{ "id": 220615, "title": "Congratulations to the second team (+690 Big shake up)", "url": "/competitions/cassava-leaf-disease-classification/discussion/220615", "author_name": "Eisa", "post_date": "2021-02-19T01:21:09.638000", "votes": 13, "comment_count": 15, "views": 0, "content": "

I want to know the story behind those three submissions of a novice Kaggler

\n

@devonstanfield

", "messages": [ { "id": 1209622, "postDate": "2021-02-19T01:21:09.640Z", "content": "

I want to know the story behind those three submissions of a novice Kaggler

\n

@devonstanfield

", "rawMarkdown": "I want to know the story behind those three submissions of a novice Kaggler\n\n @devonstanfield", "votes": 12 }, { "id": 1211042, "postDate": "2021-02-19T23:04:16.873Z", "content": "

Hey man, I'm just as surprised as you are. I got in late to the competition, tried implementing my own model designs and learned a lot as i went. The score wasn't very high though, as expected. Near the last few days, I figured I'd move onto another kaggle problem with more time, but before I did I would do some simple fine-tuning of a TFHub model to submit. The public score was .9025 sitting at 660th place close to a bunch of other people, which made me figure they did the same thing. Anyways, here I am, 2nd place, lets see what happens.

", "rawMarkdown": "Hey man, I'm just as surprised as you are. I got in late to the competition, tried implementing my own model designs and learned a lot as i went. The score wasn't very high though, as expected. Near the last few days, I figured I'd move onto another kaggle problem with more time, but before I did I would do some simple fine-tuning of a TFHub model to submit. The public score was .9025 sitting at 660th place close to a bunch of other people, which made me figure they did the same thing. Anyways, here I am, 2nd place, lets see what happens.", "votes": 7, "replies": [ { "id": 1211058, "postDate": "2021-02-19T23:34:23.867Z", "content": "

Congratulations again. We were really curious about the impact of randomness in winning this competition. @devonstanfield

", "rawMarkdown": "Congratulations again. We were really curious about the impact of randomness in winning this competition. @devonstanfield ", "votes": 1 }, { "id": 1211113, "postDate": "2021-02-20T00:54:37.930Z", "content": "

I'm curious if anyone else tried to fine-tune that tfhub model (it was at the very end of my to-do list), without fine-tuning it gives 0.888 private … and that's actually pretty high for a single model with no k-fold/TTA and any fine-tuning on target dataset.
\nI hope there is no data leak through the model weights… is there a link to the original dataset for that pre-trained tfhub model?
\n(but I didn't try other models from the previous 2019 competition, so perhaps such a score is ok)

", "rawMarkdown": "I'm curious if anyone else tried to fine-tune that tfhub model (it was at the very end of my to-do list), without fine-tuning it gives 0.888 private ... and that's actually pretty high for a single model with no k-fold/TTA and any fine-tuning on target dataset.\nI hope there is no data leak through the model weights... is there a link to the original dataset for that pre-trained tfhub model?\n(but I didn't try other models from the previous 2019 competition, so perhaps such a score is ok)" }, { "id": 1211116, "postDate": "2021-02-20T01:03:08.800Z", "content": "

@devonstanfield Congratulation to you! Could you please provide the details on how you finetune that model in TFHub? I didn't have a chance to get at least 0.895 when I finetuned it.

", "rawMarkdown": "@devonstanfield Congratulation to you! Could you please provide the details on how you finetune that model in TFHub? I didn't have a chance to get at least 0.895 when I finetuned it.", "votes": 1 }, { "id": 1211125, "postDate": "2021-02-20T01:13:32.130Z", "content": "

@sparakhin : you may check this link. https://www.tensorflow.org/hub/tutorials/cropnet_cassava. you are correct about the model accuracy when you use the pre-trained model for prediction. the link will guide you to the dataset used in the pre-trained model

", "rawMarkdown": "@sparakhin : you may check this link. https://www.tensorflow.org/hub/tutorials/cropnet_cassava. you are correct about the model accuracy when you use the pre-trained model for prediction. the link will guide you to the dataset used in the pre-trained model" }, { "id": 1211129, "postDate": "2021-02-20T01:23:49.273Z", "content": "

Thanks! Hm, looks like it should be trained using only 2019 data… Interesting to see how @devonstanfield fine-tuned the model

", "rawMarkdown": "Thanks! Hm, looks like it should be trained using only 2019 data... Interesting to see how @devonstanfield fine-tuned the model" }, { "id": 1211131, "postDate": "2021-02-20T01:24:59.887Z", "content": "

@devonstanfield did you change the last layer of this model? the current pre-trained model classifies the 6 labels while this competition label is 5. Based on your published inference notebook, I didn't see any changes in model architecture, especially on the last layer to accommodate the 5 labels only. The pre-trained model can predict the 6th label \"Unknown\". If that is the case, this is an issue. We can assume that the pre-trained model can handle the noise issue using 6th label.

", "rawMarkdown": "@devonstanfield did you change the last layer of this model? the current pre-trained model classifies the 6 labels while this competition label is 5. Based on your published inference notebook, I didn't see any changes in model architecture, especially on the last layer to accommodate the 5 labels only. The pre-trained model can predict the 6th label \"Unknown\". If that is the case, this is an issue. We can assume that the pre-trained model can handle the noise issue using 6th label." }, { "id": 1211172, "postDate": "2021-02-20T03:06:55.450Z", "content": "

@projdev During training, I concat a zero to the training set labels in \"_preprocess_fn\". Though, during inference, I probably should have ignored that 6th column before taking the argmax. Turned out not to be much of a problem.

", "rawMarkdown": "@projdev During training, I concat a zero to the training set labels in \"_preprocess_fn\". Though, during inference, I probably should have ignored that 6th column before taking the argmax. Turned out not to be much of a problem." } ] }, { "id": 1211069, "postDate": "2021-02-19T23:42:25.563Z", "content": "

the whole thing is like a magic to me…so crazy…

", "rawMarkdown": "the whole thing is like a magic to me...so crazy...", "votes": 4 }, { "id": 1209837, "postDate": "2021-02-19T04:13:55.227Z", "content": "

\"\"

\n

Crazy activity graph too

", "rawMarkdown": "![](https://i.imgur.com/U5gJpAQ.png)\n\nCrazy activity graph too ", "votes": 4, "replies": [ { "id": 1210397, "postDate": "2021-02-19T12:00:59.227Z", "content": "

oh, I just saw some even crazier shake ups, like the 40th is came from the 1135th. It's really a lottery story.😂

", "rawMarkdown": "oh, I just saw some even crazier shake ups, like the 40th is came from the 1135th. It's really a lottery story.😂" }, { "id": 1210591, "postDate": "2021-02-19T14:41:22.980Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1211091, "postDate": "2021-02-20T00:05:34.253Z", "content": "

dude, 1135 -> 40 is not a crazier shake up … usually such performance is due to strong experience and finding good CV/LB schema…here we are talking about:

\n
    \n
  1. ~ 3 days kaggle activity
    2. \n
  2. I am doing 3 subs
    3. \n
  3. I drop the mic
    4. \n
  4. profit
    5. \n
", "rawMarkdown": "dude, 1135 -> 40 is not a crazier shake up ... usually such performance is due to strong experience and finding good CV/LB schema...here we are talking about:\n1. ~ 3 days kaggle activity\n2. I am doing 3 subs\n3. I drop the mic\n4. profit" }, { "id": 1211185, "postDate": "2021-02-20T03:26:10.983Z", "content": "

Hahaha what a roller coaster😂 Congratulations @devonstanfield

", "rawMarkdown": "Hahaha what a roller coaster😂 Congratulations @devonstanfield " } ] }, { "id": 1209627, "postDate": "2021-02-19T01:25:57.160Z", "content": "

Me too. Wait and learn.

", "rawMarkdown": "Me too. Wait and learn." } ], "comments": [ { "id": 1211042, "author_name": "Devon Stanfield", "author_url": "", "post_date": "2021-02-19T23:04:16.873000", "content": "

Hey man, I'm just as surprised as you are. I got in late to the competition, tried implementing my own model designs and learned a lot as i went. The score wasn't very high though, as expected. Near the last few days, I figured I'd move onto another kaggle problem with more time, but before I did I would do some simple fine-tuning of a TFHub model to submit. The public score was .9025 sitting at 660th place close to a bunch of other people, which made me figure they did the same thing. Anyways, here I am, 2nd place, lets see what happens.

", "votes": 7, "replies": [ { "id": 1211058, "author_name": "Eisa", "author_url": "", "post_date": "2021-02-19T23:34:23.867000", "content": "

Congratulations again. We were really curious about the impact of randomness in winning this competition. @devonstanfield

", "votes": 1, "replies": [] }, { "id": 1211113, "author_name": "Serhii Parakhin", "author_url": "", "post_date": "2021-02-20T00:54:37.930000", "content": "

I'm curious if anyone else tried to fine-tune that tfhub model (it was at the very end of my to-do list), without fine-tuning it gives 0.888 private … and that's actually pretty high for a single model with no k-fold/TTA and any fine-tuning on target dataset.
\nI hope there is no data leak through the model weights… is there a link to the original dataset for that pre-trained tfhub model?
\n(but I didn't try other models from the previous 2019 competition, so perhaps such a score is ok)

", "votes": 0, "replies": [] }, { "id": 1211116, "author_name": "FGPC", "author_url": "", "post_date": "2021-02-20T01:03:08.800000", "content": "

@devonstanfield Congratulation to you! Could you please provide the details on how you finetune that model in TFHub? I didn't have a chance to get at least 0.895 when I finetuned it.

", "votes": 1, "replies": [] }, { "id": 1211125, "author_name": "FGPC", "author_url": "", "post_date": "2021-02-20T01:13:32.130000", "content": "

@sparakhin : you may check this link. https://www.tensorflow.org/hub/tutorials/cropnet_cassava. you are correct about the model accuracy when you use the pre-trained model for prediction. the link will guide you to the dataset used in the pre-trained model

", "votes": 0, "replies": [] }, { "id": 1211129, "author_name": "Serhii Parakhin", "author_url": "", "post_date": "2021-02-20T01:23:49.273000", "content": "

Thanks! Hm, looks like it should be trained using only 2019 data… Interesting to see how @devonstanfield fine-tuned the model

", "votes": 0, "replies": [] }, { "id": 1211131, "author_name": "FGPC", "author_url": "", "post_date": "2021-02-20T01:24:59.887000", "content": "

@devonstanfield did you change the last layer of this model? the current pre-trained model classifies the 6 labels while this competition label is 5. Based on your published inference notebook, I didn't see any changes in model architecture, especially on the last layer to accommodate the 5 labels only. The pre-trained model can predict the 6th label \"Unknown\". If that is the case, this is an issue. We can assume that the pre-trained model can handle the noise issue using 6th label.

", "votes": 0, "replies": [] }, { "id": 1211172, "author_name": "Devon Stanfield", "author_url": "", "post_date": "2021-02-20T03:06:55.450000", "content": "

@projdev During training, I concat a zero to the training set labels in \"_preprocess_fn\". Though, during inference, I probably should have ignored that 6th column before taking the argmax. Turned out not to be much of a problem.

", "votes": 0, "replies": [] } ] }, { "id": 1211069, "author_name": "Atanas Atanasov", "author_url": "", "post_date": "2021-02-19T23:42:25.563000", "content": "

the whole thing is like a magic to me…so crazy…

", "votes": 4, "replies": [] }, { "id": 1209837, "author_name": "ryches", "author_url": "", "post_date": "2021-02-19T04:13:55.227000", "content": "

\"\"

\n

Crazy activity graph too

", "votes": 4, "replies": [ { "id": 1210397, "author_name": "komakiyyy", "author_url": "", "post_date": "2021-02-19T12:00:59.227000", "content": "

oh, I just saw some even crazier shake ups, like the 40th is came from the 1135th. It's really a lottery story.😂

", "votes": 0, "replies": [] }, { "id": 1210591, "author_name": "", "author_url": "", "post_date": "2021-02-19T14:41:22.980000", "content": "", "votes": 0, "replies": [] }, { "id": 1211091, "author_name": "Atanas Atanasov", "author_url": "", "post_date": "2021-02-20T00:05:34.253000", "content": "

dude, 1135 -> 40 is not a crazier shake up … usually such performance is due to strong experience and finding good CV/LB schema…here we are talking about:

\n
    \n
  1. ~ 3 days kaggle activity
    2. \n
  2. I am doing 3 subs
    3. \n
  3. I drop the mic
    4. \n
  4. profit
    5. \n
", "votes": 0, "replies": [] }, { "id": 1211185, "author_name": "Sripaad Srinivasan", "author_url": "", "post_date": "2021-02-20T03:26:10.983000", "content": "

Hahaha what a roller coaster😂 Congratulations @devonstanfield

", "votes": 0, "replies": [] } ] }, { "id": 1209627, "author_name": "komakiyyy", "author_url": "", "post_date": "2021-02-19T01:25:57.160000", "content": "

Me too. Wait and learn.

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1209622": "I want to know the story behind those three submissions of a novice Kaggler\n\n @devonstanfield", "1211042": "Hey man, I'm just as surprised as you are. I got in late to the competition, tried implementing my own model designs and learned a lot as i went. The score wasn't very high though, as expected. Near the last few days, I figured I'd move onto another kaggle problem with more time, but before I did I would do some simple fine-tuning of a TFHub model to submit. The public score was .9025 sitting at 660th place close to a bunch of other people, which made me figure they did the same thing. Anyways, here I am, 2nd place, lets see what happens.", "1211069": "the whole thing is like a magic to me...so crazy...", "1209837": "![](https://i.imgur.com/U5gJpAQ.png)\n\nCrazy activity graph too ", "1209627": "Me too. Wait and learn." } }