{ "id": 240398, "title": "[LB: 0.360] Starter YOLOv5(Image)+EfficientNet(Study)", "url": "/competitions/siim-covid19-detection/discussion/240398", "author_name": "", "post_date": "2021-05-19T16:39:20.706190400Z", "votes": 14, "comment_count": 2, "views": 0, "content": "

I've published some notebooks for getting started. I hiope this helps,

\n

Overview:

\n\n

Notebooks:

\n

Study-Level:

\n\n

Image-Level:

\n\n

Dataset:

\n

JPEG

\n\n

TFRECORD

\n\n

Happy Kaggling :)

", "messages": [ { "id": "1315229", "postDate": "05/19/2021 16:39:20", "content": "

I've published some notebooks for getting started. I hiope this helps,

\n

Overview:

\n\n

Notebooks:

\n

Study-Level:

\n\n

Image-Level:

\n\n

Dataset:

\n

JPEG

\n\n

TFRECORD

\n\n

Happy Kaggling :)

", "rawMarkdown": "I've published some notebooks for getting started. I hiope this helps,\n\n# Overview:\n* Basic idea was to use **classification** model for **Study-Level** & **detection** model for **Image-Level**,\n\n# Notebooks:\n#### Study-Level:\n* **train**: [SIIM-COVID-19: Study-Level [train] TPU🩺](https://www.kaggle.com/awsaf49/siim-covid-19-study-level-train-tpu/)\n* **infer**: [SIIM-COVID-19: Study-Level [infer]🩺](https://www.kaggle.com/awsaf49/siim-covid-19-study-level-infer) [LB: **0.360**]\n* **data**: [SIIM-COVID-19: 512x512 tfrec Data](https://www.kaggle.com/awsaf49/siim-covid-19-512x512-tfrec-data)\n#### Image-Level:\n* **train**: [SIIM-COVID-19: YOLOv5 Image-Level [train]](https://www.kaggle.com/awsaf49/siim-covid-19-yolov5-image-level-train)\n* **infer**: [SIIM-COVID-19: YOLOv5 Image-Level [infer]](https://www.kaggle.com/awsaf49/siim-covid-19-yolov5-image-level-infer) **placeholder**, seems someting is wrong with `image-level` data, gives very small score `0.051`.\n# Dataset:\n#### JPEG\n* [1024x1024](https://www.kaggle.com/awsaf49/siimcovid19-1024-jpg-image-dataset)\n* [512x512](https://www.kaggle.com/awsaf49/siimcovid19-512-jpg-image-dataset)\n* [256x256](https://www.kaggle.com/awsaf49/siimcovid19-256-jpg-image-dataset)\n#### TFRECORD\n* [1024x1024](https://www.kaggle.com/awsaf49/siimcovid19-1024x1024-tfrec-dataset)\n* [512x512](https://www.kaggle.com/awsaf49/siimcovid19-512x512-tfrec-dataset)\n* [256x256](https://www.kaggle.com/awsaf49/siimcovid19-256x256-tfrec-dataset)\n\nHappy Kaggling :)", "votes": null }, { "id": "1315257", "postDate": "05/19/2021 17:00:27", "content": "
for idx in range(4):\n        conf =  row[str(idx)]\n        if conf>thr:\n            string+=f'{label2name[idx]} {conf:0.2f} 0 0 1 1 '\n
\n

Because the score is calculated by map, there is no need to set a threshold. I also only train Study-Level, with a score of 0.389. Although I use efnb7, I think the difference in performance is because of the threshold.

", "rawMarkdown": "```\nfor idx in range(4):\n conf = row[str(idx)]\n if conf>thr:\n string+=f'{label2name[idx]} {conf:0.2f} 0 0 1 1 '\n```\n\nBecause the score is calculated by map, there is no need to set a threshold. I also only train Study-Level, with a score of 0.389. Although I use efnb7, I think the difference in performance is because of the threshold.", "votes": null }, { "id": "1315262", "postDate": "05/19/2021 17:03:13", "content": "

thanks for the info, will update it.

", "rawMarkdown": "thanks for the info, will update it.", "votes": null } ], "comments": [ { "id": 1315257, "author_name": "h053473666", "author_url": "", "post_date": "05/19/2021 17:00:27", "content": "
for idx in range(4):\n        conf =  row[str(idx)]\n        if conf>thr:\n            string+=f'{label2name[idx]} {conf:0.2f} 0 0 1 1 '\n
\n

Because the score is calculated by map, there is no need to set a threshold. I also only train Study-Level, with a score of 0.389. Although I use efnb7, I think the difference in performance is because of the threshold.

", "votes": null, "replies": [ { "id": 1315262, "author_name": "awsaf49", "author_url": "", "post_date": "05/19/2021 17:03:13", "content": "

thanks for the info, will update it.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1315229": "I've published some notebooks for getting started. I hiope this helps,\n\n# Overview:\n* Basic idea was to use **classification** model for **Study-Level** & **detection** model for **Image-Level**,\n\n# Notebooks:\n#### Study-Level:\n* **train**: [SIIM-COVID-19: Study-Level [train] TPU🩺](https://www.kaggle.com/awsaf49/siim-covid-19-study-level-train-tpu/)\n* **infer**: [SIIM-COVID-19: Study-Level [infer]🩺](https://www.kaggle.com/awsaf49/siim-covid-19-study-level-infer) [LB: **0.360**]\n* **data**: [SIIM-COVID-19: 512x512 tfrec Data](https://www.kaggle.com/awsaf49/siim-covid-19-512x512-tfrec-data)\n#### Image-Level:\n* **train**: [SIIM-COVID-19: YOLOv5 Image-Level [train]](https://www.kaggle.com/awsaf49/siim-covid-19-yolov5-image-level-train)\n* **infer**: [SIIM-COVID-19: YOLOv5 Image-Level [infer]](https://www.kaggle.com/awsaf49/siim-covid-19-yolov5-image-level-infer) **placeholder**, seems someting is wrong with `image-level` data, gives very small score `0.051`.\n# Dataset:\n#### JPEG\n* [1024x1024](https://www.kaggle.com/awsaf49/siimcovid19-1024-jpg-image-dataset)\n* [512x512](https://www.kaggle.com/awsaf49/siimcovid19-512-jpg-image-dataset)\n* [256x256](https://www.kaggle.com/awsaf49/siimcovid19-256-jpg-image-dataset)\n#### TFRECORD\n* [1024x1024](https://www.kaggle.com/awsaf49/siimcovid19-1024x1024-tfrec-dataset)\n* [512x512](https://www.kaggle.com/awsaf49/siimcovid19-512x512-tfrec-dataset)\n* [256x256](https://www.kaggle.com/awsaf49/siimcovid19-256x256-tfrec-dataset)\n\nHappy Kaggling :)", "1315257": "```\nfor idx in range(4):\n conf = row[str(idx)]\n if conf>thr:\n string+=f'{label2name[idx]} {conf:0.2f} 0 0 1 1 '\n```\n\nBecause the score is calculated by map, there is no need to set a threshold. I also only train Study-Level, with a score of 0.389. Although I use efnb7, I think the difference in performance is because of the threshold.", "1315262": "thanks for the info, will update it." }, "source": "meta" }