I created translated test set by google translate api and created inference kernel using this. (This is forked from abhishek's kernel. Please upvote original one too!)
\n\nWhat I want to say here is it's better to create private test set. This is not interesting approarch and I assume it's not what host wanted.
\n\nIt will be much more interesting if there is an unknown test set and if it contains unseen language, isn't it??
\n\n[UPDATE] And also please turn off internet access when submission. It doesn't prevent using TPU when training.
", "messages": [ { "id": "785985", "postDate": "03/25/2020 15:00:19", "content": "I created translated test set by google translate api and created inference kernel using this. (This is forked from abhishek's kernel. Please upvote original one too!)
\n\nWhat I want to say here is it's better to create private test set. This is not interesting approarch and I assume it's not what host wanted.
\n\nIt will be much more interesting if there is an unknown test set and if it contains unseen language, isn't it??
\n\n[UPDATE] And also please turn off internet access when submission. It doesn't prevent using TPU when training.
", "rawMarkdown": "I created [translated test set](https://www.kaggle.com/bamps53/test-en-df) by google translate api and created [inference kernel](https://www.kaggle.com/bamps53/inference-with-translated-test-set?scriptVersionId=30818481) using this. (This is forked from [abhishek's kernel](https://www.kaggle.com/abhishek/inference-of-bert-tpu-model-ml-w-validation). Please upvote original one too!)\n\nWhat I want to say here is it's better to create private test set. This is not interesting approarch and I assume it's not what host wanted.\n\nIt will be much more interesting if there is an unknown test set and if it contains unseen language, isn't it??\n\n[UPDATE] And also please turn off internet access when submission. It doesn't prevent using TPU when training.", "votes": null }, { "id": "785993", "postDate": "03/25/2020 15:03:35", "content": "The End Of the comp is near :(
", "rawMarkdown": "The End Of the comp is near :(", "votes": null }, { "id": "786004", "postDate": "03/25/2020 15:09:54", "content": "I hope private test set contains more than 100 languages and will be very long tail one😹
", "rawMarkdown": "I hope private test set contains more than 100 languages and will be very long tail one😹", "votes": null }, { "id": "786021", "postDate": "03/25/2020 15:25:25", "content": "As long as Internet access in inference is allowed not even private test set helps.
", "rawMarkdown": "As long as Internet access in inference is allowed not even private test set helps.", "votes": null }, { "id": "786040", "postDate": "03/25/2020 15:34:02", "content": "I totally agree with @bamps53 ! Please create a private hidden test set.
", "rawMarkdown": "I totally agree with @bamps53 ! Please create a private hidden test set.", "votes": null }, { "id": "786051", "postDate": "03/25/2020 15:42:21", "content": "@philippsinger Hummm, that's true... It seems Kaggle chose spreading TPU than getting better solution🤔
", "rawMarkdown": "philippsinger Hummm, that's true... It seems Kaggle chose spreading TPU than getting better solution🤔", "votes": null }, { "id": "786054", "postDate": "03/25/2020 15:44:23", "content": "I hope it will be re-new the comp!
", "rawMarkdown": "I hope it will be re-new the comp!", "votes": null }, { "id": "786060", "postDate": "03/25/2020 15:47:22", "content": "@bamps53 I was doing the same with Yandex Translate API :) and it was a bit hacky. Is there a nice way to bypass API limitations with Google Translate that you can share? Seems like none of publicly available hacks worked for me.
\n\nI primarily work with Dutch that's why it's also interesting for me, it's not only about backtranslation but also about using better English-pretrained models.
\n\nps. totally agree that it's not the point of this competition, to do it all in English.
", "rawMarkdown": "bamps53 I was doing the same with Yandex Translate API :) and it was a bit hacky. Is there a nice way to bypass API limitations with Google Translate that you can share? Seems like none of publicly available hacks worked for me. \n\nI primarily work with Dutch that's why it's also interesting for me, it's not only about backtranslation but also about using better English-pretrained models. \n\nps. totally agree that it's not the point of this competition, to do it all in English.", "votes": null }, { "id": "786066", "postDate": "03/25/2020 15:52:24", "content": "Wow, now you're on top!! btw, thanks for sharing great kernels!
", "rawMarkdown": "Wow, now you're on top!! btw, thanks for sharing great kernels!", "votes": null }, { "id": "786068", "postDate": "03/25/2020 15:54:03", "content": "We don't need TPUs for inference! :)
", "rawMarkdown": "We don't need TPUs for inference! :)", "votes": null }, { "id": "786069", "postDate": "03/25/2020 15:54:19", "content": "I used the data you shared. :) So, thanks to you too :)
", "rawMarkdown": "I used the data you shared. :) So, thanks to you too :)", "votes": null }, { "id": "786078", "postDate": "03/25/2020 16:02:28", "content": "Completely agree with the OP. If the whole point of the competition is to be able to build models that generalize well on an unseen multilingual dataset, it's a good idea to turn off the internet on inference.
", "rawMarkdown": "Completely agree with the OP. If the whole point of the competition is to be able to build models that ***generalize well on an unseen multilingual dataset***, it's a good idea to turn off the internet on inference.", "votes": null }, { "id": "786084", "postDate": "03/25/2020 16:09:51", "content": "@kashnitsky I also looked up machine translation model, but yeah it seems good model is not available publicly. My solution to avoid api limitation is also hacky, I uploaded csv file to google spread sheet and then use GOOGLETRANSLATE function😂 It's not scalable at all.
", "rawMarkdown": "kashnitsky I also looked up machine translation model, but yeah it seems good model is not available publicly. My solution to avoid api limitation is also hacky, I uploaded csv file to google spread sheet and then use [GOOGLETRANSLATE function](https://support.google.com/docs/answer/3093331?hl=en)😂 It's not scalable at all.", "votes": null }, { "id": "786100", "postDate": "03/25/2020 16:19:45", "content": "Yeah, but then TPU inference needs to be disabled, but there is TPU prize that also requests TPU inference. Tricky situation.
", "rawMarkdown": "Yeah, but then TPU inference needs to be disabled, but there is TPU prize that also requests TPU inference. Tricky situation.", "votes": null }, { "id": "786139", "postDate": "03/25/2020 16:50:10", "content": "haha, thanks :) indeed. Though it looks like the simplest possible solution.
", "rawMarkdown": "haha, thanks :) indeed. Though it looks like the simplest possible solution.", "votes": null }, { "id": "786152", "postDate": "03/25/2020 16:59:13", "content": "About translations, we checked with the Jigsaw team. They are interested in seeing if there is something useful that can be done through automatic translations. Their gut feeling however is that it will not be a very useful approach because automatic translations tend to de-toxify.
", "rawMarkdown": "About translations, we checked with the Jigsaw team. They are interested in seeing if there is something useful that can be done through automatic translations. Their gut feeling however is that it will not be a very useful approach because automatic translations tend to de-toxify.", "votes": null }, { "id": "786234", "postDate": "03/25/2020 18:10:42", "content": "I guess that means nothing will be changed about this competition. Sad. Originally, it looked like a really interesting task.
", "rawMarkdown": "I guess that means nothing will be changed about this competition. Sad. Originally, it looked like a really interesting task.", "votes": null }, { "id": "786250", "postDate": "03/25/2020 18:25:27", "content": "Yeah it's sad. People will just translate the test data . This competition loses its particularity compared to the past ones.
", "rawMarkdown": "Yeah it's sad. People will just translate the test data . This competition loses its particularity compared to the past ones.", "votes": null }, { "id": "786256", "postDate": "03/25/2020 18:28:48", "content": "I still hope for some creative solutions. Personally, I would have also loved to have a hidden test dataset with no Internet access and more unseen languages. This would have meant no TPU inference though.
", "rawMarkdown": "I still hope for some creative solutions. Personally, I would have also loved to have a hidden test dataset with no Internet access and more unseen languages. This would have meant no TPU inference though.", "votes": null }, { "id": "786320", "postDate": "03/25/2020 19:53:25", "content": "I totally agree, I thought it's a good opportunity to make use of what I learned from Bengali, but it turned out not like that🙄 A little sad, hopefully what jigsaw team said is true and other creative solution surpass just translation!
", "rawMarkdown": "I totally agree, I thought it's a good opportunity to make use of what I learned from Bengali, but it turned out not like that🙄 A little sad, hopefully what jigsaw team said is true and other creative solution surpass just translation!", "votes": null } ], "comments": [ { "id": 785993, "author_name": "adityaecdrid", "author_url": "", "post_date": "03/25/2020 15:03:35", "content": "The End Of the comp is near :(
", "votes": null, "replies": [ { "id": 786054, "author_name": "bamps53", "author_url": "", "post_date": "03/25/2020 15:44:23", "content": "I hope it will be re-new the comp!
", "votes": null, "replies": [] } ] }, { "id": 786004, "author_name": "bamps53", "author_url": "", "post_date": "03/25/2020 15:09:54", "content": "I hope private test set contains more than 100 languages and will be very long tail one😹
", "votes": null, "replies": [] }, { "id": 786021, "author_name": "philippsinger", "author_url": "", "post_date": "03/25/2020 15:25:25", "content": "As long as Internet access in inference is allowed not even private test set helps.
", "votes": null, "replies": [ { "id": 786051, "author_name": "bamps53", "author_url": "", "post_date": "03/25/2020 15:42:21", "content": "@philippsinger Hummm, that's true... It seems Kaggle chose spreading TPU than getting better solution🤔
", "votes": null, "replies": [] }, { "id": 786068, "author_name": "abhishek", "author_url": "", "post_date": "03/25/2020 15:54:03", "content": "We don't need TPUs for inference! :)
", "votes": null, "replies": [] }, { "id": 786100, "author_name": "philippsinger", "author_url": "", "post_date": "03/25/2020 16:19:45", "content": "Yeah, but then TPU inference needs to be disabled, but there is TPU prize that also requests TPU inference. Tricky situation.
", "votes": null, "replies": [] } ] }, { "id": 786040, "author_name": "abhishek", "author_url": "", "post_date": "03/25/2020 15:34:02", "content": "I totally agree with @bamps53 ! Please create a private hidden test set.
", "votes": null, "replies": [ { "id": 786066, "author_name": "bamps53", "author_url": "", "post_date": "03/25/2020 15:52:24", "content": "Wow, now you're on top!! btw, thanks for sharing great kernels!
", "votes": null, "replies": [] }, { "id": 786069, "author_name": "abhishek", "author_url": "", "post_date": "03/25/2020 15:54:19", "content": "I used the data you shared. :) So, thanks to you too :)
", "votes": null, "replies": [] } ] }, { "id": 786060, "author_name": "kashnitsky", "author_url": "", "post_date": "03/25/2020 15:47:22", "content": "@bamps53 I was doing the same with Yandex Translate API :) and it was a bit hacky. Is there a nice way to bypass API limitations with Google Translate that you can share? Seems like none of publicly available hacks worked for me.
\n\nI primarily work with Dutch that's why it's also interesting for me, it's not only about backtranslation but also about using better English-pretrained models.
\n\nps. totally agree that it's not the point of this competition, to do it all in English.
", "votes": null, "replies": [ { "id": 786084, "author_name": "bamps53", "author_url": "", "post_date": "03/25/2020 16:09:51", "content": "@kashnitsky I also looked up machine translation model, but yeah it seems good model is not available publicly. My solution to avoid api limitation is also hacky, I uploaded csv file to google spread sheet and then use GOOGLETRANSLATE function😂 It's not scalable at all.
", "votes": null, "replies": [] }, { "id": 786139, "author_name": "kashnitsky", "author_url": "", "post_date": "03/25/2020 16:50:10", "content": "haha, thanks :) indeed. Though it looks like the simplest possible solution.
", "votes": null, "replies": [] } ] }, { "id": 786078, "author_name": "phanisrikanth", "author_url": "", "post_date": "03/25/2020 16:02:28", "content": "Completely agree with the OP. If the whole point of the competition is to be able to build models that generalize well on an unseen multilingual dataset, it's a good idea to turn off the internet on inference.
", "votes": null, "replies": [] }, { "id": 786152, "author_name": "mgorner", "author_url": "", "post_date": "03/25/2020 16:59:13", "content": "About translations, we checked with the Jigsaw team. They are interested in seeing if there is something useful that can be done through automatic translations. Their gut feeling however is that it will not be a very useful approach because automatic translations tend to de-toxify.
", "votes": null, "replies": [ { "id": 786234, "author_name": "kashnitsky", "author_url": "", "post_date": "03/25/2020 18:10:42", "content": "I guess that means nothing will be changed about this competition. Sad. Originally, it looked like a really interesting task.
", "votes": null, "replies": [] }, { "id": 786250, "author_name": "serigne", "author_url": "", "post_date": "03/25/2020 18:25:27", "content": "Yeah it's sad. People will just translate the test data . This competition loses its particularity compared to the past ones.
", "votes": null, "replies": [] }, { "id": 786256, "author_name": "philippsinger", "author_url": "", "post_date": "03/25/2020 18:28:48", "content": "I still hope for some creative solutions. Personally, I would have also loved to have a hidden test dataset with no Internet access and more unseen languages. This would have meant no TPU inference though.
", "votes": null, "replies": [] }, { "id": 786320, "author_name": "bamps53", "author_url": "", "post_date": "03/25/2020 19:53:25", "content": "I totally agree, I thought it's a good opportunity to make use of what I learned from Bengali, but it turned out not like that🙄 A little sad, hopefully what jigsaw team said is true and other creative solution surpass just translation!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "785985": "I created [translated test set](https://www.kaggle.com/bamps53/test-en-df) by google translate api and created [inference kernel](https://www.kaggle.com/bamps53/inference-with-translated-test-set?scriptVersionId=30818481) using this. (This is forked from [abhishek's kernel](https://www.kaggle.com/abhishek/inference-of-bert-tpu-model-ml-w-validation). Please upvote original one too!)\n\nWhat I want to say here is it's better to create private test set. This is not interesting approarch and I assume it's not what host wanted.\n\nIt will be much more interesting if there is an unknown test set and if it contains unseen language, isn't it??\n\n[UPDATE] And also please turn off internet access when submission. It doesn't prevent using TPU when training.", "785993": "The End Of the comp is near :(", "786004": "I hope private test set contains more than 100 languages and will be very long tail one😹", "786021": "As long as Internet access in inference is allowed not even private test set helps.", "786040": "I totally agree with @bamps53 ! Please create a private hidden test set.", "786051": "philippsinger Hummm, that's true... It seems Kaggle chose spreading TPU than getting better solution🤔", "786054": "I hope it will be re-new the comp!", "786060": "bamps53 I was doing the same with Yandex Translate API :) and it was a bit hacky. Is there a nice way to bypass API limitations with Google Translate that you can share? Seems like none of publicly available hacks worked for me. \n\nI primarily work with Dutch that's why it's also interesting for me, it's not only about backtranslation but also about using better English-pretrained models. \n\nps. totally agree that it's not the point of this competition, to do it all in English.", "786066": "Wow, now you're on top!! btw, thanks for sharing great kernels!", "786068": "We don't need TPUs for inference! :)", "786069": "I used the data you shared. :) So, thanks to you too :)", "786078": "Completely agree with the OP. If the whole point of the competition is to be able to build models that ***generalize well on an unseen multilingual dataset***, it's a good idea to turn off the internet on inference.", "786084": "kashnitsky I also looked up machine translation model, but yeah it seems good model is not available publicly. My solution to avoid api limitation is also hacky, I uploaded csv file to google spread sheet and then use [GOOGLETRANSLATE function](https://support.google.com/docs/answer/3093331?hl=en)😂 It's not scalable at all.", "786100": "Yeah, but then TPU inference needs to be disabled, but there is TPU prize that also requests TPU inference. Tricky situation.", "786139": "haha, thanks :) indeed. Though it looks like the simplest possible solution.", "786152": "About translations, we checked with the Jigsaw team. They are interested in seeing if there is something useful that can be done through automatic translations. Their gut feeling however is that it will not be a very useful approach because automatic translations tend to de-toxify.", "786234": "I guess that means nothing will be changed about this competition. Sad. Originally, it looked like a really interesting task.", "786250": "Yeah it's sad. People will just translate the test data . This competition loses its particularity compared to the past ones.", "786256": "I still hope for some creative solutions. Personally, I would have also loved to have a hidden test dataset with no Internet access and more unseen languages. This would have meant no TPU inference though.", "786320": "I totally agree, I thought it's a good opportunity to make use of what I learned from Bengali, but it turned out not like that🙄 A little sad, hopefully what jigsaw team said is true and other creative solution surpass just translation!" }, "source": "meta" }