I trained on train images only. My attempts with pseudo labeling on extra-images did not work. The reason for this is likely the lack of experimentation or quite possibly my incompetence.
\n\nI used 6fold seresnext 101\nlr: 2e-4 \nbatch size: 16\ninput size: 448x448x3\nepochs: 5\naugmentations: standard\ntta: 8x
\n\ncv: ~0.92
\n\nBy averaging folds, it scored 0.927+ on public lb and 0.935+ on private lb.
\n\nMy CV improvements did not reflect on LB.\nI believe strongly that training data is noisy.
", "messages": [ { "id": 541284, "postDate": "2019-06-02T06:45:24.993Z", "content": "I trained on train images only. My attempts with pseudo labeling on extra-images did not work. The reason for this is likely the lack of experimentation or quite possibly my incompetence.
\n\nI used 6fold seresnext 101\nlr: 2e-4 \nbatch size: 16\ninput size: 448x448x3\nepochs: 5\naugmentations: standard\ntta: 8x
\n\ncv: ~0.92
\n\nBy averaging folds, it scored 0.927+ on public lb and 0.935+ on private lb.
\n\nMy CV improvements did not reflect on LB.\nI believe strongly that training data is noisy.
", "rawMarkdown": "I trained on train images only. My attempts with pseudo labeling on extra-images did not work. The reason for this is likely the lack of experimentation or quite possibly my incompetence. \n\nI used 6fold seresnext 101\nlr: 2e-4 \nbatch size: 16\ninput size: 448x448x3\nepochs: 5\naugmentations: standard\ntta: 8x\n\ncv: ~0.92\n\nBy averaging folds, it scored 0.927+ on public lb and 0.935+ on private lb. \n\nMy CV improvements did not reflect on LB.\nI believe strongly that training data is noisy.", "votes": 21 }, { "id": 541335, "postDate": "2019-06-02T08:54:06.267Z", "content": "Thanks for your solution. Great work!
\n\nThanks again!
", "rawMarkdown": "Thanks for your solution. Great work!\n\n- Did you use Kaggle Kernels for running or your own machine / cloud ?\n- How much time did it take to train ?\n\nThanks again!", "replies": [ { "id": 544646, "postDate": "2019-06-05T18:20:01.343Z", "content": "I only have access to kernels. \ncca 9 hours is training and inference combined
", "rawMarkdown": "I only have access to kernels. \ncca 9 hours is training and inference combined", "votes": 1 }, { "id": 545941, "postDate": "2019-06-06T04:09:50.597Z", "content": "Thanks for your reply @valanm. Would it be possible to share code for the same ? I am trying with fastai and the results are not so high. I will shortly share the kernel with the results .
", "rawMarkdown": "Thanks for your reply @valanm. Would it be possible to share code for the same ? I am trying with fastai and the results are not so high. I will shortly share the kernel with the results .", "votes": 1 }, { "id": 550685, "postDate": "2019-06-11T23:24:08.710Z", "content": "@ambarish Do you have your kernal to share? We could discuss different approaches.
", "rawMarkdown": "@ambarish Do you have your kernal to share? We could discuss different approaches.", "votes": 1 }, { "id": 550772, "postDate": "2019-06-12T02:41:02.013Z", "content": "\n\nhttps://www.kaggle.com/ambarish/icassava-2019-cadene-models?scriptVersionId=15125391
\n\nThis has the private score of 0.91386
\n\nhttps://www.kaggle.com/ambarish/icassava-2019-cadene-models-2.
\n\nThis uses pretrained cadene models using se_resnext50_32x4d.
", "rawMarkdown": "@jedhassell \n\nhttps://www.kaggle.com/ambarish/icassava-2019-cadene-models?scriptVersionId=15125391 \n\nThis has the private score of 0.91386\n\n-----------------------------------------------------------\n\nhttps://www.kaggle.com/ambarish/icassava-2019-cadene-models-2.\n\nThis uses pretrained cadene models using se_resnext50_32x4d.", "votes": 1 }, { "id": 551610, "postDate": "2019-06-12T23:23:54.960Z", "content": "awesome, thank you!
", "rawMarkdown": "awesome, thank you!", "votes": 1 }, { "id": 551793, "postDate": "2019-06-13T05:57:29.833Z", "content": "@jedhassell would it be possible to share your solution ?
", "rawMarkdown": "@jedhassell would it be possible to share your solution ?" }, { "id": 554144, "postDate": "2019-06-17T04:12:50.823Z", "content": "@ambarish Sorry for the delay. I just started playing around with Kaggle Kernals and wanted to get a sharable one. I basically just did the standard Fastai approach using a Resnet152 pre-trained model. I can't remember the exact epochs I used but playing around with it now, it seems 10 is a good start before it begins overfitting. I played around with a lot of data transformation but the only one that seemed to benefit was the max_zoom attribute.
Also, I played around with Sqeezenet after reading your post but I couldn't get any good results from it. Not sure what I'm doing wrong there.
\n\nOh, and looking at the notebook, line 7 seems to have cut off the results. I saw that on your notebook too. Do you know of anyway around this? It's very annoying not to see the validation error at the end of each epoch.
", "rawMarkdown": "@ambarish Sorry for the delay. I just started playing around with Kaggle Kernals and wanted to get a sharable one. I basically just did the standard [Fastai approach using a Resnet152 pre-trained model](https://www.kaggle.com/jedhassell/cassava-resnet152?scriptVersionId=15809730). I can't remember the exact epochs I used but playing around with it now, it seems 10 is a good start before it begins overfitting. I played around with a lot of data transformation but the only one that seemed to benefit was the `max_zoom` attribute. \n\nAlso, I played around with [Sqeezenet](https://github.com/fastai/fastai/blob/master/fastai/vision/models/__init__.py#L3) after reading your post but I couldn't get any good results from it. Not sure what I'm doing wrong there. \n\nOh, and looking at the notebook, line 7 seems to have cut off the results. I saw that on your notebook too. Do you know of anyway around this? It's very annoying not to see the validation error at the end of each epoch. ", "votes": 1 }, { "id": 554151, "postDate": "2019-06-17T04:24:13.003Z", "content": "@jedhassell The link Fastai approach using a Resnet152 pre-trained model gives a 404 error.
Would you please provide again ?
", "rawMarkdown": "@jedhassell The link Fastai approach using a Resnet152 pre-trained model gives a `404 error`.\n\nWould you please provide again ? " }, { "id": 554565, "postDate": "2019-06-17T16:42:02.667Z", "content": "@ambarish Sorry, please try again.
", "rawMarkdown": "@ambarish [Sorry, please try again.](https://www.kaggle.com/jedhassell/cassava-resnet152?scriptVersionId=15809730)", "votes": 1 } ] }, { "id": 541346, "postDate": "2019-06-02T09:21:50.437Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 544644, "postDate": "2019-06-05T18:18:04.077Z", "content": "Please note that I have not confirm this on this dataset but I believe it until someone disprove it because that's what my logic tells me (and my background in veterinary medicine)
", "rawMarkdown": "1. humans assign labels and humans are prone to errors. (too general but always true)\n2. when dealing with categories which involve one healthy category and several diseases I always expect higher rate of noise: \n- it is sometimes hard to notice symptoms in the early stage of disease so it looks like healthy\n- if two diseases are present symptoms of the one which is less severe (a) has high chance to be overlooked and (b) in these type of challenges people are almost always asked to assign only one label and they assign only dominant disease\n\nPlease note that I have not confirm this on this dataset but I believe it until someone disprove it because that's what my logic tells me (and my background in veterinary medicine)", "votes": 1 }, { "id": 550659, "postDate": "2019-06-11T22:15:51.607Z", "content": "Hello, can you please post your email address here? For some reason Kaggle isn't allowing me to contact you. I wanted to send you instructions for presenting your work, prizes etc.
", "rawMarkdown": "Hello, can you please post your email address here? For some reason Kaggle isn't allowing me to contact you. I wanted to send you instructions for presenting your work, prizes etc." }, { "id": 550778, "postDate": "2019-06-12T02:52:49.237Z", "content": "@tgebru123 \nCan everyone see the code for the winning prizes ?
", "rawMarkdown": "@tgebru123 \nCan everyone see the code for the winning prizes ?" }, { "id": 551432, "postDate": "2019-06-12T17:42:37.363Z", "content": "Thats the plan as soon as I figure out how to contact everyone and send out instructions :-) I also suggest everyone else posts their code/discuss their solutions on Kaggle too, not just the top 3 submissions.
", "rawMarkdown": "Thats the plan as soon as I figure out how to contact everyone and send out instructions :-) I also suggest everyone else posts their code/discuss their solutions on Kaggle too, not just the top 3 submissions.", "votes": 1 }, { "id": 551449, "postDate": "2019-06-12T18:20:44.437Z", "content": "Awesome , hope to learn a lot!
", "rawMarkdown": "Awesome , hope to learn a lot!" }, { "id": 553045, "postDate": "2019-06-15T03:38:51.817Z", "content": "@valanm please respond to my message with instructions when you get a chance.
", "rawMarkdown": "@valanm please respond to my message with instructions when you get a chance." } ] } ], "comments": [ { "id": 541335, "author_name": "Ambarish Ganguly", "author_url": "", "post_date": "2019-06-02T08:54:06.267000", "content": "Thanks for your solution. Great work!
\n\nThanks again!
", "votes": 0, "replies": [ { "id": 544646, "author_name": "Miroslav Valan", "author_url": "", "post_date": "2019-06-05T18:20:01.343000", "content": "I only have access to kernels. \ncca 9 hours is training and inference combined
", "votes": 1, "replies": [] }, { "id": 545941, "author_name": "Ambarish Ganguly", "author_url": "", "post_date": "2019-06-06T04:09:50.597000", "content": "Thanks for your reply @valanm. Would it be possible to share code for the same ? I am trying with fastai and the results are not so high. I will shortly share the kernel with the results .
", "votes": 1, "replies": [] }, { "id": 550685, "author_name": "jed hassell", "author_url": "", "post_date": "2019-06-11T23:24:08.710000", "content": "@ambarish Do you have your kernal to share? We could discuss different approaches.
", "votes": 1, "replies": [] }, { "id": 550772, "author_name": "Ambarish Ganguly", "author_url": "", "post_date": "2019-06-12T02:41:02.013000", "content": "\n\nhttps://www.kaggle.com/ambarish/icassava-2019-cadene-models?scriptVersionId=15125391
\n\nThis has the private score of 0.91386
\n\nhttps://www.kaggle.com/ambarish/icassava-2019-cadene-models-2.
\n\nThis uses pretrained cadene models using se_resnext50_32x4d.
", "votes": 1, "replies": [] }, { "id": 551610, "author_name": "jed hassell", "author_url": "", "post_date": "2019-06-12T23:23:54.960000", "content": "awesome, thank you!
", "votes": 1, "replies": [] }, { "id": 551793, "author_name": "Ambarish Ganguly", "author_url": "", "post_date": "2019-06-13T05:57:29.833000", "content": "@jedhassell would it be possible to share your solution ?
", "votes": 0, "replies": [] }, { "id": 554144, "author_name": "jed hassell", "author_url": "", "post_date": "2019-06-17T04:12:50.823000", "content": "@ambarish Sorry for the delay. I just started playing around with Kaggle Kernals and wanted to get a sharable one. I basically just did the standard Fastai approach using a Resnet152 pre-trained model. I can't remember the exact epochs I used but playing around with it now, it seems 10 is a good start before it begins overfitting. I played around with a lot of data transformation but the only one that seemed to benefit was the max_zoom attribute.
Also, I played around with Sqeezenet after reading your post but I couldn't get any good results from it. Not sure what I'm doing wrong there.
\n\nOh, and looking at the notebook, line 7 seems to have cut off the results. I saw that on your notebook too. Do you know of anyway around this? It's very annoying not to see the validation error at the end of each epoch.
", "votes": 1, "replies": [] }, { "id": 554151, "author_name": "Ambarish Ganguly", "author_url": "", "post_date": "2019-06-17T04:24:13.003000", "content": "@jedhassell The link Fastai approach using a Resnet152 pre-trained model gives a 404 error.
Would you please provide again ?
", "votes": 0, "replies": [] }, { "id": 554565, "author_name": "jed hassell", "author_url": "", "post_date": "2019-06-17T16:42:02.667000", "content": "@ambarish Sorry, please try again.
", "votes": 1, "replies": [] } ] }, { "id": 541346, "author_name": "", "author_url": "", "post_date": "2019-06-02T09:21:50.437000", "content": "", "votes": 0, "replies": [ { "id": 544644, "author_name": "Miroslav Valan", "author_url": "", "post_date": "2019-06-05T18:18:04.077000", "content": "Please note that I have not confirm this on this dataset but I believe it until someone disprove it because that's what my logic tells me (and my background in veterinary medicine)
", "votes": 1, "replies": [] }, { "id": 550659, "author_name": "tgebru123", "author_url": "", "post_date": "2019-06-11T22:15:51.607000", "content": "Hello, can you please post your email address here? For some reason Kaggle isn't allowing me to contact you. I wanted to send you instructions for presenting your work, prizes etc.
", "votes": 0, "replies": [] }, { "id": 550778, "author_name": "Ambarish Ganguly", "author_url": "", "post_date": "2019-06-12T02:52:49.237000", "content": "@tgebru123 \nCan everyone see the code for the winning prizes ?
", "votes": 0, "replies": [] }, { "id": 551432, "author_name": "tgebru123", "author_url": "", "post_date": "2019-06-12T17:42:37.363000", "content": "Thats the plan as soon as I figure out how to contact everyone and send out instructions :-) I also suggest everyone else posts their code/discuss their solutions on Kaggle too, not just the top 3 submissions.
", "votes": 1, "replies": [] }, { "id": 551449, "author_name": "Ambarish Ganguly", "author_url": "", "post_date": "2019-06-12T18:20:44.437000", "content": "Awesome , hope to learn a lot!
", "votes": 0, "replies": [] }, { "id": 553045, "author_name": "tgebru123", "author_url": "", "post_date": "2019-06-15T03:38:51.817000", "content": "@valanm please respond to my message with instructions when you get a chance.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "541284": "I trained on train images only. My attempts with pseudo labeling on extra-images did not work. The reason for this is likely the lack of experimentation or quite possibly my incompetence. \n\nI used 6fold seresnext 101\nlr: 2e-4 \nbatch size: 16\ninput size: 448x448x3\nepochs: 5\naugmentations: standard\ntta: 8x\n\ncv: ~0.92\n\nBy averaging folds, it scored 0.927+ on public lb and 0.935+ on private lb. \n\nMy CV improvements did not reflect on LB.\nI believe strongly that training data is noisy.", "541335": "Thanks for your solution. Great work!\n\n- Did you use Kaggle Kernels for running or your own machine / cloud ?\n- How much time did it take to train ?\n\nThanks again!", "541346": "" } }