Hey, this is kind of my first kaggle competition, so sorry in advance if I am asking trivial question:\nTraining a GAN in 9h is quite challenging (if you want to use a little deeper stuff than standard DCGAN), so I was wondering if there is some sort of way to save weights and reuse them when committing or other tricks how I can use a model that was trained longer than 9h?\nIf the 9h limit is part of the challenge requirements that is of course fine with me.
", "messages": [ { "id": "567407", "postDate": "07/03/2019 13:23:43", "content": "Hey, this is kind of my first kaggle competition, so sorry in advance if I am asking trivial question:\nTraining a GAN in 9h is quite challenging (if you want to use a little deeper stuff than standard DCGAN), so I was wondering if there is some sort of way to save weights and reuse them when committing or other tricks how I can use a model that was trained longer than 9h?\nIf the 9h limit is part of the challenge requirements that is of course fine with me.
", "rawMarkdown": "Hey, this is kind of my first kaggle competition, so sorry in advance if I am asking trivial question:\nTraining a GAN in 9h is quite challenging (if you want to use a little deeper stuff than standard DCGAN), so I was wondering if there is some sort of way to save weights and reuse them when committing or other tricks how I can use a model that was trained longer than 9h?\nIf the 9h limit is part of the challenge requirements that is of course fine with me.", "votes": null }, { "id": "567429", "postDate": "07/03/2019 14:01:07", "content": "Make sure that you turn GPU on. (It is off by default). If you use GPU and an efficient GAN design, you can generate quality images in under 9 hours.
\n\nUnfortunately you can not save and load anything into your kernel for this competition. (In general, you can in Kaggle kernels).
", "rawMarkdown": "Make sure that you turn GPU on. (It is off by default). If you use GPU and an efficient GAN design, you can generate quality images in under 9 hours.\n\nUnfortunately you can not save and load anything into your kernel for this competition. (In general, you can in Kaggle kernels).", "votes": null }, { "id": "567604", "postDate": "07/03/2019 18:37:41", "content": "Thanks @cdeotte for the quick reply!\nYeah, GPU was on, but still my WGAN-GP was quite slow. Probably my implementation is quite bad so lots of room to improve. That's what makes a challenge a challenge, I guess.
", "rawMarkdown": "Thanks @cdeotte for the quick reply!\nYeah, GPU was on, but still my WGAN-GP was quite slow. Probably my implementation is quite bad so lots of room to improve. That's what makes a challenge a challenge, I guess.", "votes": null }, { "id": "567647", "postDate": "07/03/2019 19:48:32", "content": "It provides the challenge to create models that create fast results. \nSo thankfully, in the end not just someone with the deepest wallet wins the competition.
", "rawMarkdown": "It provides the challenge to create models that create fast results. \nSo thankfully, in the end not just someone with the deepest wallet wins the competition.", "votes": null } ], "comments": [ { "id": 567429, "author_name": "cdeotte", "author_url": "", "post_date": "07/03/2019 14:01:07", "content": "Make sure that you turn GPU on. (It is off by default). If you use GPU and an efficient GAN design, you can generate quality images in under 9 hours.
\n\nUnfortunately you can not save and load anything into your kernel for this competition. (In general, you can in Kaggle kernels).
", "votes": null, "replies": [] }, { "id": 567604, "author_name": "flocfschulz", "author_url": "", "post_date": "07/03/2019 18:37:41", "content": "Thanks @cdeotte for the quick reply!\nYeah, GPU was on, but still my WGAN-GP was quite slow. Probably my implementation is quite bad so lots of room to improve. That's what makes a challenge a challenge, I guess.
", "votes": null, "replies": [] }, { "id": 567647, "author_name": "ilu000", "author_url": "", "post_date": "07/03/2019 19:48:32", "content": "It provides the challenge to create models that create fast results. \nSo thankfully, in the end not just someone with the deepest wallet wins the competition.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "567407": "Hey, this is kind of my first kaggle competition, so sorry in advance if I am asking trivial question:\nTraining a GAN in 9h is quite challenging (if you want to use a little deeper stuff than standard DCGAN), so I was wondering if there is some sort of way to save weights and reuse them when committing or other tricks how I can use a model that was trained longer than 9h?\nIf the 9h limit is part of the challenge requirements that is of course fine with me.", "567429": "Make sure that you turn GPU on. (It is off by default). If you use GPU and an efficient GAN design, you can generate quality images in under 9 hours.\n\nUnfortunately you can not save and load anything into your kernel for this competition. (In general, you can in Kaggle kernels).", "567604": "Thanks @cdeotte for the quick reply!\nYeah, GPU was on, but still my WGAN-GP was quite slow. Probably my implementation is quite bad so lots of room to improve. That's what makes a challenge a challenge, I guess.", "567647": "It provides the challenge to create models that create fast results. \nSo thankfully, in the end not just someone with the deepest wallet wins the competition." }, "source": "meta" }