{ "id": 238371, "title": "A simple approach and only 2 cell-wise models to get a silver [LB Private 0.436]", "url": "/competitions/hpa-single-cell-image-classification/discussion/238371", "author_name": "wayfarer", "post_date": "2021-05-12T03:12:40.154000", "votes": 30, "comment_count": 6, "views": 0, "content": "

Thanks to my teammates and hosts.

\n

I trained cell-wise models using competition and external data (RGB). To get cell-tiles I used a similar code as (https://www.kaggle.com/thedrcat/hpa-cell-tiles-sample-balanced).

\n
    \n
  1. 5-fold models: efficientnet-b0 and efficientnet-b1
    2. \n
  2. Input-shape to CNNs (3x224x224)
    3. \n
  3. Trained for 5 epochs but always chose 1 epoch (it was always better on LB, I kind of found a correlation between local loss and LB score).
    4. \n
\n

The full code can be found here: https://github.com/jovenwayfarer/human_protein_atlas
\nThe inference kernel: https://www.kaggle.com/joven1997/full-data-timm-ensemble

", "messages": [ { "id": 1303350, "postDate": "2021-05-12T03:12:40.153Z", "content": "

Thanks to my teammates and hosts.

\n

I trained cell-wise models using competition and external data (RGB). To get cell-tiles I used a similar code as (https://www.kaggle.com/thedrcat/hpa-cell-tiles-sample-balanced).

\n
    \n
  1. 5-fold models: efficientnet-b0 and efficientnet-b1
    2. \n
  2. Input-shape to CNNs (3x224x224)
    3. \n
  3. Trained for 5 epochs but always chose 1 epoch (it was always better on LB, I kind of found a correlation between local loss and LB score).
    4. \n
\n

The full code can be found here: https://github.com/jovenwayfarer/human_protein_atlas
\nThe inference kernel: https://www.kaggle.com/joven1997/full-data-timm-ensemble

", "rawMarkdown": "Thanks to my teammates and hosts.\n\nI trained cell-wise models using competition and external data (RGB). To get cell-tiles I used a similar code as (https://www.kaggle.com/thedrcat/hpa-cell-tiles-sample-balanced). \n1. 5-fold models: efficientnet-b0 and efficientnet-b1\n2. Input-shape to CNNs (3x224x224)\n3. Trained for 5 epochs but always chose 1 epoch (it was always better on LB, I kind of found a correlation between local loss and LB score).\n\nThe full code can be found here: https://github.com/jovenwayfarer/human_protein_atlas\nThe inference kernel: https://www.kaggle.com/joven1997/full-data-timm-ensemble\n", "votes": 30 }, { "id": 1542289, "postDate": "2021-10-12T10:44:37.860Z", "content": "

Congrats, nice job)

", "rawMarkdown": "Congrats, nice job)", "votes": 1 }, { "id": 1303777, "postDate": "2021-05-12T08:29:20.297Z", "content": "

Thanks for sharing, you used ensemble Effnet B0 and B1 ?

", "rawMarkdown": "Thanks for sharing, you used ensemble Effnet B0 and B1 ?", "votes": 1, "replies": [ { "id": 1303792, "postDate": "2021-05-12T08:38:52.967Z", "content": "

yes, it is right

", "rawMarkdown": "yes, it is right" } ] }, { "id": 1303367, "postDate": "2021-05-12T03:30:56.173Z", "content": "

I used a very similar approach. How many images did you use?

", "rawMarkdown": "I used a very similar approach. How many images did you use?", "votes": 1, "replies": [ { "id": 1303369, "postDate": "2021-05-12T03:34:19.043Z", "content": "

882469 cell-tiles

", "rawMarkdown": "882469 cell-tiles", "votes": 1 } ] }, { "id": 1303398, "postDate": "2021-05-12T03:54:41.287Z", "content": "

@joven1997 Congratulations and Thanks for sharing the approach

", "rawMarkdown": "@joven1997 Congratulations and Thanks for sharing the approach", "votes": 1 } ], "comments": [ { "id": 1542289, "author_name": "Dake[dsmlkz]", "author_url": "", "post_date": "2021-10-12T10:44:37.860000", "content": "

Congrats, nice job)

", "votes": 1, "replies": [] }, { "id": 1303777, "author_name": "Salim Khazem", "author_url": "", "post_date": "2021-05-12T08:29:20.297000", "content": "

Thanks for sharing, you used ensemble Effnet B0 and B1 ?

", "votes": 1, "replies": [ { "id": 1303792, "author_name": "wayfarer", "author_url": "", "post_date": "2021-05-12T08:38:52.967000", "content": "

yes, it is right

", "votes": 0, "replies": [] } ] }, { "id": 1303367, "author_name": "novice03", "author_url": "", "post_date": "2021-05-12T03:30:56.173000", "content": "

I used a very similar approach. How many images did you use?

", "votes": 1, "replies": [ { "id": 1303369, "author_name": "wayfarer", "author_url": "", "post_date": "2021-05-12T03:34:19.043000", "content": "

882469 cell-tiles

", "votes": 1, "replies": [] } ] }, { "id": 1303398, "author_name": "Tensor Girl", "author_url": "", "post_date": "2021-05-12T03:54:41.287000", "content": "

@joven1997 Congratulations and Thanks for sharing the approach

", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "1303350": "Thanks to my teammates and hosts.\n\nI trained cell-wise models using competition and external data (RGB). To get cell-tiles I used a similar code as (https://www.kaggle.com/thedrcat/hpa-cell-tiles-sample-balanced). \n1. 5-fold models: efficientnet-b0 and efficientnet-b1\n2. Input-shape to CNNs (3x224x224)\n3. Trained for 5 epochs but always chose 1 epoch (it was always better on LB, I kind of found a correlation between local loss and LB score).\n\nThe full code can be found here: https://github.com/jovenwayfarer/human_protein_atlas\nThe inference kernel: https://www.kaggle.com/joven1997/full-data-timm-ensemble\n", "1542289": "Congrats, nice job)", "1303777": "Thanks for sharing, you used ensemble Effnet B0 and B1 ?", "1303367": "I used a very similar approach. How many images did you use?", "1303398": "@joven1997 Congratulations and Thanks for sharing the approach" } }