It seems that HPACellSegmentator cost large GPU memory. It is easy for me to lead to OOM issue. For now, I find that use the weight from @rdizzl3 could help (thanks for your work!!!). Is there any way else that work?
", "messages": [ { "id": 1287287, "postDate": "2021-04-28T21:52:56.820Z", "content": "I tried many things. You can decrease 'scale factor' in the CellSegmentator but it does decrease the cell mask ability to some extent: .25 seems to be optimal. I am using a tweaked version of the label_cell function from https://www.kaggle.com/samusram/even-faster-hpa-cell-segmentation in place of the imported package from CellSegmentator. I'll be interested to hear if others have ideas in this area.
", "rawMarkdown": "I tried many things. You can decrease 'scale factor' in the [CellSegmentator](https://github.com/CellProfiling/HPA-Cell-Segmentation/tree/master/hpacellseg) but it does decrease the cell mask ability to some extent: .25 seems to be optimal. I am using a tweaked version of the label_cell function from https://www.kaggle.com/samusram/even-faster-hpa-cell-segmentation in place of the imported package from [CellSegmentator](https://github.com/CellProfiling/HPA-Cell-Segmentation/blob/master/hpacellseg/utils.py). I'll be interested to hear if others have ideas in this area.", "votes": 1 }, { "id": 1285653, "postDate": "2021-04-27T06:46:32.483Z", "content": "Try using uint8 dtypes wherever possible , that reduces a huge memory overhead. I'm using batch size 8 for my inference notebooks , due to similar issues. CellSegmentator does use up lot of memory :/ If you use cell tile for inference ( I assume you do because you use segmentator) , load images , cast them to uint8 , segment , cast to float32 if needed.
", "rawMarkdown": "Try using uint8 dtypes wherever possible , that reduces a huge memory overhead. I'm using batch size 8 for my inference notebooks , due to similar issues. CellSegmentator does use up lot of memory :/ If you use cell tile for inference ( I assume you do because you use segmentator) , load images , cast them to uint8 , segment , cast to float32 if needed.", "votes": 2 }, { "id": 1284286, "postDate": "2021-04-25T18:57:29.987Z", "content": "Use a smaller batch size? Could that help?
", "rawMarkdown": "Use a smaller batch size? Could that help?", "votes": 2, "replies": [ { "id": 1284465, "postDate": "2021-04-26T00:32:27.857Z", "content": "I'll try that. Thanks. But I just doubt about the time cost.
", "rawMarkdown": "I'll try that. Thanks. But I just doubt about the time cost.", "votes": 1 }, { "id": 1287639, "postDate": "2021-04-29T08:54:50.233Z", "content": "Use an even smaller batch size. I agree
", "rawMarkdown": "Use an even smaller batch size. I agree" } ] }, { "id": 1283960, "postDate": "2021-04-25T12:49:32.193Z", "content": "It seems that HPACellSegmentator cost large GPU memory. It is easy for me to lead to OOM issue. For now, I find that use the weight from @rdizzl3 could help (thanks for your work!!!). Is there any way else that work?
", "rawMarkdown": "It seems that HPACellSegmentator cost large GPU memory. It is easy for me to lead to OOM issue. For now, I find that use the weight from @rdizzl3 could help (thanks for your work!!!). Is there any way else that work?", "votes": 2 } ], "comments": [ { "id": 1287287, "author_name": "Amanda K Kimball", "author_url": "", "post_date": "2021-04-28T21:52:56.820000", "content": "I tried many things. You can decrease 'scale factor' in the CellSegmentator but it does decrease the cell mask ability to some extent: .25 seems to be optimal. I am using a tweaked version of the label_cell function from https://www.kaggle.com/samusram/even-faster-hpa-cell-segmentation in place of the imported package from CellSegmentator. I'll be interested to hear if others have ideas in this area.
", "votes": 1, "replies": [] }, { "id": 1285653, "author_name": "Satwik", "author_url": "", "post_date": "2021-04-27T06:46:32.483000", "content": "Try using uint8 dtypes wherever possible , that reduces a huge memory overhead. I'm using batch size 8 for my inference notebooks , due to similar issues. CellSegmentator does use up lot of memory :/ If you use cell tile for inference ( I assume you do because you use segmentator) , load images , cast them to uint8 , segment , cast to float32 if needed.
", "votes": 2, "replies": [] }, { "id": 1284286, "author_name": "CroDoc", "author_url": "", "post_date": "2021-04-25T18:57:29.987000", "content": "Use a smaller batch size? Could that help?
", "votes": 2, "replies": [ { "id": 1284465, "author_name": "Shihao Shao", "author_url": "", "post_date": "2021-04-26T00:32:27.857000", "content": "I'll try that. Thanks. But I just doubt about the time cost.
", "votes": 1, "replies": [] }, { "id": 1287639, "author_name": "Marek Nurzynski", "author_url": "", "post_date": "2021-04-29T08:54:50.233000", "content": "Use an even smaller batch size. I agree
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1287287": "I tried many things. You can decrease 'scale factor' in the [CellSegmentator](https://github.com/CellProfiling/HPA-Cell-Segmentation/tree/master/hpacellseg) but it does decrease the cell mask ability to some extent: .25 seems to be optimal. I am using a tweaked version of the label_cell function from https://www.kaggle.com/samusram/even-faster-hpa-cell-segmentation in place of the imported package from [CellSegmentator](https://github.com/CellProfiling/HPA-Cell-Segmentation/blob/master/hpacellseg/utils.py). I'll be interested to hear if others have ideas in this area.", "1285653": "Try using uint8 dtypes wherever possible , that reduces a huge memory overhead. I'm using batch size 8 for my inference notebooks , due to similar issues. CellSegmentator does use up lot of memory :/ If you use cell tile for inference ( I assume you do because you use segmentator) , load images , cast them to uint8 , segment , cast to float32 if needed.", "1284286": "Use a smaller batch size? Could that help?", "1283960": "It seems that HPACellSegmentator cost large GPU memory. It is easy for me to lead to OOM issue. For now, I find that use the weight from @rdizzl3 could help (thanks for your work!!!). Is there any way else that work?" } }