Even though I changed my models, it is still same. Am I doing something wrong with my submission file? Is anyone having same issue?
", "messages": [ { "id": 2146826, "postDate": "2023-02-16T07:28:26.433Z", "content": "B/c you are classifying everything to be positive.
", "rawMarkdown": "B/c you are classifying everything to be positive.", "votes": 1 }, { "id": 2136008, "postDate": "2023-02-09T03:21:18.973Z", "content": "Even though I changed my models, it is still same. Am I doing something wrong with my submission file? Is anyone having same issue?
", "rawMarkdown": "Even though I changed my models, it is still same. Am I doing something wrong with my submission file? Is anyone having same issue?", "votes": 2 }, { "id": 2147933, "postDate": "2023-02-17T02:58:13.460Z", "content": "Do some oversample work (copy 'cancer == 1' lines in csv file)
\nIncrease image resolution (at least 512x512)
Now I am trying using 720 x720 images. and can you please tell me what do you mean by \" upsample work (copy 'cancer == 1' lines in csv file)\". I am not sure what does that mean.
", "rawMarkdown": "Now I am trying using 720 x720 images. and can you please tell me what do you mean by \" upsample work (copy 'cancer == 1' lines in csv file)\". I am not sure what does that mean." }, { "id": 2147942, "postDate": "2023-02-17T03:08:08.583Z", "content": "upsample work: do you mean augment the images and create more like them?
", "rawMarkdown": "upsample work: do you mean augment the images and create more like them?", "replies": [ { "id": 2148199, "postDate": "2023-02-17T08:28:32.130Z", "content": "Sorry, It`s a mistake. I mean oversampling😂. You need to put at least one positive case in each batch, or the model can not learn enough features and make bad prediction.
", "rawMarkdown": "Sorry, It`s a mistake. I mean oversampling😂. You need to put at least one positive case in each batch, or the model can not learn enough features and make bad prediction." } ] } ] }, { "id": 2147918, "postDate": "2023-02-17T02:32:04.447Z", "content": "Can you share your notebook?
", "rawMarkdown": "Can you share your notebook?", "replies": [ { "id": 2147941, "postDate": "2023-02-17T03:07:09.580Z", "content": "yes but its messy. I have shared the code with you. I am training the model in my own computer and then using those weights to just only predict. all the processes are disabled using 'train = False\" in my code.
", "rawMarkdown": "yes but its messy. I have shared the code with you. I am training the model in my own computer and then using those weights to just only predict. all the processes are disabled using 'train = False\" in my code.", "replies": [ { "id": 2147961, "postDate": "2023-02-17T04:00:07.743Z", "content": "Thank you for sharing, and I suggest some options for debugging your code :
\nBefore making csv, check out your metrics again. (e.g. after loading your model in the last section)
Make sure that the format of your submission file is weird. (e.g. the negative value of cancer is more than 99%), which means that the value of threshold is not great to separate data. Therefore, change your threshold to fit the region of data.
Thank you for your suggestion. I didn't understand the first suggestion. 'check out your metrics again'? you mean the evaluation metrics during training?
", "rawMarkdown": "Thank you for your suggestion. I didn't understand the first suggestion. 'check out your metrics again'? you mean the evaluation metrics during training?" }, { "id": 2147977, "postDate": "2023-02-17T04:20:00.353Z", "content": "And also, can you please share me your code? Plese hide the preprocessing part and the model structure. I just want to check the submission parts. Did you train model on kaggle notebook or on your local computer?
", "rawMarkdown": "And also, can you please share me your code? Plese hide the preprocessing part and the model structure. I just want to check the submission parts. Did you train model on kaggle notebook or on your local computer?" }, { "id": 2148043, "postDate": "2023-02-17T05:31:17.883Z", "content": "Compare the evaluation values of the final epochs and the ones after loading your model to see the difference. If they are not different significantly, there is no need to revise the section.
", "rawMarkdown": "Compare the evaluation values of the final epochs and the ones after loading your model to see the difference. If they are not different significantly, there is no need to revise the section." }, { "id": 2148045, "postDate": "2023-02-17T05:34:45.597Z", "content": "Refer to this : https://www.kaggle.com/code/markwijkhuizen/rsna-convnextv2-inference-tensorflow
", "rawMarkdown": "Refer to this : https://www.kaggle.com/code/markwijkhuizen/rsna-convnextv2-inference-tensorflow\n" }, { "id": 2148076, "postDate": "2023-02-17T06:00:37.563Z", "content": "thank you very much. I will try this.
", "rawMarkdown": "thank you very much. I will try this." }, { "id": 2162159, "postDate": "2023-02-28T03:14:40.240Z", "content": "hello, As the competition is over. Can you please share your code with me? I want to learn where I am doing wrong. I will not share your code with anyone. Also the competition is over, so I hope you don't mind.
", "rawMarkdown": "hello, As the competition is over. Can you please share your code with me? I want to learn where I am doing wrong. I will not share your code with anyone. Also the competition is over, so I hope you don't mind." } ] } ] }, { "id": 2139981, "postDate": "2023-02-11T10:41:38.523Z", "content": "There is a quick way to find out if the model learns something. I use let's say 500 positive images, 500 neg images and train. The pf beta should go up, from 0.5, a neutral value, if that doesn't happen then the model doesn't learn anything. So at least how I test is to overfit on a small dataset to pre-filter models.
", "rawMarkdown": "There is a quick way to find out if the model learns something. I use let's say 500 positive images, 500 neg images and train. The pf beta should go up, from 0.5, a neutral value, if that doesn't happen then the model doesn't learn anything. So at least how I test is to overfit on a small dataset to pre-filter models.", "replies": [ { "id": 2142318, "postDate": "2023-02-13T12:47:56Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 2136765, "postDate": "2023-02-09T14:33:41.617Z", "content": "Did you try thresholding on predictions?
", "rawMarkdown": "Did you try thresholding on predictions?", "replies": [ { "id": 2136819, "postDate": "2023-02-09T15:00:02.563Z", "content": "yes. if prediction > 0.5 then output equals to 1.0, else if prediction < 0.1 then output equals to 0.0 else output equals to prediction as it is.
\nCan you please tell me if this if ok? should I just limit the output as 0 or 1 only?
", "rawMarkdown": "yes. if prediction > 0.5 then output equals to 1.0, else if prediction < 0.1 then output equals to 0.0 else output equals to prediction as it is.\n\nCan you please tell me if this if ok? should I just limit the output as 0 or 1 only?", "replies": [ { "id": 2136834, "postDate": "2023-02-09T15:15:18.113Z", "content": "there should be only 1 thresholding value, meaning if pred > 0.5, its 1 else if pred < = 0.5, its 0. The way you implemented, there are a lot of predictions between 0.1 & 0.5 which is left out.
", "rawMarkdown": "there should be only 1 thresholding value, meaning if pred > 0.5, its 1 else if pred < = 0.5, its 0. The way you implemented, there are a lot of predictions between 0.1 & 0.5 which is left out. " }, { "id": 2136843, "postDate": "2023-02-09T15:20:20.953Z", "content": "I thought as the evaluation mentioned in the competition, it will calculate every prediction probability and gives the public score. Thank you for your suggestion. Now, I will try this thresholding method to obtain either 0 or 1 only.
", "rawMarkdown": "I thought as the evaluation mentioned in the competition, it will calculate every prediction probability and gives the public score. Thank you for your suggestion. Now, I will try this thresholding method to obtain either 0 or 1 only. " }, { "id": 2136864, "postDate": "2023-02-09T15:37:15.170Z", "content": "what is your implementation: Backbone, cross-validation strategy etc., if I may ask?
", "rawMarkdown": "what is your implementation: Backbone, cross-validation strategy etc., if I may ask?" }, { "id": 2136873, "postDate": "2023-02-09T15:41:38.640Z", "content": "I used simple CNN, resnet with 3 blocks, resnet50 and tried to increase and decrease filters in the network. Trained it from the scratch. But always, same result 0.04. I did not use any cross-validation for now. No matter what I use, the result is always 0.04. I used only early stopping.
", "rawMarkdown": "I used simple CNN, resnet with 3 blocks, resnet50 and tried to increase and decrease filters in the network. Trained it from the scratch. But always, same result 0.04. I did not use any cross-validation for now. No matter what I use, the result is always 0.04. I used only early stopping.", "votes": 1 }, { "id": 2136875, "postDate": "2023-02-09T15:45:36.357Z", "content": "Ok. Try scoring your implementation. Do let me know the score if you like to. Try looking at others implementation as well. Best wishes.
", "rawMarkdown": "Ok. Try scoring your implementation. Do let me know the score if you like to. Try looking at others implementation as well. Best wishes." }, { "id": 2137319, "postDate": "2023-02-09T21:14:28.147Z", "content": "the main problem could be that u are training from scratch, this make the model take more time to converge and more easy to overfitting
", "rawMarkdown": "the main problem could be that u are training from scratch, this make the model take more time to converge and more easy to overfitting" }, { "id": 2137414, "postDate": "2023-02-10T01:15:17.010Z", "content": "So, I should use pre-trained weights?
", "rawMarkdown": "So, I should use pre-trained weights?" }, { "id": 2137415, "postDate": "2023-02-10T01:15:42.680Z", "content": "Try to use pre-trained weight and set a small lr, maybe it will help you.
", "rawMarkdown": "Try to use pre-trained weight and set a small lr, maybe it will help you." }, { "id": 2139496, "postDate": "2023-02-10T19:58:29.383Z", "content": "I hear about a small lr alot 👀, but what is a small lr :) 1e-4, 1e-5, 1e-6 ?
", "rawMarkdown": "I hear about a small lr alot 👀, but what is a small lr :) 1e-4, 1e-5, 1e-6 ?" }, { "id": 2139523, "postDate": "2023-02-10T20:27:11.633Z", "content": "If you don't have any validation, try to add some. You will easily see different kinds of problems on your val set. My bet here - overfitting.
", "rawMarkdown": "If you don't have any validation, try to add some. You will easily see different kinds of problems on your val set. My bet here - overfitting. " }, { "id": 2142342, "postDate": "2023-02-13T13:13:17.263Z", "content": "name = []
\ncancer_score = []
\nfor i in range(test images):
\n p_img = pre_process(i)
\n prediction = model(p_img)
\n name.append(i)
\n cancer_score.append(e if prediction.item()>threshold else0)
output_sub = pd.DataFrame({'prediction_id': name, 'cancer': cancer_score})
\noutput_sub.to_csv('submission.csv', index=False)
This is my general code for output submission. There is no error while submission. Also, all image id and cancer scores are there in the table. Now I am confused, is it because of my submission code? Can anyone help me?
", "rawMarkdown": "*********************************************\nname = []\ncancer_score = []\nfor i in range(test images):\n p_img = pre_process(i)\n prediction = model(p_img)\n name.append(i)\n cancer_score.append(e if prediction.item()>threshold else0)\n\noutput_sub = pd.DataFrame({'prediction_id': name, 'cancer': cancer_score})\noutput_sub.to_csv('submission.csv', index=False)\n************************************************\n\nThis is my general code for output submission. There is no error while submission. Also, all image id and cancer scores are there in the table. Now I am confused, is it because of my submission code? Can anyone help me?" } ] } ] } ], "comments": [ { "id": 2146826, "author_name": "ThinkSimple", "author_url": "", "post_date": "2023-02-16T07:28:26.433000", "content": "B/c you are classifying everything to be positive.
", "votes": 1, "replies": [] }, { "id": 2147933, "author_name": "RickyLu", "author_url": "", "post_date": "2023-02-17T02:58:13.460000", "content": "Do some oversample work (copy 'cancer == 1' lines in csv file)
\nIncrease image resolution (at least 512x512)
Now I am trying using 720 x720 images. and can you please tell me what do you mean by \" upsample work (copy 'cancer == 1' lines in csv file)\". I am not sure what does that mean.
", "votes": 0, "replies": [] }, { "id": 2147942, "author_name": "Rukesh Prajapati", "author_url": "", "post_date": "2023-02-17T03:08:08.583000", "content": "upsample work: do you mean augment the images and create more like them?
", "votes": 0, "replies": [ { "id": 2148199, "author_name": "RickyLu", "author_url": "", "post_date": "2023-02-17T08:28:32.130000", "content": "Sorry, It`s a mistake. I mean oversampling😂. You need to put at least one positive case in each batch, or the model can not learn enough features and make bad prediction.
", "votes": 0, "replies": [] } ] } ] }, { "id": 2147918, "author_name": "Hyunsoo Lee 1010", "author_url": "", "post_date": "2023-02-17T02:32:04.447000", "content": "Can you share your notebook?
", "votes": 0, "replies": [ { "id": 2147941, "author_name": "Rukesh Prajapati", "author_url": "", "post_date": "2023-02-17T03:07:09.580000", "content": "yes but its messy. I have shared the code with you. I am training the model in my own computer and then using those weights to just only predict. all the processes are disabled using 'train = False\" in my code.
", "votes": 0, "replies": [ { "id": 2147961, "author_name": "Hyunsoo Lee 1010", "author_url": "", "post_date": "2023-02-17T04:00:07.743000", "content": "Thank you for sharing, and I suggest some options for debugging your code :
\nBefore making csv, check out your metrics again. (e.g. after loading your model in the last section)
Make sure that the format of your submission file is weird. (e.g. the negative value of cancer is more than 99%), which means that the value of threshold is not great to separate data. Therefore, change your threshold to fit the region of data.
Thank you for your suggestion. I didn't understand the first suggestion. 'check out your metrics again'? you mean the evaluation metrics during training?
", "votes": 0, "replies": [] }, { "id": 2147977, "author_name": "Rukesh Prajapati", "author_url": "", "post_date": "2023-02-17T04:20:00.353000", "content": "And also, can you please share me your code? Plese hide the preprocessing part and the model structure. I just want to check the submission parts. Did you train model on kaggle notebook or on your local computer?
", "votes": 0, "replies": [] }, { "id": 2148043, "author_name": "Hyunsoo Lee 1010", "author_url": "", "post_date": "2023-02-17T05:31:17.883000", "content": "Compare the evaluation values of the final epochs and the ones after loading your model to see the difference. If they are not different significantly, there is no need to revise the section.
", "votes": 0, "replies": [] }, { "id": 2148045, "author_name": "Hyunsoo Lee 1010", "author_url": "", "post_date": "2023-02-17T05:34:45.597000", "content": "Refer to this : https://www.kaggle.com/code/markwijkhuizen/rsna-convnextv2-inference-tensorflow
", "votes": 0, "replies": [] }, { "id": 2148076, "author_name": "Rukesh Prajapati", "author_url": "", "post_date": "2023-02-17T06:00:37.563000", "content": "thank you very much. I will try this.
", "votes": 0, "replies": [] }, { "id": 2162159, "author_name": "Rukesh Prajapati", "author_url": "", "post_date": "2023-02-28T03:14:40.240000", "content": "hello, As the competition is over. Can you please share your code with me? I want to learn where I am doing wrong. I will not share your code with anyone. Also the competition is over, so I hope you don't mind.
", "votes": 0, "replies": [] } ] } ] }, { "id": 2139981, "author_name": "FlaviuPaul", "author_url": "", "post_date": "2023-02-11T10:41:38.523000", "content": "There is a quick way to find out if the model learns something. I use let's say 500 positive images, 500 neg images and train. The pf beta should go up, from 0.5, a neutral value, if that doesn't happen then the model doesn't learn anything. So at least how I test is to overfit on a small dataset to pre-filter models.
", "votes": 0, "replies": [ { "id": 2142318, "author_name": "", "author_url": "", "post_date": "2023-02-13T12:47:56", "content": "", "votes": 0, "replies": [] } ] }, { "id": 2136765, "author_name": "Sandy", "author_url": "", "post_date": "2023-02-09T14:33:41.617000", "content": "Did you try thresholding on predictions?
", "votes": 0, "replies": [ { "id": 2136819, "author_name": "Rukesh Prajapati", "author_url": "", "post_date": "2023-02-09T15:00:02.563000", "content": "yes. if prediction > 0.5 then output equals to 1.0, else if prediction < 0.1 then output equals to 0.0 else output equals to prediction as it is.
\nCan you please tell me if this if ok? should I just limit the output as 0 or 1 only?
", "votes": 0, "replies": [ { "id": 2136834, "author_name": "Sandy", "author_url": "", "post_date": "2023-02-09T15:15:18.113000", "content": "there should be only 1 thresholding value, meaning if pred > 0.5, its 1 else if pred < = 0.5, its 0. The way you implemented, there are a lot of predictions between 0.1 & 0.5 which is left out.
", "votes": 0, "replies": [] }, { "id": 2136843, "author_name": "Rukesh Prajapati", "author_url": "", "post_date": "2023-02-09T15:20:20.953000", "content": "I thought as the evaluation mentioned in the competition, it will calculate every prediction probability and gives the public score. Thank you for your suggestion. Now, I will try this thresholding method to obtain either 0 or 1 only.
", "votes": 0, "replies": [] }, { "id": 2136864, "author_name": "Sandy", "author_url": "", "post_date": "2023-02-09T15:37:15.170000", "content": "what is your implementation: Backbone, cross-validation strategy etc., if I may ask?
", "votes": 0, "replies": [] }, { "id": 2136873, "author_name": "Rukesh Prajapati", "author_url": "", "post_date": "2023-02-09T15:41:38.640000", "content": "I used simple CNN, resnet with 3 blocks, resnet50 and tried to increase and decrease filters in the network. Trained it from the scratch. But always, same result 0.04. I did not use any cross-validation for now. No matter what I use, the result is always 0.04. I used only early stopping.
", "votes": 1, "replies": [] }, { "id": 2136875, "author_name": "Sandy", "author_url": "", "post_date": "2023-02-09T15:45:36.357000", "content": "Ok. Try scoring your implementation. Do let me know the score if you like to. Try looking at others implementation as well. Best wishes.
", "votes": 0, "replies": [] }, { "id": 2137319, "author_name": "adriel cabral", "author_url": "", "post_date": "2023-02-09T21:14:28.147000", "content": "the main problem could be that u are training from scratch, this make the model take more time to converge and more easy to overfitting
", "votes": 0, "replies": [] }, { "id": 2137414, "author_name": "Rukesh Prajapati", "author_url": "", "post_date": "2023-02-10T01:15:17.010000", "content": "So, I should use pre-trained weights?
", "votes": 0, "replies": [] }, { "id": 2137415, "author_name": "JimmyLiao", "author_url": "", "post_date": "2023-02-10T01:15:42.680000", "content": "Try to use pre-trained weight and set a small lr, maybe it will help you.
", "votes": 0, "replies": [] }, { "id": 2139496, "author_name": "FlaviuPaul", "author_url": "", "post_date": "2023-02-10T19:58:29.383000", "content": "I hear about a small lr alot 👀, but what is a small lr :) 1e-4, 1e-5, 1e-6 ?
", "votes": 0, "replies": [] }, { "id": 2139523, "author_name": "GubadiaCake", "author_url": "", "post_date": "2023-02-10T20:27:11.633000", "content": "If you don't have any validation, try to add some. You will easily see different kinds of problems on your val set. My bet here - overfitting.
", "votes": 0, "replies": [] }, { "id": 2142342, "author_name": "Rukesh Prajapati", "author_url": "", "post_date": "2023-02-13T13:13:17.263000", "content": "name = []
\ncancer_score = []
\nfor i in range(test images):
\n p_img = pre_process(i)
\n prediction = model(p_img)
\n name.append(i)
\n cancer_score.append(e if prediction.item()>threshold else0)
output_sub = pd.DataFrame({'prediction_id': name, 'cancer': cancer_score})
\noutput_sub.to_csv('submission.csv', index=False)
This is my general code for output submission. There is no error while submission. Also, all image id and cancer scores are there in the table. Now I am confused, is it because of my submission code? Can anyone help me?
", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2146826": "B/c you are classifying everything to be positive.", "2136008": "Even though I changed my models, it is still same. Am I doing something wrong with my submission file? Is anyone having same issue?", "2147933": "Do some oversample work (copy 'cancer == 1' lines in csv file)\nIncrease image resolution (at least 512x512)", "2147918": "Can you share your notebook?", "2139981": "There is a quick way to find out if the model learns something. I use let's say 500 positive images, 500 neg images and train. The pf beta should go up, from 0.5, a neutral value, if that doesn't happen then the model doesn't learn anything. So at least how I test is to overfit on a small dataset to pre-filter models.", "2136765": "Did you try thresholding on predictions?" } }