I tried using a generative model such as conditional GAN to augment the dataset. I used the Wasserstein GAN architecture as basis, which I implemented from scratch.
\nI'm not sure that for this task it really gives a significant increase in metrics, but I think someone might find this technique interesting for more complex tasks.
I think it's worth trying to train more the generative model and work with CNN architectures, or using another approaches with pre-trained model to embeddings and e.g. GBDT models on it.
\nGood luck everyone!
\nP.s. I will be glad to receive any feedback
", "messages": [ { "id": 2995898, "postDate": "2024-09-22T18:13:42.633Z", "content": "With part of score give Data augmentation?
", "rawMarkdown": "With part of score give Data augmentation?", "votes": 1, "replies": [ { "id": 2995925, "postDate": "2024-09-22T18:56:11.147Z", "content": "If you mean how much augmentation has improved the performance of the model, then I can't say an honest assessment. This is an interesting question, as it seems to me it is worth trying the same CNN model on a default train part for fixed N epoches, try use standard data augmentation by torch.transforms and on extended train part (but i am not sure for choose count of epoches. For example if we increase our train part for two times, i think we need to reduce the number of epochs by half - N // 2).
\nIn any case, the dataset is very simple, and even a simple CNN model from a notebook gives a good result. Rather, I was just interested in such a technique. It is worth checking it for more complex data.
", "rawMarkdown": "If you mean how much augmentation has improved the performance of the model, then I can't say an honest assessment. This is an interesting question, as it seems to me it is worth trying the same CNN model on a default train part for fixed N epoches, try use standard data augmentation by torch.transforms and on extended train part (but i am not sure for choose count of epoches. For example if we increase our train part for two times, i think we need to reduce the number of epochs by half - N // 2). \n\nIn any case, the dataset is very simple, and even a simple CNN model from a notebook gives a good result. Rather, I was just interested in such a technique. It is worth checking it for more complex data." } ] }, { "id": 2989855, "postDate": "2024-09-15T15:56:20.363Z", "content": "I tried using a generative model such as conditional GAN to augment the dataset. I used the Wasserstein GAN architecture as basis, which I implemented from scratch.
\nI'm not sure that for this task it really gives a significant increase in metrics, but I think someone might find this technique interesting for more complex tasks.
I think it's worth trying to train more the generative model and work with CNN architectures, or using another approaches with pre-trained model to embeddings and e.g. GBDT models on it.
\nGood luck everyone!
\nP.s. I will be glad to receive any feedback
", "rawMarkdown": "I tried using a generative model such as conditional GAN to augment the dataset. I used the Wasserstein GAN architecture as basis, which I implemented from scratch. \nI'm not sure that for this task it really gives a significant increase in metrics, but I think someone might find this technique interesting for more complex tasks.\n\n[Notebook](https://www.kaggle.com/code/latypovramil/data-augmentation-using-cgan-pytorch)\n\nI think it's worth trying to train more the generative model and work with CNN architectures, or using another approaches with pre-trained model to embeddings and e.g. GBDT models on it. \n\nGood luck everyone!\n\nP.s. I will be glad to receive any feedback", "votes": 1 } ], "comments": [ { "id": 2995898, "author_name": "", "author_url": "", "post_date": "2024-09-22T18:13:42.633000", "content": "With part of score give Data augmentation?
", "votes": 1, "replies": [ { "id": 2995925, "author_name": "Latypov Ramil", "author_url": "", "post_date": "2024-09-22T18:56:11.147000", "content": "If you mean how much augmentation has improved the performance of the model, then I can't say an honest assessment. This is an interesting question, as it seems to me it is worth trying the same CNN model on a default train part for fixed N epoches, try use standard data augmentation by torch.transforms and on extended train part (but i am not sure for choose count of epoches. For example if we increase our train part for two times, i think we need to reduce the number of epochs by half - N // 2).
\nIn any case, the dataset is very simple, and even a simple CNN model from a notebook gives a good result. Rather, I was just interested in such a technique. It is worth checking it for more complex data.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "2995898": "With part of score give Data augmentation?", "2989855": "I tried using a generative model such as conditional GAN to augment the dataset. I used the Wasserstein GAN architecture as basis, which I implemented from scratch. \nI'm not sure that for this task it really gives a significant increase in metrics, but I think someone might find this technique interesting for more complex tasks.\n\n[Notebook](https://www.kaggle.com/code/latypovramil/data-augmentation-using-cgan-pytorch)\n\nI think it's worth trying to train more the generative model and work with CNN architectures, or using another approaches with pre-trained model to embeddings and e.g. GBDT models on it. \n\nGood luck everyone!\n\nP.s. I will be glad to receive any feedback" } }