Hi,
\n\nI have seen in other posts people talking about 20 minute models. My models have been taking 2-3 hours to commit. People who have 20 minute models, what are you using ? What libraries (Keras? Fast.ai? Pytorch ?) What image sets (Resnet ?, Densenet? Efficientnet?)
\n\nIf you can point to an open Kernel, that would be great, but if you just want to give general ideas that's great too. \nThe most frustrating part of this competition has been the length of time it takes to run each model.
", "messages": [ { "id": "619356", "postDate": "09/06/2019 05:47:36", "content": "Hi,
\n\nI have seen in other posts people talking about 20 minute models. My models have been taking 2-3 hours to commit. People who have 20 minute models, what are you using ? What libraries (Keras? Fast.ai? Pytorch ?) What image sets (Resnet ?, Densenet? Efficientnet?)
\n\nIf you can point to an open Kernel, that would be great, but if you just want to give general ideas that's great too. \nThe most frustrating part of this competition has been the length of time it takes to run each model.
", "rawMarkdown": "Hi, \n\nI have seen in other posts people talking about 20 minute models. My models have been taking 2-3 hours to commit. People who have 20 minute models, what are you using ? What libraries (Keras? Fast.ai? Pytorch ?) What image sets (Resnet ?, Densenet? Efficientnet?)\n\nIf you can point to an open Kernel, that would be great, but if you just want to give general ideas that's great too. \nThe most frustrating part of this competition has been the length of time it takes to run each model.", "votes": null }, { "id": "619358", "postDate": "09/06/2019 05:50:56", "content": "You only need to put inference part in your submitted version. I also believe lots of people do the deterministic preprocessing only once to including more models.
", "rawMarkdown": "You only need to put inference part in your submitted version. I also believe lots of people do the deterministic preprocessing only once to including more models.", "votes": null }, { "id": "619361", "postDate": "09/06/2019 05:53:15", "content": "My all Efficient Nets have 1-2 minutes of commit time and 25-30 minutes of submission time. They are trained on their default image sizes using Fastai.
", "rawMarkdown": "My all Efficient Nets have 1-2 minutes of commit time and 25-30 minutes of submission time. They are trained on their default image sizes using Fastai.", "votes": null }, { "id": "619404", "postDate": "09/06/2019 06:56:55", "content": "I don't really understand what you mean. Could you explain that a bit more ? I would appreciate it.
", "rawMarkdown": "I don't really understand what you mean. Could you explain that a bit more ? I would appreciate it.", "votes": null }, { "id": "619406", "postDate": "09/06/2019 06:57:20", "content": "When the competition is over, I would love to look at some of your kernels.
", "rawMarkdown": "When the competition is over, I would love to look at some of your kernels.", "votes": null }, { "id": "619426", "postDate": "09/06/2019 07:35:56", "content": "In the submitted model, you only need to load your trained model and do the prediction.
\n\nAs for the preprocessing part, you can refer to my kernel.
", "rawMarkdown": "In the submitted model, you only need to load your trained model and do the prediction. \n\nAs for the preprocessing part, you can refer to my [kernel](https://www.kaggle.com/naivelamb/process-test-images-parallelly).", "votes": null }, { "id": "619583", "postDate": "09/06/2019 10:15:26", "content": "Take a look at this kernel, it runs under 3 minutes while commiting, will surely give you some idea.
\n\nhttps://www.kaggle.com/drhabib/starter-kernel-for-0-79
", "rawMarkdown": "Take a look at this kernel, it runs under 3 minutes while commiting, will surely give you some idea.\n\nhttps://www.kaggle.com/drhabib/starter-kernel-for-0-79", "votes": null }, { "id": "619710", "postDate": "09/06/2019 13:51:59", "content": "I have been wondering opposite of this. My models take around 4 minute to commit and 15 minutes to submit. They are just inference models. How are yours taking so long if it is just an inference kernel? Is that because of TTA?
", "rawMarkdown": "I have been wondering opposite of this. My models take around 4 minute to commit and 15 minutes to submit. They are just inference models. How are yours taking so long if it is just an inference kernel? Is that because of TTA?", "votes": null }, { "id": "619820", "postDate": "09/06/2019 16:02:07", "content": "They aren't just inference models. I'm pretty new (I did a competition a few years ago and then didn't until this one). So my models do all the learning in the kernel. \nMy main purpose for this competition is to become more familiar with fast.ai, so I have been using public kernels with fast.ai, so I can generally understand the kernel and then tweaking transformations and learning rates and such. Downloading the datasets and running models locally and then uploading is rather prohibitive because I have slow internet speeds. (Soon I will get optic fiber, but they haven't installed it yet). So I have been doing everything in the kernel. \nThis is not my best model, but it's about average for the time it takes each epoch to run. 8-9 minutes each. https://www.kaggle.com/chrisfs/vgg19-bn-base-model?scriptVersionId=20142317
", "rawMarkdown": "They aren't just inference models. I'm pretty new (I did a competition a few years ago and then didn't until this one). So my models do all the learning in the kernel. \nMy main purpose for this competition is to become more familiar with fast.ai, so I have been using public kernels with fast.ai, so I can generally understand the kernel and then tweaking transformations and learning rates and such. Downloading the datasets and running models locally and then uploading is rather prohibitive because I have slow internet speeds. (Soon I will get optic fiber, but they haven't installed it yet). So I have been doing everything in the kernel. \nThis is not my best model, but it's about average for the time it takes each epoch to run. 8-9 minutes each. https://www.kaggle.com/chrisfs/vgg19-bn-base-model?scriptVersionId=20142317", "votes": null }, { "id": "620216", "postDate": "09/07/2019 06:53:09", "content": "Try to use at least resnets. Fastai has resnets built-in, if i remember correctly. If you could, try efficientnets, effnetb2 is 29 mb, b3 is about 42 mb. Also they are a lot faster than vgg.
", "rawMarkdown": "Try to use at least resnets. Fastai has resnets built-in, if i remember correctly. If you could, try efficientnets, effnetb2 is 29 mb, b3 is about 42 mb. Also they are a lot faster than vgg.", "votes": null }, { "id": "620664", "postDate": "09/07/2019 20:01:18", "content": "\n\n\nIs that because of TTA?
\n
Ensembling several models and using TTA for each of them 😁. Mine took around 35 mins to run and 7h once I submit.
", "rawMarkdown": "> Is that because of TTA?\n\nEnsembling several models and using TTA for each of them 😁. Mine took around 35 mins to run and 7h once I submit.", "votes": null } ], "comments": [ { "id": 619358, "author_name": "naivelamb", "author_url": "", "post_date": "09/06/2019 05:50:56", "content": "You only need to put inference part in your submitted version. I also believe lots of people do the deterministic preprocessing only once to including more models.
", "votes": null, "replies": [ { "id": 619404, "author_name": "chrisfs", "author_url": "", "post_date": "09/06/2019 06:56:55", "content": "I don't really understand what you mean. Could you explain that a bit more ? I would appreciate it.
", "votes": null, "replies": [] }, { "id": 619426, "author_name": "naivelamb", "author_url": "", "post_date": "09/06/2019 07:35:56", "content": "In the submitted model, you only need to load your trained model and do the prediction.
\n\nAs for the preprocessing part, you can refer to my kernel.
", "votes": null, "replies": [] } ] }, { "id": 619361, "author_name": "salilm23", "author_url": "", "post_date": "09/06/2019 05:53:15", "content": "My all Efficient Nets have 1-2 minutes of commit time and 25-30 minutes of submission time. They are trained on their default image sizes using Fastai.
", "votes": null, "replies": [ { "id": 619406, "author_name": "chrisfs", "author_url": "", "post_date": "09/06/2019 06:57:20", "content": "When the competition is over, I would love to look at some of your kernels.
", "votes": null, "replies": [] }, { "id": 619583, "author_name": "salilm23", "author_url": "", "post_date": "09/06/2019 10:15:26", "content": "Take a look at this kernel, it runs under 3 minutes while commiting, will surely give you some idea.
\n\nhttps://www.kaggle.com/drhabib/starter-kernel-for-0-79
", "votes": null, "replies": [] } ] }, { "id": 619710, "author_name": "mearex", "author_url": "", "post_date": "09/06/2019 13:51:59", "content": "I have been wondering opposite of this. My models take around 4 minute to commit and 15 minutes to submit. They are just inference models. How are yours taking so long if it is just an inference kernel? Is that because of TTA?
", "votes": null, "replies": [ { "id": 619820, "author_name": "chrisfs", "author_url": "", "post_date": "09/06/2019 16:02:07", "content": "They aren't just inference models. I'm pretty new (I did a competition a few years ago and then didn't until this one). So my models do all the learning in the kernel. \nMy main purpose for this competition is to become more familiar with fast.ai, so I have been using public kernels with fast.ai, so I can generally understand the kernel and then tweaking transformations and learning rates and such. Downloading the datasets and running models locally and then uploading is rather prohibitive because I have slow internet speeds. (Soon I will get optic fiber, but they haven't installed it yet). So I have been doing everything in the kernel. \nThis is not my best model, but it's about average for the time it takes each epoch to run. 8-9 minutes each. https://www.kaggle.com/chrisfs/vgg19-bn-base-model?scriptVersionId=20142317
", "votes": null, "replies": [] }, { "id": 620216, "author_name": "mearex", "author_url": "", "post_date": "09/07/2019 06:53:09", "content": "Try to use at least resnets. Fastai has resnets built-in, if i remember correctly. If you could, try efficientnets, effnetb2 is 29 mb, b3 is about 42 mb. Also they are a lot faster than vgg.
", "votes": null, "replies": [] }, { "id": 620664, "author_name": "raimonds1993", "author_url": "", "post_date": "09/07/2019 20:01:18", "content": "\n\n\nIs that because of TTA?
\n
Ensembling several models and using TTA for each of them 😁. Mine took around 35 mins to run and 7h once I submit.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "619356": "Hi, \n\nI have seen in other posts people talking about 20 minute models. My models have been taking 2-3 hours to commit. People who have 20 minute models, what are you using ? What libraries (Keras? Fast.ai? Pytorch ?) What image sets (Resnet ?, Densenet? Efficientnet?)\n\nIf you can point to an open Kernel, that would be great, but if you just want to give general ideas that's great too. \nThe most frustrating part of this competition has been the length of time it takes to run each model.", "619358": "You only need to put inference part in your submitted version. I also believe lots of people do the deterministic preprocessing only once to including more models.", "619361": "My all Efficient Nets have 1-2 minutes of commit time and 25-30 minutes of submission time. They are trained on their default image sizes using Fastai.", "619404": "I don't really understand what you mean. Could you explain that a bit more ? I would appreciate it.", "619406": "When the competition is over, I would love to look at some of your kernels.", "619426": "In the submitted model, you only need to load your trained model and do the prediction. \n\nAs for the preprocessing part, you can refer to my [kernel](https://www.kaggle.com/naivelamb/process-test-images-parallelly).", "619583": "Take a look at this kernel, it runs under 3 minutes while commiting, will surely give you some idea.\n\nhttps://www.kaggle.com/drhabib/starter-kernel-for-0-79", "619710": "I have been wondering opposite of this. My models take around 4 minute to commit and 15 minutes to submit. They are just inference models. How are yours taking so long if it is just an inference kernel? Is that because of TTA?", "619820": "They aren't just inference models. I'm pretty new (I did a competition a few years ago and then didn't until this one). So my models do all the learning in the kernel. \nMy main purpose for this competition is to become more familiar with fast.ai, so I have been using public kernels with fast.ai, so I can generally understand the kernel and then tweaking transformations and learning rates and such. Downloading the datasets and running models locally and then uploading is rather prohibitive because I have slow internet speeds. (Soon I will get optic fiber, but they haven't installed it yet). So I have been doing everything in the kernel. \nThis is not my best model, but it's about average for the time it takes each epoch to run. 8-9 minutes each. https://www.kaggle.com/chrisfs/vgg19-bn-base-model?scriptVersionId=20142317", "620216": "Try to use at least resnets. Fastai has resnets built-in, if i remember correctly. If you could, try efficientnets, effnetb2 is 29 mb, b3 is about 42 mb. Also they are a lot faster than vgg.", "620664": "> Is that because of TTA?\n\nEnsembling several models and using TTA for each of them 😁. Mine took around 35 mins to run and 7h once I submit." }, "source": "meta" }