I just use the inference notebook from
\n@Rabbit https://www.kaggle.com/code/nghiahoangtrung/hubmap-mmdet3-1-single-fold-inference.
\nThis submission usually works fine, but sometimes it gets [Submission CSV Not Found], especially when the file size is large.
Finally, I got it, It's a matter of formatting. You can found the different with https://www.kaggle.com/code/atom1231/hubmap-mmdet-2-26-public-inference.
", "rawMarkdown": "Finally, I got it, It's a matter of formatting. You can found the different with https://www.kaggle.com/code/atom1231/hubmap-mmdet-2-26-public-inference.", "votes": 1, "replies": [ { "id": 2359863, "postDate": "2023-07-26T12:03:29.817Z", "content": "I can't find the difference from them.Could you tell me?
", "rawMarkdown": "I can't find the difference from them.Could you tell me?", "replies": [ { "id": 2360756, "postDate": "2023-07-27T02:07:19.473Z", "content": "m = 0
\nfor label, score, mask in zip(labels, scores, masks):
\n if(label==0): #blood case
\n sg = mask.astype(np.uint8)
\n kernel = np.ones(shape=(3, 3), dtype=np.uint8)
\n binary_mask = cv2.dilate(sg, kernel, 3)
\n binary_mask = binary_mask.astype(bool)
\n #-------------------------
\n encoded = encode_binary_mask(binary_mask)
\n if m==0: # difference
\n pred_string += f\"0 {score} {encoded.decode('utf-8')}\"
\n m=m+1
\n else:
\n pred_string += f\" 0 {score} {encoded.decode('utf-8')}\"
`pred_strings = []
\nfor label, score, mask in zip(labels,scores, masks):
\npred_strings_txt = \n\n label==:\n\n pred_strings_txt += \n\n:\n\n pred_strings_txt += \n\npred_strings.append(pred_strings_txt)\nprediction_string.append(\" \".join(pred_strings)) # here, I think this is equal to 'm==0' in result,but 'submission csv not found still exists'`
", "rawMarkdown": "`pred_strings = []\n\nfor label, score, mask in zip(labels,scores, masks):\n\n pred_strings_txt = \"\"\n\n if label==0:\n\n pred_strings_txt += f\"0 {score} {encode_binary_mask(mask.astype(np.bool_)).decode('utf-8')}\"\n\n else:\n\n pred_strings_txt += f\"0 {score} {encode_binary_mask(mask.astype(np.bool_)).decode('utf-8')}\"\n\n pred_strings.append(pred_strings_txt)\n\nprediction_string.append(\" \".join(pred_strings)) # here, I think this is equal to 'm==0' in result,but 'submission csv not found still exists'`" }, { "id": 2360784, "postDate": "2023-07-27T02:49:05.730Z", "content": "\n", "rawMarkdown": "> I think this is equal to 'm==0' in result\nno, `lable = 0` only mean it's blood case, `m=0` mean it's the result's first mask." }, { "id": 2360821, "postDate": "2023-07-27T03:34:20.773Z", "content": "I think this is equal to 'm==0' in result
\n
\n no,lable = 0only mean it's blood case,m=0mean it's the result's first mask.
Yes.But \" \".join(pred_strings) make every mask separate by blank space
", "rawMarkdown": "Yes.But \" \".join(pred_strings) make every mask separate by blank space" }, { "id": 2360873, "postDate": "2023-07-27T04:30:39.417Z", "content": "I checked it again, guess the error like that.Pay attention to the number of spaces.
\n\n", "rawMarkdown": "I checked it again, guess the error like that.Pay attention to the number of spaces.\n\n>>> \"\"+\"how\"+\" are\"+\"\"+\" you\"\n'how are you'\n>>> \" \".join([\"how\", \"are\",\"\", \"you\"])\n'how are you'\n" } ] } ] }, { "id": 2352872, "postDate": "2023-07-21T10:09:57.480Z", "content": "\n\n\n\n\"\"+\"how\"+\" are\"+\"\"+\" you\"
\n
\n 'how are you'
\n \" \".join([\"how\", \"are\",\"\", \"you\"])
\n 'how are you'
It doesn’t seem to be a GPU problem.I encountered this problem when I used the same model and configuration but only changed the rcnn score_thr.
", "rawMarkdown": "It doesn’t seem to be a GPU problem.I encountered this problem when I used the same model and configuration but only changed the rcnn score_thr.", "votes": 1 }, { "id": 2352849, "postDate": "2023-07-21T09:58:07.667Z", "content": "I just use the inference notebook from
\n@Rabbit https://www.kaggle.com/code/nghiahoangtrung/hubmap-mmdet3-1-single-fold-inference.
\nThis submission usually works fine, but sometimes it gets [Submission CSV Not Found], especially when the file size is large.
have you solved the problem? I'm facing the same error.
", "rawMarkdown": "have you solved the problem? I'm facing the same error. ", "replies": [ { "id": 2359123, "postDate": "2023-07-26T02:30:27.097Z", "content": "It's a matter of formatting. You can found the different with https://www.kaggle.com/code/atom1231/hubmap-mmdet-2-26-public-inference.
", "rawMarkdown": "It's a matter of formatting. You can found the different with https://www.kaggle.com/code/atom1231/hubmap-mmdet-2-26-public-inference." } ] }, { "id": 2354518, "postDate": "2023-07-22T14:25:55.243Z", "content": "Out of curiosity: did you mange to make TTA work with 3.x version of mmdet?
", "rawMarkdown": "Out of curiosity: did you mange to make TTA work with 3.x version of mmdet?", "replies": [ { "id": 2354541, "postDate": "2023-07-22T14:45:16.650Z", "content": "I tried to add the TTA module, but there seems to be some errors,andLB score dropped to 0.09 instead. I am a Kaggle novice.I want to share the result that I added TTA module and It didn't work as expected to boost LB score, maybe something wrong as I add and modify code. Here is my code and I am working on finding the problem.Link: https://www.kaggle.com/shangjianzhong/try-adding-tta-to-maskrcnn-but-failed-lb-0-09
", "rawMarkdown": "I tried to add the TTA module, but there seems to be some errors,andLB score dropped to 0.09 instead. I am a Kaggle novice.I want to share the result that I added TTA module and It didn't work as expected to boost LB score, maybe something wrong as I add and modify code. Here is my code and I am working on finding the problem.Link: https://www.kaggle.com/shangjianzhong/try-adding-tta-to-maskrcnn-but-failed-lb-0-09", "isDeleted": true }, { "id": 2354551, "postDate": "2023-07-22T14:53:17.503Z", "content": "My original idea was to flip the image from multiple angles before inputting the model. After inputting the model, I rotated the mask and box respectively without changing the score and label. Did I drop the LB because I didn't consider Non-Maximum Suppression?
", "rawMarkdown": "My original idea was to flip the image from multiple angles before inputting the model. After inputting the model, I rotated the mask and box respectively without changing the score and label. Did I drop the LB because I didn't consider Non-Maximum Suppression?", "votes": 1, "isDeleted": true, "replies": [ { "id": 2354568, "postDate": "2023-07-22T15:14:59.057Z", "content": "That's a good question, it could be. I am also new to TTA for segmentation. My idea was to use the TTA from mmdet MultiScaleFlipAug but it seems to be broken in the 3.x version.
\nI will give this more tries and share if I find something interesting.
I found this migration guide https://mmdetection.readthedocs.io/en/latest/migration/config_migration.html.
\nIt seems that the MultiScaleFlipAug has been removed.
I try tta use the solution.https://github.com/open-mmlab/mmyolo/blob/dc85144fab20a970341550794857a2f2f9b11564/demo/large_image_demo.py#L109-L129.
\nbut get the error AssertionError: TTA of instance segmentation does not support now.
\nmaybe should try this. https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/424919
The situation may be related to the wrong environment of some packages. I also encountered such a situation. You can put your notebook log in the comment area. The original author updated the name of the installation package dataset, etc., and the code needs to be adjusted accordingly.
", "rawMarkdown": "The situation may be related to the wrong environment of some packages. I also encountered such a situation. You can put your notebook log in the comment area. The original author updated the name of the installation package dataset, etc., and the code needs to be adjusted accordingly.", "isDeleted": true, "replies": [ { "id": 2353856, "postDate": "2023-07-22T04:43:29.697Z", "content": "Thanks, i 'll try it later.
", "rawMarkdown": "Thanks, i 'll try it later." } ] } ], "comments": [ { "id": 2356222, "author_name": "Ksana", "author_url": "", "post_date": "2023-07-24T03:41:36.227000", "content": "Finally, I got it, It's a matter of formatting. You can found the different with https://www.kaggle.com/code/atom1231/hubmap-mmdet-2-26-public-inference.
", "votes": 1, "replies": [ { "id": 2359863, "author_name": "yinan-kaggle", "author_url": "", "post_date": "2023-07-26T12:03:29.817000", "content": "I can't find the difference from them.Could you tell me?
", "votes": 0, "replies": [ { "id": 2360756, "author_name": "Ksana", "author_url": "", "post_date": "2023-07-27T02:07:19.473000", "content": "m = 0
\nfor label, score, mask in zip(labels, scores, masks):
\n if(label==0): #blood case
\n sg = mask.astype(np.uint8)
\n kernel = np.ones(shape=(3, 3), dtype=np.uint8)
\n binary_mask = cv2.dilate(sg, kernel, 3)
\n binary_mask = binary_mask.astype(bool)
\n #-------------------------
\n encoded = encode_binary_mask(binary_mask)
\n if m==0: # difference
\n pred_string += f\"0 {score} {encoded.decode('utf-8')}\"
\n m=m+1
\n else:
\n pred_string += f\" 0 {score} {encoded.decode('utf-8')}\"
`pred_strings = []
\nfor label, score, mask in zip(labels,scores, masks):
\npred_strings_txt = \n\n label==:\n\n pred_strings_txt += \n\n:\n\n pred_strings_txt += \n\npred_strings.append(pred_strings_txt)\nprediction_string.append(\" \".join(pred_strings)) # here, I think this is equal to 'm==0' in result,but 'submission csv not found still exists'`
", "votes": 0, "replies": [] }, { "id": 2360784, "author_name": "Ksana", "author_url": "", "post_date": "2023-07-27T02:49:05.730000", "content": "\n", "votes": 0, "replies": [] }, { "id": 2360821, "author_name": "yinan-kaggle", "author_url": "", "post_date": "2023-07-27T03:34:20.773000", "content": "I think this is equal to 'm==0' in result
\n
\n no,lable = 0only mean it's blood case,m=0mean it's the result's first mask.
Yes.But \" \".join(pred_strings) make every mask separate by blank space
", "votes": 0, "replies": [] }, { "id": 2360873, "author_name": "Ksana", "author_url": "", "post_date": "2023-07-27T04:30:39.417000", "content": "I checked it again, guess the error like that.Pay attention to the number of spaces.
\n\n", "votes": 0, "replies": [] } ] } ] }, { "id": 2352872, "author_name": "Ksana", "author_url": "", "post_date": "2023-07-21T10:09:57.480000", "content": "\n\n\n\n\"\"+\"how\"+\" are\"+\"\"+\" you\"
\n
\n 'how are you'
\n \" \".join([\"how\", \"are\",\"\", \"you\"])
\n 'how are you'
It doesn’t seem to be a GPU problem.I encountered this problem when I used the same model and configuration but only changed the rcnn score_thr.
", "votes": 1, "replies": [] }, { "id": 2358876, "author_name": "CocaKoala77", "author_url": "", "post_date": "2023-07-25T20:04:06.087000", "content": "have you solved the problem? I'm facing the same error.
", "votes": 0, "replies": [ { "id": 2359123, "author_name": "Ksana", "author_url": "", "post_date": "2023-07-26T02:30:27.097000", "content": "It's a matter of formatting. You can found the different with https://www.kaggle.com/code/atom1231/hubmap-mmdet-2-26-public-inference.
", "votes": 0, "replies": [] } ] }, { "id": 2354518, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2023-07-22T14:25:55.243000", "content": "Out of curiosity: did you mange to make TTA work with 3.x version of mmdet?
", "votes": 0, "replies": [ { "id": 2354541, "author_name": "", "author_url": "", "post_date": "2023-07-22T14:45:16.650000", "content": "I tried to add the TTA module, but there seems to be some errors,andLB score dropped to 0.09 instead. I am a Kaggle novice.I want to share the result that I added TTA module and It didn't work as expected to boost LB score, maybe something wrong as I add and modify code. Here is my code and I am working on finding the problem.Link: https://www.kaggle.com/shangjianzhong/try-adding-tta-to-maskrcnn-but-failed-lb-0-09
", "votes": 0, "replies": [] }, { "id": 2354551, "author_name": "", "author_url": "", "post_date": "2023-07-22T14:53:17.503000", "content": "My original idea was to flip the image from multiple angles before inputting the model. After inputting the model, I rotated the mask and box respectively without changing the score and label. Did I drop the LB because I didn't consider Non-Maximum Suppression?
", "votes": 1, "replies": [ { "id": 2354568, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2023-07-22T15:14:59.057000", "content": "That's a good question, it could be. I am also new to TTA for segmentation. My idea was to use the TTA from mmdet MultiScaleFlipAug but it seems to be broken in the 3.x version.
\nI will give this more tries and share if I find something interesting.
I found this migration guide https://mmdetection.readthedocs.io/en/latest/migration/config_migration.html.
\nIt seems that the MultiScaleFlipAug has been removed.
I try tta use the solution.https://github.com/open-mmlab/mmyolo/blob/dc85144fab20a970341550794857a2f2f9b11564/demo/large_image_demo.py#L109-L129.
\nbut get the error AssertionError: TTA of instance segmentation does not support now.
\nmaybe should try this. https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/424919
The situation may be related to the wrong environment of some packages. I also encountered such a situation. You can put your notebook log in the comment area. The original author updated the name of the installation package dataset, etc., and the code needs to be adjusted accordingly.
", "votes": 0, "replies": [ { "id": 2353856, "author_name": "Ksana", "author_url": "", "post_date": "2023-07-22T04:43:29.697000", "content": "Thanks, i 'll try it later.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "2356222": "Finally, I got it, It's a matter of formatting. You can found the different with https://www.kaggle.com/code/atom1231/hubmap-mmdet-2-26-public-inference.", "2352872": "It doesn’t seem to be a GPU problem.I encountered this problem when I used the same model and configuration but only changed the rcnn score_thr.", "2352849": "I just use the inference notebook from \n@Rabbit https://www.kaggle.com/code/nghiahoangtrung/hubmap-mmdet3-1-single-fold-inference.\nThis submission usually works fine, but sometimes it gets [Submission CSV Not Found], especially when the file size is large.", "2358876": "have you solved the problem? I'm facing the same error. ", "2354518": "Out of curiosity: did you mange to make TTA work with 3.x version of mmdet?", "2353809": "The situation may be related to the wrong environment of some packages. I also encountered such a situation. You can put your notebook log in the comment area. The original author updated the name of the installation package dataset, etc., and the code needs to be adjusted accordingly." } }