{ "id": 215469, "title": "How to submit a non-zero score?", "url": "/competitions/hpa-single-cell-image-classification/discussion/215469", "author_name": "Darek Kłeczek", "post_date": "2021-01-30T02:30:19.976000", "votes": 5, "comment_count": 2, "views": 0, "content": "

This has taken me a while to figure out, the first baseline I was able to submit uses the following approach:

\n
    \n
  1. Use cell segmentation model shared by the organizer team
    2. \n
  2. Apply some tricks to make it work without memory issues within the competition constraints (specifically predict in batches of the same image sizes)
    3. \n
  3. Predict the most common label (0) for each predicted cell
    4. \n
\n

My kernel is published here: https://www.kaggle.com/thedrcat/hpa-baseline-cell-segmentation

\n

The next step for me is to make running the cell segmentation model more efficient, so that I have time to also predict classes of each segmented cell.

", "messages": [ { "id": 1177050, "postDate": "2021-01-30T02:30:19.977Z", "content": "

This has taken me a while to figure out, the first baseline I was able to submit uses the following approach:

\n
    \n
  1. Use cell segmentation model shared by the organizer team
    2. \n
  2. Apply some tricks to make it work without memory issues within the competition constraints (specifically predict in batches of the same image sizes)
    3. \n
  3. Predict the most common label (0) for each predicted cell
    4. \n
\n

My kernel is published here: https://www.kaggle.com/thedrcat/hpa-baseline-cell-segmentation

\n

The next step for me is to make running the cell segmentation model more efficient, so that I have time to also predict classes of each segmented cell.

", "rawMarkdown": "This has taken me a while to figure out, the first baseline I was able to submit uses the following approach:\n1. Use cell segmentation model shared by the organizer team\n2. Apply some tricks to make it work without memory issues within the competition constraints (specifically predict in batches of the same image sizes)\n3. Predict the most common label (0) for each predicted cell\n\nMy kernel is published here: https://www.kaggle.com/thedrcat/hpa-baseline-cell-segmentation\n\nThe next step for me is to make running the cell segmentation model more efficient, so that I have time to also predict classes of each segmented cell. ", "votes": 5 }, { "id": 1266076, "postDate": "2021-04-07T13:13:12.413Z", "content": "

Thank you Darek for these tips, it is usefill, however in my team we have an issue with our submission.

\n

We will be very gratefull if you have 10 minutes to give a feedback to our code, if not, thanks again for your reviews.

\n

Our submission is here: https://www.kaggle.com/olgapuntous/hpa-final-submission-1?scriptVersionId=58993057

", "rawMarkdown": "Thank you Darek for these tips, it is usefill, however in my team we have an issue with our submission.\n\nWe will be very gratefull if you have 10 minutes to give a feedback to our code, if not, thanks again for your reviews.\n\n Our submission is here: https://www.kaggle.com/olgapuntous/hpa-final-submission-1?scriptVersionId=58993057" }, { "id": 1253431, "postDate": "2021-03-26T18:01:57.973Z", "content": "

I observed that if the assumption is \"Predict the most common label (0) for each predicted cell\" as you mentioned, the score is around 0.029-0.016.

\n

I trained a classifier in my case but I am doing nothing to mitigate class imbalance. So it must be predicting 0 for every image. Also, I assigned each cell to have predicted image-level labels. This is somewhat similar to hard code the prediction to 0. Thus low score.

\n

@thedrcat

", "rawMarkdown": "I observed that if the assumption is \"Predict the most common label (0) for each predicted cell\" as you mentioned, the score is around 0.029-0.016. \n\nI trained a classifier in my case but I am doing nothing to mitigate class imbalance. So it must be predicting 0 for every image. Also, I assigned each cell to have predicted image-level labels. This is somewhat similar to hard code the prediction to 0. Thus low score. \n\n@thedrcat " } ], "comments": [ { "id": 1266076, "author_name": "olga puntous", "author_url": "", "post_date": "2021-04-07T13:13:12.413000", "content": "

Thank you Darek for these tips, it is usefill, however in my team we have an issue with our submission.

\n

We will be very gratefull if you have 10 minutes to give a feedback to our code, if not, thanks again for your reviews.

\n

Our submission is here: https://www.kaggle.com/olgapuntous/hpa-final-submission-1?scriptVersionId=58993057

", "votes": 0, "replies": [] }, { "id": 1253431, "author_name": "Ayush Thakur", "author_url": "", "post_date": "2021-03-26T18:01:57.973000", "content": "

I observed that if the assumption is \"Predict the most common label (0) for each predicted cell\" as you mentioned, the score is around 0.029-0.016.

\n

I trained a classifier in my case but I am doing nothing to mitigate class imbalance. So it must be predicting 0 for every image. Also, I assigned each cell to have predicted image-level labels. This is somewhat similar to hard code the prediction to 0. Thus low score.

\n

@thedrcat

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1177050": "This has taken me a while to figure out, the first baseline I was able to submit uses the following approach:\n1. Use cell segmentation model shared by the organizer team\n2. Apply some tricks to make it work without memory issues within the competition constraints (specifically predict in batches of the same image sizes)\n3. Predict the most common label (0) for each predicted cell\n\nMy kernel is published here: https://www.kaggle.com/thedrcat/hpa-baseline-cell-segmentation\n\nThe next step for me is to make running the cell segmentation model more efficient, so that I have time to also predict classes of each segmented cell. ", "1266076": "Thank you Darek for these tips, it is usefill, however in my team we have an issue with our submission.\n\nWe will be very gratefull if you have 10 minutes to give a feedback to our code, if not, thanks again for your reviews.\n\n Our submission is here: https://www.kaggle.com/olgapuntous/hpa-final-submission-1?scriptVersionId=58993057", "1253431": "I observed that if the assumption is \"Predict the most common label (0) for each predicted cell\" as you mentioned, the score is around 0.029-0.016. \n\nI trained a classifier in my case but I am doing nothing to mitigate class imbalance. So it must be predicting 0 for every image. Also, I assigned each cell to have predicted image-level labels. This is somewhat similar to hard code the prediction to 0. Thus low score. \n\n@thedrcat " } }