Radek, quick question. how long does it take to train this model?
", "rawMarkdown": "Hi Radek, quick question. how long does it take to train this model?" }, { "id": 641862, "postDate": "2019-10-05T08:47:40.500Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 641934, "postDate": "2019-10-05T10:50:19.833Z", "content": "I have not saved the weights.
", "rawMarkdown": "I have not saved the weights." }, { "id": 641935, "postDate": "2019-10-05T10:52:08.040Z", "content": "The kernel I created has saved models from various phases of the training: https://www.kaggle.com/radek1/fastai-starter-pack-train-basic-model-and-submit/output
\n\nIt uses different data though.
", "rawMarkdown": "The kernel I created has saved models from various phases of the training: https://www.kaggle.com/radek1/fastai-starter-pack-train-basic-model-and-submit/output\n\nIt uses different data though." }, { "id": 642056, "postDate": "2019-10-05T13:58:43.817Z", "rawMarkdown": "", "votes": 1, "isDeleted": true } ] }, { "id": 631686, "postDate": "2019-09-22T13:32:09.143Z", "rawMarkdown": "", "votes": 1, "isDeleted": true }, { "id": 637096, "postDate": "2019-09-30T16:51:57.413Z", "content": "Hi Radek. Thank you
", "rawMarkdown": "Hi Radek. Thank you", "votes": 1 }, { "id": 633226, "postDate": "2019-09-24T15:19:16.457Z", "content": "thanks alot!
", "rawMarkdown": "thanks alot!", "votes": 1 }, { "id": 632823, "postDate": "2019-09-24T05:30:37.440Z", "content": "thank your for sharing
", "rawMarkdown": "thank your for sharing", "votes": 1 }, { "id": 631701, "postDate": "2019-09-22T14:14:16.623Z", "content": "Thanks!
", "rawMarkdown": "Thanks!", "votes": 1 }, { "id": 631384, "postDate": "2019-09-21T23:29:24.457Z", "content": "So awesome Radek! Thank you!
", "rawMarkdown": "So awesome Radek! Thank you!", "votes": 1 } ], "comments": [ { "id": 659619, "author_name": "Niranjan_NKC", "author_url": "", "post_date": "2019-10-28T01:41:31.583000", "content": "thanks for the starter pack. learned a lot
", "votes": 1, "replies": [] }, { "id": 651858, "author_name": "LongYin/杰少", "author_url": "", "post_date": "2019-10-18T02:52:31.007000", "content": "Thanks for an amazing starter kit using fastai!
", "votes": 1, "replies": [] }, { "id": 642513, "author_name": "Michael Pieler", "author_url": "", "post_date": "2019-10-06T07:52:38.200000", "content": "Thank you very much for contributing so much to the community, Radek!\nYour notebooks helped me to get started faster with the data preprocessing & Co.\nKeep up the great work! :-D
", "votes": 1, "replies": [] }, { "id": 634830, "author_name": "Hao He", "author_url": "", "post_date": "2019-09-26T20:26:37.850000", "content": "Good to see you in kaggle Radek, thanks for the great insight. I hope I can learn lots from you again just as following you in the whale competition :)
", "votes": 1, "replies": [] }, { "id": 634065, "author_name": "Tinashe Kadiki", "author_url": "", "post_date": "2019-09-25T19:38:47.293000", "content": "Great will take a look at that
", "votes": 1, "replies": [] }, { "id": 632255, "author_name": "Rohit Gupta", "author_url": "", "post_date": "2019-09-23T11:49:26.267000", "content": "Awesome!!
", "votes": 1, "replies": [] }, { "id": 631394, "author_name": "Melissa Rajaram", "author_url": "", "post_date": "2019-09-22T00:04:50.693000", "content": "This is a great help, thanks for sharing your skills!
", "votes": 1, "replies": [] }, { "id": 631229, "author_name": "Tom H.", "author_url": "", "post_date": "2019-09-21T16:22:22.553000", "content": "Much appreciated!
", "votes": 1, "replies": [] }, { "id": 631089, "author_name": "Venkat Sai", "author_url": "", "post_date": "2019-09-21T10:27:44.277000", "content": "Thanks @radek1 for the amazing starter pack.
", "votes": 1, "replies": [] }, { "id": 630959, "author_name": "Ailurophile", "author_url": "", "post_date": "2019-09-21T04:42:09.233000", "content": "Wow.... FastAI...!!
\n\nThanks for the Starter Kit...!! @radek1
", "votes": 1, "replies": [] }, { "id": 630879, "author_name": "ilovescience", "author_url": "", "post_date": "2019-09-20T22:55:45.830000", "content": "Thanks for an amazing starter kit using fastai!
", "votes": 1, "replies": [] }, { "id": 631076, "author_name": "Sanyam Bhutani", "author_url": "", "post_date": "2019-09-21T09:51:05.160000", "content": "Thanks for creating and sharing this @radek1!\nDoes this mean, we’ll be seeing you on Kaggle LB(s) again? 😄
", "votes": 2, "replies": [ { "id": 631095, "author_name": "Radek Osmulski", "author_url": "", "post_date": "2019-09-21T10:48:13.333000", "content": "I'm not sure 🙂 This competition seems like it might be a lot of fun plus I feel I can create something useful to others.
\n\nOn the other hand, these things take a lot of time which is hard to justify in the context of kids / work / life in general 🙂 Let's see 🙂
", "votes": 3, "replies": [] } ] }, { "id": 1019683, "author_name": "LIU Jundong", "author_url": "", "post_date": "2020-09-20T15:41:27.310000", "content": "Dear Sir I have some questions:
\nIn your '08_CAM_binary_classifier', may I know where does 'train_labels_as_strings_with_label_class.csv' come from? I do not see its definition in this or others files. Is it the same as 'train_labels_as_strings.csv'?
I have installed the fastai_dev via 'python setup.py install' from https://github.com/fastai/fastai_dev
\nand then set: path_to_fastai_dev_local = 'C:/Users/Administrator/Desktop/fastai_dev-master/dev'
\n sys.path.append(path_to_fastai_dev_local)
Then I have error:
\nIn [19]: from local.torch_basics import *
\nTraceback (most recent call last):
File \"\", line 1, in
\n from local.torch_basics import *
ModuleNotFoundError: No module named 'local'
\nCould you give some suggestions?
", "votes": 0, "replies": [] }, { "id": 652799, "author_name": "danny iskandar", "author_url": "", "post_date": "2019-10-19T12:37:16.497000", "content": "Hi Radek, quick question. how long does it take to train this model?
", "votes": 0, "replies": [] }, { "id": 641862, "author_name": "", "author_url": "", "post_date": "2019-10-05T08:47:40.500000", "content": "", "votes": 0, "replies": [ { "id": 641934, "author_name": "Radek Osmulski", "author_url": "", "post_date": "2019-10-05T10:50:19.833000", "content": "I have not saved the weights.
", "votes": 0, "replies": [] }, { "id": 641935, "author_name": "Radek Osmulski", "author_url": "", "post_date": "2019-10-05T10:52:08.040000", "content": "The kernel I created has saved models from various phases of the training: https://www.kaggle.com/radek1/fastai-starter-pack-train-basic-model-and-submit/output
\n\nIt uses different data though.
", "votes": 0, "replies": [] }, { "id": 642056, "author_name": "", "author_url": "", "post_date": "2019-10-05T13:58:43.817000", "content": "", "votes": 1, "replies": [] } ] }, { "id": 631686, "author_name": "", "author_url": "", "post_date": "2019-09-22T13:32:09.143000", "content": "", "votes": 1, "replies": [] }, { "id": 637096, "author_name": "danny iskandar", "author_url": "", "post_date": "2019-09-30T16:51:57.413000", "content": "Hi Radek. Thank you
", "votes": 1, "replies": [] }, { "id": 633226, "author_name": "Krishna Katyal", "author_url": "", "post_date": "2019-09-24T15:19:16.457000", "content": "thanks alot!
", "votes": 1, "replies": [] }, { "id": 632823, "author_name": "Hafid", "author_url": "", "post_date": "2019-09-24T05:30:37.440000", "content": "thank your for sharing
", "votes": 1, "replies": [] }, { "id": 631701, "author_name": "Kevin Weng", "author_url": "", "post_date": "2019-09-22T14:14:16.623000", "content": "Thanks!
", "votes": 1, "replies": [] }, { "id": 631384, "author_name": "Adam Miller", "author_url": "", "post_date": "2019-09-21T23:29:24.457000", "content": "So awesome Radek! Thank you!
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "630848": "[Github repository](https://github.com/radekosmulski/rsna-intracranial)\nFastai [forum post](https://forums.fast.ai/t/share-your-work-here-part-2/41392/129?u=radek)\n\nHi Everyone!\n\nPlease allow me the pleasure of sharing with you a starter pack using [FastAi](https://github.com/fastai/fastai)\n\nThe starter pack begins with downloading and processing the data. I extract images from the DICOM files and resize them to 112x112. I then randomly split the data assigning 20% of examples to the validation set. I use pretrained resnet-18 and a custom classification head that comes with fastai out of the box with default values.\n\nThe training is done with progressive unfreezing and the last phase of training with a completely unfrozen model leverages discriminative learning rates. \n\nThe model is trained for a total of 3 epochs without data augmentation. On a 1080ti training completes under 15 minutes.\n\nThe techniques I use are covered in the [ULMFIT paper](https://arxiv.org/abs/1801.06146). From what I remember I believe they are also introduced in the first three lectures of the FastAi [Deep Learning for Coders v2 course](https://course.fast.ai/) (it's free). \n\nBest of luck in the competition and have fun!\n\n**EDIT**: Please note, `window_and_normalize` should divide `window_width` by 2 in order to align with the 'brain window' traditionally used by radiologists for visualization. As is right now, the window is twice as wide. Good discussion on relevance of windowing to visualization / modelling in this [thread](https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/109328#latest-630565)", "659619": "thanks for the starter pack. learned a lot", "651858": "Thanks for an amazing starter kit using fastai!", "642513": "Thank you very much for contributing so much to the community, Radek!\nYour notebooks helped me to get started faster with the data preprocessing & Co.\nKeep up the great work! :-D", "634830": "Good to see you in kaggle Radek, thanks for the great insight. I hope I can learn lots from you again just as following you in the whale competition :) ", "634065": "Great will take a look at that", "632255": "Awesome!!", "631394": "This is a great help, thanks for sharing your skills!", "631229": "Much appreciated!", "631089": "Thanks @radek1 for the amazing starter pack.", "630959": "Wow.... FastAI...!!\n\nThanks for the Starter Kit...!! @radek1 ", "630879": "Thanks for an amazing starter kit using fastai!", "631076": "Thanks for creating and sharing this @radek1!\nDoes this mean, we’ll be seeing you on Kaggle LB(s) again? 😄 ", "1019683": "Dear Sir I have some questions:\n\n1. In your '08_CAM_binary_classifier', may I know where does 'train_labels_as_strings_with_label_class.csv' come from? I do not see its definition in this or others files. Is it the same as 'train_labels_as_strings.csv'?\n\n2. I have installed the fastai_dev via 'python setup.py install' from https://github.com/fastai/fastai_dev\nand then set: path_to_fastai_dev_local = 'C:/Users/Administrator/Desktop/fastai_dev-master/dev'\n sys.path.append(path_to_fastai_dev_local)\n\nThen I have error: \n\nIn [19]: from local.torch_basics import *\nTraceback (most recent call last):\n\n File \"