My model works very well on layout xla dataset, but has no effect at all on layout nlp. What should I do? Can you tell me some models that handle layout nlp dataset? Thank you so much！
", "messages": [ { "id": 2433284, "postDate": "2023-09-11T13:15:11.117Z", "content": "My model works very well on layout xla dataset, but has no effect at all on layout nlp. What should I do? Can you tell me some models that handle layout nlp dataset? Thank you so much！
", "rawMarkdown": "My model works very well on layout xla dataset, but has no effect at all on layout nlp. What should I do? Can you tell me some models that handle layout nlp dataset? Thank you so much！", "votes": 8 }, { "id": 2454517, "postDate": "2023-09-24T21:48:00.920Z", "content": "I also have very poor performance across training, validation and testing on the NLP data, which is surprising since very similar results between NLP and XLA are reported in the paper. @mangpophothilimthana is there any chance that the data might be corrupted?
", "rawMarkdown": "I also have very poor performance across training, validation and testing on the NLP data, which is surprising since very similar results between NLP and XLA are reported in the paper. @mangpophothilimthana is there any chance that the data might be corrupted?", "votes": 5, "replies": [ { "id": 2454541, "postDate": "2023-09-24T22:47:11.750Z", "content": "I was thinking the same thing… In the paper they show similar results between nlp and xla default and even suggest that xla random is harder to train than nlp default which in reality is not the case.
", "rawMarkdown": "I was thinking the same thing... In the paper they show similar results between nlp and xla default and even suggest that xla random is harder to train than nlp default which in reality is not the case. ", "replies": [ { "id": 2454547, "postDate": "2023-09-24T22:54:47.850Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 2458961, "postDate": "2023-09-27T23:43:34.910Z", "content": "Thank you so much for raising this concern! We checked the data, and some parts of the data were indeed corrupted. We have regenerated the data, and tried our best to verified that the new data is correct. Please see this post about the new data. Sorry for the inconvenient this may have caused.
", "rawMarkdown": "Thank you so much for raising this concern! We checked the data, and some parts of the data were indeed corrupted. We have regenerated the data, and tried our best to verified that the new data is correct. Please see this [post](https://www.kaggle.com/competitions/predict-ai-model-runtime/discussion/443581) about the new data. Sorry for the inconvenient this may have caused.", "votes": 3 } ] }, { "id": 2438162, "postDate": "2023-09-14T05:06:10.550Z", "content": "If you are using one model for both NLP and XLA datasets, then this is a good reason to use different models for them.
", "rawMarkdown": "If you are using one model for both NLP and XLA datasets, then this is a good reason to use different models for them. ", "replies": [ { "id": 2440430, "postDate": "2023-09-15T14:00:36.363Z", "content": "Thanks I will try
", "rawMarkdown": "Thanks I will try" } ] }, { "id": 2433928, "postDate": "2023-09-12T01:19:25.600Z", "content": "Same problem. I used SAGE and GCN, but they didn't work well in NLP. I also tried to train the layout files together but the effect was worse
", "rawMarkdown": "Same problem. I used SAGE and GCN, but they didn't work well in NLP. I also tried to train the layout files together but the effect was worse\n\n\n", "replies": [ { "id": 2433949, "postDate": "2023-09-12T02:09:30.387Z", "content": "Hello, have you ever tried the github code provided by the organizer? I tried it but it was difficult to run on kaggle.
", "rawMarkdown": "Hello, have you ever tried the github code provided by the organizer? I tried it but it was difficult to run on kaggle.", "replies": [ { "id": 2433959, "postDate": "2023-09-12T02:18:55.257Z", "content": "No, I refer to the kaggle code and improve it. I'll try the github code
", "rawMarkdown": "No, I refer to the kaggle code and improve it. I'll try the github code" } ] }, { "id": 2440032, "postDate": "2023-09-15T08:22:07.950Z", "content": "Did you implement GST in your xla/nlp model?
", "rawMarkdown": "Did you implement GST in your xla/nlp model?", "replies": [ { "id": 2440432, "postDate": "2023-09-15T14:01:28.193Z", "content": "I haven't use GST. I will try later
", "rawMarkdown": "I haven't use GST. I will try later", "votes": 1 } ] } ] } ], "comments": [ { "id": 2454517, "author_name": "Peiyuan Liao", "author_url": "", "post_date": "2023-09-24T21:48:00.920000", "content": "I also have very poor performance across training, validation and testing on the NLP data, which is surprising since very similar results between NLP and XLA are reported in the paper. @mangpophothilimthana is there any chance that the data might be corrupted?
", "votes": 5, "replies": [ { "id": 2454541, "author_name": "Amit Aharoni", "author_url": "", "post_date": "2023-09-24T22:47:11.750000", "content": "I was thinking the same thing… In the paper they show similar results between nlp and xla default and even suggest that xla random is harder to train than nlp default which in reality is not the case.
", "votes": 0, "replies": [ { "id": 2454547, "author_name": "", "author_url": "", "post_date": "2023-09-24T22:54:47.850000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 2458961, "author_name": "Mangpo Phothilimthana", "author_url": "", "post_date": "2023-09-27T23:43:34.910000", "content": "Thank you so much for raising this concern! We checked the data, and some parts of the data were indeed corrupted. We have regenerated the data, and tried our best to verified that the new data is correct. Please see this post about the new data. Sorry for the inconvenient this may have caused.
", "votes": 3, "replies": [] } ] }, { "id": 2438162, "author_name": "PassengerC07", "author_url": "", "post_date": "2023-09-14T05:06:10.550000", "content": "If you are using one model for both NLP and XLA datasets, then this is a good reason to use different models for them.
", "votes": 0, "replies": [ { "id": 2440430, "author_name": "ljjsxx", "author_url": "", "post_date": "2023-09-15T14:00:36.363000", "content": "Thanks I will try
", "votes": 0, "replies": [] } ] }, { "id": 2433928, "author_name": "MiHu", "author_url": "", "post_date": "2023-09-12T01:19:25.600000", "content": "Same problem. I used SAGE and GCN, but they didn't work well in NLP. I also tried to train the layout files together but the effect was worse
", "votes": 0, "replies": [ { "id": 2433949, "author_name": "ljjsxx", "author_url": "", "post_date": "2023-09-12T02:09:30.387000", "content": "Hello, have you ever tried the github code provided by the organizer? I tried it but it was difficult to run on kaggle.
", "votes": 0, "replies": [ { "id": 2433959, "author_name": "MiHu", "author_url": "", "post_date": "2023-09-12T02:18:55.257000", "content": "No, I refer to the kaggle code and improve it. I'll try the github code
", "votes": 0, "replies": [] } ] }, { "id": 2440032, "author_name": "Amit Aharoni", "author_url": "", "post_date": "2023-09-15T08:22:07.950000", "content": "Did you implement GST in your xla/nlp model?
", "votes": 0, "replies": [ { "id": 2440432, "author_name": "ljjsxx", "author_url": "", "post_date": "2023-09-15T14:01:28.193000", "content": "I haven't use GST. I will try later
", "votes": 1, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2433284": "My model works very well on layout xla dataset, but has no effect at all on layout nlp. What should I do? Can you tell me some models that handle layout nlp dataset? Thank you so much！", "2454517": "I also have very poor performance across training, validation and testing on the NLP data, which is surprising since very similar results between NLP and XLA are reported in the paper. @mangpophothilimthana is there any chance that the data might be corrupted?", "2438162": "If you are using one model for both NLP and XLA datasets, then this is a good reason to use different models for them. ", "2433928": "Same problem. I used SAGE and GCN, but they didn't work well in NLP. I also tried to train the layout files together but the effect was worse\n\n\n" } }