The width and height for image_id 6281 in the train.csv is (75116, 48293), but the loaded image has shape (50970, 43763).
\nVerified by the output of this notebook: https://www.kaggle.com/maxchen303/check-img-size
@sohier @masadia Please take a look at this issue. I am no sure if this is the only error in the label files.
", "messages": [ { "id": 2548990, "postDate": "2023-12-04T23:14:56.813Z", "content": "The width and height for image_id 6281 in the train.csv is (75116, 48293), but the loaded image has shape (50970, 43763).
\nVerified by the output of this notebook: https://www.kaggle.com/maxchen303/check-img-size
@sohier @masadia Please take a look at this issue. I am no sure if this is the only error in the label files.
", "rawMarkdown": "The width and height for image_id `6281` in the `train.csv` is `(75116, 48293)`, but the loaded image has shape `(50970, 43763)`. \nVerified by the output of this notebook: https://www.kaggle.com/maxchen303/check-img-size \n\n@sohier @masadia Please take a look at this issue. I am no sure if this is the only error in the label files.\n", "votes": 8 }, { "id": 2549130, "postDate": "2023-12-05T03:32:45.863Z", "content": "No, it's not the only error. In fact, as I reported here, the train.csv file has not been updated after the 49 faulty WSIs have been updated a few weeks ago. And I am starting to suspect that there could also be erroneous values in the hidden test.csv, as I am able to perform inference in a GPU notebook on the full training sample using train.csv (except for those 49 WSIs) without exceptions thrown, while I can't make any successful submissions using the hidden test.csv (the notebook always throws exceptions). @sohier @masadia, could you also please double-check that the image_width/image_height in the hidden test.csv file are all correct?
", "rawMarkdown": "No, it's not the only error. In fact, as I reported [here](https://www.kaggle.com/competitions/UBC-OCEAN/discussion/451892#2511306), the train.csv file has not been updated after the 49 faulty WSIs have been updated a few weeks ago. And I am starting to suspect that there could also be erroneous values in the hidden test.csv, as I am able to perform inference in a GPU notebook on the full training sample using train.csv (except for those 49 WSIs) without exceptions thrown, while I can't make any successful submissions using the hidden test.csv (the notebook always throws exceptions). @sohier @masadia, could you also please double-check that the image_width/image_height in the hidden test.csv file are all correct?", "votes": 2, "replies": [ { "id": 2549501, "postDate": "2023-12-05T09:59:37.857Z", "content": "You are on the spot. There is one of 49 where the size stayed consistent but the other 48: all changed size after update. One thing you can do is to not rely on test.csv sizes and always make sure you check the size of what you read. Good thing is for pyvips for example: you can check the size before doing the timeconsuming conversion to numpy so an easy fix there.
\nBut I see a bigger issue, 'the dilemma of kaggle' so to speak: 'you have some problems in data. you correct those, but what about other things that are affected? do you want to update them too or stay put to stay 'consistent'. I dont think this is solved always optimally but 'such is life' :-)
", "rawMarkdown": "You are on the spot. There is one of 49 where the size stayed consistent but the other 48: all changed size after update. One thing you can do is to not rely on test.csv sizes and always make sure you check the size of what you read. Good thing is for pyvips for example: you can check the size before doing the timeconsuming conversion to numpy so an easy fix there.\n\nBut I see a bigger issue, 'the dilemma of kaggle' so to speak: 'you have some problems in data. you correct those, but what about other things that are affected? do you want to update them too or stay put to stay 'consistent'. I dont think this is solved always optimally but 'such is life' :-)", "votes": 3 } ] }, { "id": 2551911, "postDate": "2023-12-07T04:04:56.783Z", "content": "I have fixed problem with train image size in train.csv but score not improve :((
", "rawMarkdown": "I have fixed problem with train image size in train.csv but score not improve :((" } ], "comments": [ { "id": 2549130, "author_name": "Patrick Robitaille", "author_url": "", "post_date": "2023-12-05T03:32:45.863000", "content": "No, it's not the only error. In fact, as I reported here, the train.csv file has not been updated after the 49 faulty WSIs have been updated a few weeks ago. And I am starting to suspect that there could also be erroneous values in the hidden test.csv, as I am able to perform inference in a GPU notebook on the full training sample using train.csv (except for those 49 WSIs) without exceptions thrown, while I can't make any successful submissions using the hidden test.csv (the notebook always throws exceptions). @sohier @masadia, could you also please double-check that the image_width/image_height in the hidden test.csv file are all correct?
", "votes": 2, "replies": [ { "id": 2549501, "author_name": "GUNER", "author_url": "", "post_date": "2023-12-05T09:59:37.857000", "content": "You are on the spot. There is one of 49 where the size stayed consistent but the other 48: all changed size after update. One thing you can do is to not rely on test.csv sizes and always make sure you check the size of what you read. Good thing is for pyvips for example: you can check the size before doing the timeconsuming conversion to numpy so an easy fix there.
\nBut I see a bigger issue, 'the dilemma of kaggle' so to speak: 'you have some problems in data. you correct those, but what about other things that are affected? do you want to update them too or stay put to stay 'consistent'. I dont think this is solved always optimally but 'such is life' :-)
", "votes": 3, "replies": [] } ] }, { "id": 2551911, "author_name": "Quan Vu", "author_url": "", "post_date": "2023-12-07T04:04:56.783000", "content": "I have fixed problem with train image size in train.csv but score not improve :((
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2548990": "The width and height for image_id `6281` in the `train.csv` is `(75116, 48293)`, but the loaded image has shape `(50970, 43763)`. \nVerified by the output of this notebook: https://www.kaggle.com/maxchen303/check-img-size \n\n@sohier @masadia Please take a look at this issue. I am no sure if this is the only error in the label files.\n", "2549130": "No, it's not the only error. In fact, as I reported [here](https://www.kaggle.com/competitions/UBC-OCEAN/discussion/451892#2511306), the train.csv file has not been updated after the 49 faulty WSIs have been updated a few weeks ago. And I am starting to suspect that there could also be erroneous values in the hidden test.csv, as I am able to perform inference in a GPU notebook on the full training sample using train.csv (except for those 49 WSIs) without exceptions thrown, while I can't make any successful submissions using the hidden test.csv (the notebook always throws exceptions). @sohier @masadia, could you also please double-check that the image_width/image_height in the hidden test.csv file are all correct?", "2551911": "I have fixed problem with train image size in train.csv but score not improve :((" } }