{ "id": 114437, "title": "A Late Joiner's Understanding and Notes", "url": "/competitions/understanding_cloud_organization/discussion/114437", "author_name": "", "post_date": "2019-10-26T08:14:44.072743700Z", "votes": 14, "comment_count": 9, "views": 0, "content": "

Hi All ,

\n\n

As a late joiners to this competition , I started to read the discussions by habit and started trying out vanilla architectures , what I could gather from starter code in probably an hour of coding and an hour of paperspace set-up . Observations so far .

\n\n

*Gist from Discussions : *

\n\n
    \n
  1. Four Classes of Cloud : Fish , Flower , Gravel, Sugar .
    2. \n
  2. Apparently there are not much class imbalance , completely empty images devoid of cloud are not there . Multiple clouds , sometimes 2,3 or even all 4 can be together . This makes it quite different from Sevestral Competition .
    3. \n

  3. Here is an image that I have taken from \"Find me in the Clouds\" EDA notebook to give an idea about frequency of classes . \n\"\"

    4. \n

  4. There are major difference from Sevestral is the Noisy Masks . Many times masks dont cover specific areas of cloud . Many times they overlap . These masks looks like are hand drawn approximate rectangular patches around the clouds . We need to learn and match these masks . That could be the biggest challenge to crack . Here are some example taken from the aforementioned notebook . \n\"\"\n\"\"

    5. \n

  5. Dice Coefficient is \n\"The Dice coefficient is defined to be 1 when both X and Y are empty. The leaderboard score is the mean of the Dice coefficients for each (Image, Label) pair in the test set.\" . Same rule as Sevestral .

    6. \n
\n\n

*NOTEBOOKS : *

\n\n

Few Notebooks I started with : \nEDA:\nhttps://www.kaggle.com/ekhtiar/eda-find-me-in-the-clouds\nClassifier and Post Processing:\nhttps://www.kaggle.com/samusram/cloud-classifier-for-post-processing\nhttps://www.kaggle.com/ratthachat/cloud-convexhull-polygon-postprocessing-no-gpu

\n\n

Starter Code(pytorch) : \nhttps://www.kaggle.com/dhananjay3/image-segmentation-from-scratch-in-pytorch\nhttps://www.kaggle.com/artgor/segmentation-in-pytorch-using-convenient-tools

\n\n

My Trial

\n\n
    \n

  1. I started with Vanilla Resnet34 -Unet-1/10 Stratified KFold - 20 epochs -Only Segmentation - got .642

    2. \n

  2. Changed the Unet to FPN in 1 and ran for 40 epochs , got .646 .

    3. \n

  3. Now running 5 fold FPN with Resnet34 and will see the results .

    4. \n

  4. With the starter code hyperparameters the model is converging too early (almost in 30/40 epochs) . Need to figure out right loss function etc so that it can train longer . I am using BCE-Dice from the starter code .\n\"\"

    5. \n
\n\n

I can add as I find out more, It has just been couple of days for me to look into this competition, whoever was there in the competition for long , please correct if any understanding is wrong *

", "messages": [ { "id": "658591", "postDate": "10/26/2019 08:14:44", "content": "

Hi All ,

\n\n

As a late joiners to this competition , I started to read the discussions by habit and started trying out vanilla architectures , what I could gather from starter code in probably an hour of coding and an hour of paperspace set-up . Observations so far .

\n\n

*Gist from Discussions : *

\n\n
    \n
  1. Four Classes of Cloud : Fish , Flower , Gravel, Sugar .
    2. \n
  2. Apparently there are not much class imbalance , completely empty images devoid of cloud are not there . Multiple clouds , sometimes 2,3 or even all 4 can be together . This makes it quite different from Sevestral Competition .
    3. \n

  3. Here is an image that I have taken from \"Find me in the Clouds\" EDA notebook to give an idea about frequency of classes . \n\"\"

    4. \n

  4. There are major difference from Sevestral is the Noisy Masks . Many times masks dont cover specific areas of cloud . Many times they overlap . These masks looks like are hand drawn approximate rectangular patches around the clouds . We need to learn and match these masks . That could be the biggest challenge to crack . Here are some example taken from the aforementioned notebook . \n\"\"\n\"\"

    5. \n

  5. Dice Coefficient is \n\"The Dice coefficient is defined to be 1 when both X and Y are empty. The leaderboard score is the mean of the Dice coefficients for each (Image, Label) pair in the test set.\" . Same rule as Sevestral .

    6. \n
\n\n

*NOTEBOOKS : *

\n\n

Few Notebooks I started with : \nEDA:\nhttps://www.kaggle.com/ekhtiar/eda-find-me-in-the-clouds\nClassifier and Post Processing:\nhttps://www.kaggle.com/samusram/cloud-classifier-for-post-processing\nhttps://www.kaggle.com/ratthachat/cloud-convexhull-polygon-postprocessing-no-gpu

\n\n

Starter Code(pytorch) : \nhttps://www.kaggle.com/dhananjay3/image-segmentation-from-scratch-in-pytorch\nhttps://www.kaggle.com/artgor/segmentation-in-pytorch-using-convenient-tools

\n\n

My Trial

\n\n
    \n

  1. I started with Vanilla Resnet34 -Unet-1/10 Stratified KFold - 20 epochs -Only Segmentation - got .642

    2. \n

  2. Changed the Unet to FPN in 1 and ran for 40 epochs , got .646 .

    3. \n

  3. Now running 5 fold FPN with Resnet34 and will see the results .

    4. \n

  4. With the starter code hyperparameters the model is converging too early (almost in 30/40 epochs) . Need to figure out right loss function etc so that it can train longer . I am using BCE-Dice from the starter code .\n\"\"

    5. \n
\n\n

I can add as I find out more, It has just been couple of days for me to look into this competition, whoever was there in the competition for long , please correct if any understanding is wrong *

", "rawMarkdown": "Hi All ,\n\nAs a late joiners to this competition , I started to read the discussions by habit and started trying out vanilla architectures , what I could gather from starter code in probably an hour of coding and an hour of paperspace set-up . Observations so far .\n\n****Gist from Discussions : ****\n\n1. Four Classes of Cloud : Fish , Flower , Gravel, Sugar . \n2. Apparently there are not much class imbalance , completely empty images devoid of cloud are not there . Multiple clouds , sometimes 2,3 or even all 4 can be together . This makes it quite different from Sevestral Competition . \n3. Here is an image that I have taken from \"Find me in the Clouds\" EDA notebook to give an idea about frequency of classes . \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2234817%2F1deeaec9f36522941ef7026b3ee0c860%2Fnewplot.png?generation=1572076763159115&alt=media)\n\n4. There are major difference from Sevestral is the Noisy Masks . Many times masks dont cover specific areas of cloud . Many times they overlap . These masks looks like are hand drawn approximate rectangular patches around the clouds . We need to learn and match these masks . That could be the biggest challenge to crack . Here are some example taken from the aforementioned notebook . \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2234817%2F78b0908494bcace9b76b9288f7d8aabc%2FFlower_Fish_Sugar_Orig.png?generation=1572076941476252&alt=media)\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2234817%2Fbdcd2cb28a55cd259e30638ed9c1ec14%2FOverlapping_Mask.png?generation=1572076946037240&alt=media)\n\n5. Dice Coefficient is \n\"The Dice coefficient is defined to be 1 when both X and Y are empty. The leaderboard score is the mean of the Dice coefficients for each (Image, Label) pair in the test set.\" . Same rule as Sevestral . \n\n****NOTEBOOKS : ****\n\nFew Notebooks I started with : \nEDA:\nhttps://www.kaggle.com/ekhtiar/eda-find-me-in-the-clouds\nClassifier and Post Processing:\nhttps://www.kaggle.com/samusram/cloud-classifier-for-post-processing\nhttps://www.kaggle.com/ratthachat/cloud-convexhull-polygon-postprocessing-no-gpu\n\nStarter Code(pytorch) : \nhttps://www.kaggle.com/dhananjay3/image-segmentation-from-scratch-in-pytorch\nhttps://www.kaggle.com/artgor/segmentation-in-pytorch-using-convenient-tools\n\n**My Trial**\n\n1. I started with Vanilla Resnet34 -Unet-1/10 Stratified KFold - 20 epochs -Only Segmentation - got .642\n\n2. Changed the Unet to FPN in 1 and ran for 40 epochs , got .646 . \n\n3. Now running 5 fold FPN with Resnet34 and will see the results . \n\n4. With the starter code hyperparameters the model is converging too early (almost in 30/40 epochs) . Need to figure out right loss function etc so that it can train longer . I am using BCE-Dice from the starter code .\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2234817%2F23fbb77fd946ff12a2d87c5c5632289b%2FLoss%20Functions.png?generation=1572077622646906&alt=media)\n\n\n**** I can add as I find out more, It has just been couple of days for me to look into this competition, whoever was there in the competition for long , please correct if any understanding is wrong *****", "votes": null }, { "id": "658628", "postDate": "10/26/2019 09:26:55", "content": "

nice. result for 5 fold FPN with Resnet34?

", "rawMarkdown": "nice. result for 5 fold FPN with Resnet34?", "votes": null }, { "id": "658649", "postDate": "10/26/2019 10:02:57", "content": "

Just now got it . It's .647 . No TTA, few common traintime augmentation .. I am still struggling a bit with the loss function and runtime dice calculation . So there could be small mistakes here and there . However , to me FPN looked slightly better than UNET . Also 1/10 stratifiedfold looked better than 1/5 Kfold.

", "rawMarkdown": "Just now got it . It's .647 . No TTA, few common traintime augmentation .. I am still struggling a bit with the loss function and runtime dice calculation . So there could be small mistakes here and there . However , to me FPN looked slightly better than UNET . Also 1/10 stratifiedfold looked better than 1/5 Kfold.", "votes": null }, { "id": "658680", "postDate": "10/26/2019 10:55:16", "content": "

it's good to know,did you try linknet?

", "rawMarkdown": "it's good to know,did you try linknet?", "votes": null }, { "id": "658857", "postDate": "10/26/2019 16:27:26", "content": "

Did you remove small masks to get 642?

", "rawMarkdown": "Did you remove small masks to get 642?", "votes": null }, { "id": "658878", "postDate": "10/26/2019 17:12:08", "content": "

Yes Xuan . Used the post processing from starter code .

", "rawMarkdown": "Yes Xuan . Used the post processing from starter code .", "votes": null }, { "id": "658879", "postDate": "10/26/2019 17:13:48", "content": "

Was going to. I had high hopes on PSP as this is something for which PSP was created for . Scene Parsing . But first pass didn't do anything good . Need to try more .

", "rawMarkdown": "Was going to. I had high hopes on PSP as this is something for which PSP was created for . Scene Parsing . But first pass didn't do anything good . Need to try more .", "votes": null }, { "id": "659070", "postDate": "10/27/2019 02:59:59", "content": "

Just to clarify. Is this a kernels only comp, or a normal comp? Are the 3698 test images that we can download all the test images (or are there hidden private ones)? Are there any restrictions on kernel runtimes etc?

", "rawMarkdown": "Just to clarify. Is this a kernels only comp, or a normal comp? Are the 3698 test images that we can download all the test images (or are there hidden private ones)? Are there any restrictions on kernel runtimes etc?", "votes": null }, { "id": "659071", "postDate": "10/27/2019 03:01:35", "content": "

no this is not kernel only competition,you just generate submission.csv and submit to competition,waiting for your first kernel for this competition :)

", "rawMarkdown": "no this is not kernel only competition,you just generate submission.csv and submit to competition,waiting for your first kernel for this competition :)", "votes": null }, { "id": "659072", "postDate": "10/27/2019 03:07:31", "content": "

1.This is a normal comp . I have trained everything outside and uploaded submission.csv in Kernel and submitted . There is also one good thing . There is no time or internet restriction in the Inference /submission Kernel , so if we want we can train the full pipeline in one kernel and submit .

\n\n
    \n
  1. Please see this discussion . Looks like there is no private data.\"\"
    2. \n
", "rawMarkdown": "1.This is a normal comp . I have trained everything outside and uploaded submission.csv in Kernel and submitted . There is also one good thing . There is no time or internet restriction in the Inference /submission Kernel , so if we want we can train the full pipeline in one kernel and submit . \n\n2. Please see this discussion . Looks like there is no private data.![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2234817%2F4f52bd7a49d9d3fb8254ab9d7c72c195%2FLeaderborad.PNG?generation=1572145646004175&alt=media)", "votes": null } ], "comments": [ { "id": 658628, "author_name": "mobassir", "author_url": "", "post_date": "10/26/2019 09:26:55", "content": "

nice. result for 5 fold FPN with Resnet34?

", "votes": null, "replies": [ { "id": 658649, "author_name": "phoenix9032", "author_url": "", "post_date": "10/26/2019 10:02:57", "content": "

Just now got it . It's .647 . No TTA, few common traintime augmentation .. I am still struggling a bit with the loss function and runtime dice calculation . So there could be small mistakes here and there . However , to me FPN looked slightly better than UNET . Also 1/10 stratifiedfold looked better than 1/5 Kfold.

", "votes": null, "replies": [] }, { "id": 658680, "author_name": "mobassir", "author_url": "", "post_date": "10/26/2019 10:55:16", "content": "

it's good to know,did you try linknet?

", "votes": null, "replies": [] }, { "id": 658879, "author_name": "phoenix9032", "author_url": "", "post_date": "10/26/2019 17:13:48", "content": "

Was going to. I had high hopes on PSP as this is something for which PSP was created for . Scene Parsing . But first pass didn't do anything good . Need to try more .

", "votes": null, "replies": [] } ] }, { "id": 658857, "author_name": "naivelamb", "author_url": "", "post_date": "10/26/2019 16:27:26", "content": "

Did you remove small masks to get 642?

", "votes": null, "replies": [ { "id": 658878, "author_name": "phoenix9032", "author_url": "", "post_date": "10/26/2019 17:12:08", "content": "

Yes Xuan . Used the post processing from starter code .

", "votes": null, "replies": [] } ] }, { "id": 659070, "author_name": "cdeotte", "author_url": "", "post_date": "10/27/2019 02:59:59", "content": "

Just to clarify. Is this a kernels only comp, or a normal comp? Are the 3698 test images that we can download all the test images (or are there hidden private ones)? Are there any restrictions on kernel runtimes etc?

", "votes": null, "replies": [ { "id": 659071, "author_name": "mobassir", "author_url": "", "post_date": "10/27/2019 03:01:35", "content": "

no this is not kernel only competition,you just generate submission.csv and submit to competition,waiting for your first kernel for this competition :)

", "votes": null, "replies": [] }, { "id": 659072, "author_name": "phoenix9032", "author_url": "", "post_date": "10/27/2019 03:07:31", "content": "

1.This is a normal comp . I have trained everything outside and uploaded submission.csv in Kernel and submitted . There is also one good thing . There is no time or internet restriction in the Inference /submission Kernel , so if we want we can train the full pipeline in one kernel and submit .

\n\n
    \n
  1. Please see this discussion . Looks like there is no private data.\"\"
    2. \n
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "658591": "Hi All ,\n\nAs a late joiners to this competition , I started to read the discussions by habit and started trying out vanilla architectures , what I could gather from starter code in probably an hour of coding and an hour of paperspace set-up . Observations so far .\n\n****Gist from Discussions : ****\n\n1. Four Classes of Cloud : Fish , Flower , Gravel, Sugar . \n2. Apparently there are not much class imbalance , completely empty images devoid of cloud are not there . Multiple clouds , sometimes 2,3 or even all 4 can be together . This makes it quite different from Sevestral Competition . \n3. Here is an image that I have taken from \"Find me in the Clouds\" EDA notebook to give an idea about frequency of classes . \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2234817%2F1deeaec9f36522941ef7026b3ee0c860%2Fnewplot.png?generation=1572076763159115&alt=media)\n\n4. There are major difference from Sevestral is the Noisy Masks . Many times masks dont cover specific areas of cloud . Many times they overlap . These masks looks like are hand drawn approximate rectangular patches around the clouds . We need to learn and match these masks . That could be the biggest challenge to crack . Here are some example taken from the aforementioned notebook . \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2234817%2F78b0908494bcace9b76b9288f7d8aabc%2FFlower_Fish_Sugar_Orig.png?generation=1572076941476252&alt=media)\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2234817%2Fbdcd2cb28a55cd259e30638ed9c1ec14%2FOverlapping_Mask.png?generation=1572076946037240&alt=media)\n\n5. Dice Coefficient is \n\"The Dice coefficient is defined to be 1 when both X and Y are empty. The leaderboard score is the mean of the Dice coefficients for each (Image, Label) pair in the test set.\" . Same rule as Sevestral . \n\n****NOTEBOOKS : ****\n\nFew Notebooks I started with : \nEDA:\nhttps://www.kaggle.com/ekhtiar/eda-find-me-in-the-clouds\nClassifier and Post Processing:\nhttps://www.kaggle.com/samusram/cloud-classifier-for-post-processing\nhttps://www.kaggle.com/ratthachat/cloud-convexhull-polygon-postprocessing-no-gpu\n\nStarter Code(pytorch) : \nhttps://www.kaggle.com/dhananjay3/image-segmentation-from-scratch-in-pytorch\nhttps://www.kaggle.com/artgor/segmentation-in-pytorch-using-convenient-tools\n\n**My Trial**\n\n1. I started with Vanilla Resnet34 -Unet-1/10 Stratified KFold - 20 epochs -Only Segmentation - got .642\n\n2. Changed the Unet to FPN in 1 and ran for 40 epochs , got .646 . \n\n3. Now running 5 fold FPN with Resnet34 and will see the results . \n\n4. With the starter code hyperparameters the model is converging too early (almost in 30/40 epochs) . Need to figure out right loss function etc so that it can train longer . I am using BCE-Dice from the starter code .\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2234817%2F23fbb77fd946ff12a2d87c5c5632289b%2FLoss%20Functions.png?generation=1572077622646906&alt=media)\n\n\n**** I can add as I find out more, It has just been couple of days for me to look into this competition, whoever was there in the competition for long , please correct if any understanding is wrong *****", "658628": "nice. result for 5 fold FPN with Resnet34?", "658649": "Just now got it . It's .647 . No TTA, few common traintime augmentation .. I am still struggling a bit with the loss function and runtime dice calculation . So there could be small mistakes here and there . However , to me FPN looked slightly better than UNET . Also 1/10 stratifiedfold looked better than 1/5 Kfold.", "658680": "it's good to know,did you try linknet?", "658857": "Did you remove small masks to get 642?", "658878": "Yes Xuan . Used the post processing from starter code .", "658879": "Was going to. I had high hopes on PSP as this is something for which PSP was created for . Scene Parsing . But first pass didn't do anything good . Need to try more .", "659070": "Just to clarify. Is this a kernels only comp, or a normal comp? Are the 3698 test images that we can download all the test images (or are there hidden private ones)? Are there any restrictions on kernel runtimes etc?", "659071": "no this is not kernel only competition,you just generate submission.csv and submit to competition,waiting for your first kernel for this competition :)", "659072": "1.This is a normal comp . I have trained everything outside and uploaded submission.csv in Kernel and submitted . There is also one good thing . There is no time or internet restriction in the Inference /submission Kernel , so if we want we can train the full pipeline in one kernel and submit . \n\n2. Please see this discussion . Looks like there is no private data.![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2234817%2F4f52bd7a49d9d3fb8254ab9d7c72c195%2FLeaderborad.PNG?generation=1572145646004175&alt=media)" }, "source": "meta" }