Here is my submission\nhttps://www.kaggle.com/code/ykotoba/notebook79f1ce2f96\nI've checked that my code haven't read sample_submission, and I've compared my output with the requirement(and other's submission), but I cannot find any mistakes. I dont know how to fix it then(sad)
", "messages": [ { "id": "3388814", "postDate": "01/09/2026 16:37:46", "content": "Here is my submission\nhttps://www.kaggle.com/code/ykotoba/notebook79f1ce2f96\nI've checked that my code haven't read sample_submission, and I've compared my output with the requirement(and other's submission), but I cannot find any mistakes. I dont know how to fix it then(sad)
", "rawMarkdown": "Here is my submission\nhttps://www.kaggle.com/code/ykotoba/notebook79f1ce2f96\nI've checked that my code haven't read sample_submission, and I've compared my output with the requirement(and other's submission), but I cannot find any mistakes. I dont know how to fix it then(sad)", "votes": null }, { "id": "3388820", "postDate": "01/09/2026 16:48:30", "content": "Its hard to fork and run your notebook when you have private dataset. Make it public and I will take a look.
", "rawMarkdown": "Its hard to fork and run your notebook when you have private dataset. Make it public and I will take a look.", "votes": null }, { "id": "3388828", "postDate": "01/09/2026 16:56:02", "content": "You cannot read the test images from a data set - which is what you appear to be doing. The correct folder is\n/kaggle/input/csiro-biomass
TEST_IMG_DIR = \"/kaggle/input/dataset/test\" \ntest_df = pd.read_csv('/kaggle/input/dataset/test.csv')\nDon't tell kaggle what folder to save the submission file. Its often going to fail. Running this change and correct folders now.
\nsubmission_df.to_csv('/kaggle/working/submission.csv', index=False)\nDo this
\nsubmission_df.to_csv('submission.csv', index=False)
Not sure what LLM your using to help write the code. I am not real fond of code that does not take advantage of notebooks by breaking the code into cells. Makes it very hard to take advantage of notebook features when all the code is in a single cell. In your case you can't run the code in the editior up to the prediction section without doing the whole thing. Time is Not Your Friend in life or in competitions
\nSo I almost never would use a py file for kaggle. But that's just my preference.
", "rawMarkdown": "Not sure what LLM your using to help write the code. I am not real fond of code that does not take advantage of notebooks by breaking the code into cells. Makes it very hard to take advantage of notebook features when all the code is in a single cell. In your case you can't run the code in the editior up to the prediction section without doing the whole thing. Time is Not Your Friend in life or in competitions \n\nSo I almost never would use a py file for kaggle. But that's just my preference.", "votes": null }, { "id": "3388857", "postDate": "01/09/2026 17:53:28", "content": "Independent of the errors that are causing the submission to fail, your predictions for the single test image that we can see are very low. The best shared notebook with 0.70 LB has 55 as the total dry.
\nI had this myself on a model or two. My fix was to convert the values to log and than reverse for the predictions. In another model I decided my initial learning rate was too low and the model never started to converge, so predictions all close to zero. I forked your code on local and will try the log approach. You might try 0.001 learning rate.
", "rawMarkdown": "Independent of the errors that are causing the submission to fail, your predictions for the single test image that we can see are very low. The best shared notebook with 0.70 LB has 55 as the total dry.\n\nI had this myself on a model or two. My fix was to convert the values to log and than reverse for the predictions. In another model I decided my initial learning rate was too low and the model never started to converge, so predictions all close to zero. I forked your code on local and will try the log approach. You might try 0.001 learning rate.", "votes": null }, { "id": "3388858", "postDate": "01/09/2026 17:54:56", "content": "With the two changes I suggest below the notebook did run and a successful submission resulted. I only trained the model for 2 epochs so my LB score is very low. Let me know if you can't get it running and I will make my copy of your notebook public.
", "rawMarkdown": "With the two changes I suggest below the notebook did run and a successful submission resulted. I only trained the model for 2 epochs so my LB score is very low. Let me know if you can't get it running and I will make my copy of your notebook public.", "votes": null }, { "id": "3388879", "postDate": "01/09/2026 18:35:53", "content": "Got my reminder that I hate tensor flow - getting the right TF version / cuda version to run my newer dual GPU's is too painful. Will fire up an older computer later today or tomorrow and try the log fix than.
", "rawMarkdown": "Got my reminder that I hate tensor flow - getting the right TF version / cuda version to run my newer dual GPU's is too painful. Will fire up an older computer later today or tomorrow and try the log fix than.", "votes": null }, { "id": "3388911", "postDate": "01/09/2026 19:21:03", "content": "Made the learning rate change - values look better but might still be under estimated. LB submission in process.
", "rawMarkdown": "Made the learning rate change - values look better but might still be under estimated. LB submission in process.", "votes": null }, { "id": "3388957", "postDate": "01/09/2026 21:36:46", "content": "LB -0.06 - ran more epochs and learning rate of 0.001.
", "rawMarkdown": "LB -0.06 - ran more epochs and learning rate of 0.001.", "votes": null } ], "comments": [ { "id": 3388820, "author_name": "pcjimmmy", "author_url": "", "post_date": "01/09/2026 16:48:30", "content": "Its hard to fork and run your notebook when you have private dataset. Make it public and I will take a look.
", "votes": null, "replies": [ { "id": 3388828, "author_name": "pcjimmmy", "author_url": "", "post_date": "01/09/2026 16:56:02", "content": "You cannot read the test images from a data set - which is what you appear to be doing. The correct folder is\n/kaggle/input/csiro-biomass
TEST_IMG_DIR = \"/kaggle/input/dataset/test\" \ntest_df = pd.read_csv('/kaggle/input/dataset/test.csv')\nDon't tell kaggle what folder to save the submission file. Its often going to fail. Running this change and correct folders now.
\nsubmission_df.to_csv('/kaggle/working/submission.csv', index=False)\nDo this
\nsubmission_df.to_csv('submission.csv', index=False)
Not sure what LLM your using to help write the code. I am not real fond of code that does not take advantage of notebooks by breaking the code into cells. Makes it very hard to take advantage of notebook features when all the code is in a single cell. In your case you can't run the code in the editior up to the prediction section without doing the whole thing. Time is Not Your Friend in life or in competitions
\nSo I almost never would use a py file for kaggle. But that's just my preference.
", "votes": null, "replies": [] } ] }, { "id": 3388857, "author_name": "pcjimmmy", "author_url": "", "post_date": "01/09/2026 17:53:28", "content": "Independent of the errors that are causing the submission to fail, your predictions for the single test image that we can see are very low. The best shared notebook with 0.70 LB has 55 as the total dry.
\nI had this myself on a model or two. My fix was to convert the values to log and than reverse for the predictions. In another model I decided my initial learning rate was too low and the model never started to converge, so predictions all close to zero. I forked your code on local and will try the log approach. You might try 0.001 learning rate.
", "votes": null, "replies": [ { "id": 3388879, "author_name": "pcjimmmy", "author_url": "", "post_date": "01/09/2026 18:35:53", "content": "Got my reminder that I hate tensor flow - getting the right TF version / cuda version to run my newer dual GPU's is too painful. Will fire up an older computer later today or tomorrow and try the log fix than.
", "votes": null, "replies": [] }, { "id": 3388911, "author_name": "pcjimmmy", "author_url": "", "post_date": "01/09/2026 19:21:03", "content": "Made the learning rate change - values look better but might still be under estimated. LB submission in process.
", "votes": null, "replies": [ { "id": 3388957, "author_name": "pcjimmmy", "author_url": "", "post_date": "01/09/2026 21:36:46", "content": "LB -0.06 - ran more epochs and learning rate of 0.001.
", "votes": null, "replies": [] } ] } ] }, { "id": 3388858, "author_name": "pcjimmmy", "author_url": "", "post_date": "01/09/2026 17:54:56", "content": "With the two changes I suggest below the notebook did run and a successful submission resulted. I only trained the model for 2 epochs so my LB score is very low. Let me know if you can't get it running and I will make my copy of your notebook public.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3388814": "Here is my submission\nhttps://www.kaggle.com/code/ykotoba/notebook79f1ce2f96\nI've checked that my code haven't read sample_submission, and I've compared my output with the requirement(and other's submission), but I cannot find any mistakes. I dont know how to fix it then(sad)", "3388820": "Its hard to fork and run your notebook when you have private dataset. Make it public and I will take a look.", "3388828": "You cannot read the test images from a data set - which is what you appear to be doing. The correct folder is\n`/kaggle/input/csiro-biomass`\n\n\n\n```python\nTEST_IMG_DIR = \"/kaggle/input/dataset/test\" \ntest_df = pd.read_csv('/kaggle/input/dataset/test.csv')\n```", "3388833": "Don't tell kaggle what folder to save the submission file. Its often going to fail. Running this change and correct folders now.\n\n```python\nsubmission_df.to_csv('/kaggle/working/submission.csv', index=False)\n\n```\nDo this\n\n`submission_df.to_csv('submission.csv', index=False)`", "3388837": "Not sure what LLM your using to help write the code. I am not real fond of code that does not take advantage of notebooks by breaking the code into cells. Makes it very hard to take advantage of notebook features when all the code is in a single cell. In your case you can't run the code in the editior up to the prediction section without doing the whole thing. Time is Not Your Friend in life or in competitions \n\nSo I almost never would use a py file for kaggle. But that's just my preference.", "3388857": "Independent of the errors that are causing the submission to fail, your predictions for the single test image that we can see are very low. The best shared notebook with 0.70 LB has 55 as the total dry.\n\nI had this myself on a model or two. My fix was to convert the values to log and than reverse for the predictions. In another model I decided my initial learning rate was too low and the model never started to converge, so predictions all close to zero. I forked your code on local and will try the log approach. You might try 0.001 learning rate.", "3388858": "With the two changes I suggest below the notebook did run and a successful submission resulted. I only trained the model for 2 epochs so my LB score is very low. Let me know if you can't get it running and I will make my copy of your notebook public.", "3388879": "Got my reminder that I hate tensor flow - getting the right TF version / cuda version to run my newer dual GPU's is too painful. Will fire up an older computer later today or tomorrow and try the log fix than.", "3388911": "Made the learning rate change - values look better but might still be under estimated. LB submission in process.", "3388957": "LB -0.06 - ran more epochs and learning rate of 0.001." }, "source": "meta" }