Hi, sorry for the silly question. I have never worked with image datasets before.
\n\nMy question to those who have worked with this kind of data... will it be useful (good learning experience) to participate in this competition with 16GB system?
\n\nI am assuming that deep learning will play a big role in this competition, and with my current system specification it will be difficult to train deep learning models on this data. Any thoughts?
", "messages": [ { "id": "207829", "postDate": "07/27/2017 15:55:46", "content": "Hi, sorry for the silly question. I have never worked with image datasets before.
\n\nMy question to those who have worked with this kind of data... will it be useful (good learning experience) to participate in this competition with 16GB system?
\n\nI am assuming that deep learning will play a big role in this competition, and with my current system specification it will be difficult to train deep learning models on this data. Any thoughts?
", "rawMarkdown": "Hi, sorry for the silly question. I have never worked with image datasets before. \n\nMy question to those who have worked with this kind of data... **will it be useful (good learning experience) to participate in this competition with 16GB system?**\n\nI am assuming that deep learning will play a big role in this competition, and with my current system specification it will be difficult to train deep learning models on this data. Any thoughts?", "votes": null }, { "id": "207831", "postDate": "07/27/2017 15:58:14", "content": "16GB RAM will be fine, but i guess to use deep learning you should probably make use of GPU computing.
", "rawMarkdown": "16GB RAM will be fine, but i guess to use deep learning you should probably make use of GPU computing.", "votes": null }, { "id": "207836", "postDate": "07/27/2017 16:06:56", "content": "Hi Anubhav,
\n\nI am no expert by any means but most vision related problems today including image segmentation use convolutional neural nets to get state of the art results. I share your frustration in the inability to train such models on large datasets without having to chose a very small batch size (which in turns increases my training time) on my personal machine (which is 32GB mind you).
\n\nI have found Google Compute Engine very handy for such tasks. You get $300 worth of credits on free signup and can get access to Tesla K80 cards on your instances for free by filling up a simple request to increase quota.
\n\nIf you would like some help in setting up your work environment, the people organizing the CS231n: Convolutional Neural Networks for Visual Recognition over at Stanford have got your back. Hope this helps.
\n\nBest of luck for the competition.
\n\nBest Regards,
\nShashank
Hi Shashank,
\n\nThanks for the suggestion. I've already set up Google cloud using those instructions. You mentioned above that we can get access to Tesla K80 cards. Can you provide some details?
", "rawMarkdown": "Hi Shashank,\n\nThanks for the suggestion. I've already set up Google cloud using those instructions. You mentioned above that we can get access to Tesla K80 cards. Can you provide some details?", "votes": null }, { "id": "207870", "postDate": "07/27/2017 18:22:17", "content": "Once you have a Google Cloud account, you can access the IAM & Admin -> Quotas from the left panel. You can then increase the quota for GPU for a particular project from zero to as much as 4.
\n\nPlease keep in mind thought that Google will ask for a small fee ($35) for increasing beyond 2 GPUs just to make sure you are genuine in your request. This fee will be added to your credits. You can increase upto 2 GPUs without having to pay anything.
", "rawMarkdown": "Once you have a Google Cloud account, you can access the [IAM & Admin -> Quotas][1] from the left panel. You can then increase the quota for GPU for a particular project from zero to as much as 4.\n\nPlease keep in mind thought that Google will ask for a small fee ($35) for increasing beyond 2 GPUs just to make sure you are genuine in your request. This fee will be added to your credits. You can increase upto 2 GPUs without having to pay anything.\n\n\n [1]: https://console.cloud.google.com/projectselector/iam-admin/quotas", "votes": null }, { "id": "208021", "postDate": "07/28/2017 07:59:47", "content": "There is also Google Cloud ML, which is specifically for running TensorFlow jobs.
", "rawMarkdown": "There is also Google Cloud ML, which is specifically for running TensorFlow jobs.", "votes": null }, { "id": "208091", "postDate": "07/28/2017 14:16:35", "content": "If you have a decent Nvidia GPU then installing tensorflow-gpu to your python environment might be a very good option https://www.tensorflow.org/install/ and setting up Keras on top of it, which is really easy to use if you have used scikit-learn before.\nhttps://keras.io/#installation
Here is a very good guide on building a system for Deep Learning: http://timdettmers.com/2015/03/09/deep-learning-hardware-guide/ \n
And a guide on selecting GPUs: http://timdettmers.com/2017/04/09/which-gpu-for-deep-learning/
Also getting a nice system and training NNs by yourself will probably be cheaper than using Cloud services, if you plan to do it for a while.
", "rawMarkdown": "If you have a decent Nvidia GPU then installing `tensorflow-gpu` to your python environment might be a very good option https://www.tensorflow.org/install/ and setting up Keras on top of it, which is really easy to use if you have used scikit-learn before.\nhttps://keras.io/#installation\n\n\nHere is a very good guide on building a system for Deep Learning: http://timdettmers.com/2015/03/09/deep-learning-hardware-guide/ \nYou can also try retraining an existing suitable Tensorflow model.
", "rawMarkdown": "You can also try retraining an existing suitable Tensorflow model.", "votes": null } ], "comments": [ { "id": 207831, "author_name": "crailtap", "author_url": "", "post_date": "07/27/2017 15:58:14", "content": "16GB RAM will be fine, but i guess to use deep learning you should probably make use of GPU computing.
", "votes": null, "replies": [] }, { "id": 207836, "author_name": "sshekhar", "author_url": "", "post_date": "07/27/2017 16:06:56", "content": "Hi Anubhav,
\n\nI am no expert by any means but most vision related problems today including image segmentation use convolutional neural nets to get state of the art results. I share your frustration in the inability to train such models on large datasets without having to chose a very small batch size (which in turns increases my training time) on my personal machine (which is 32GB mind you).
\n\nI have found Google Compute Engine very handy for such tasks. You get $300 worth of credits on free signup and can get access to Tesla K80 cards on your instances for free by filling up a simple request to increase quota.
\n\nIf you would like some help in setting up your work environment, the people organizing the CS231n: Convolutional Neural Networks for Visual Recognition over at Stanford have got your back. Hope this helps.
\n\nBest of luck for the competition.
\n\nBest Regards,
\nShashank
Hi Shashank,
\n\nThanks for the suggestion. I've already set up Google cloud using those instructions. You mentioned above that we can get access to Tesla K80 cards. Can you provide some details?
", "votes": null, "replies": [] }, { "id": 207870, "author_name": "sshekhar", "author_url": "", "post_date": "07/27/2017 18:22:17", "content": "Once you have a Google Cloud account, you can access the IAM & Admin -> Quotas from the left panel. You can then increase the quota for GPU for a particular project from zero to as much as 4.
\n\nPlease keep in mind thought that Google will ask for a small fee ($35) for increasing beyond 2 GPUs just to make sure you are genuine in your request. This fee will be added to your credits. You can increase upto 2 GPUs without having to pay anything.
", "votes": null, "replies": [] }, { "id": 208021, "author_name": "humananalog", "author_url": "", "post_date": "07/28/2017 07:59:47", "content": "There is also Google Cloud ML, which is specifically for running TensorFlow jobs.
", "votes": null, "replies": [] } ] }, { "id": 208091, "author_name": "cletoreyes", "author_url": "", "post_date": "07/28/2017 14:16:35", "content": "If you have a decent Nvidia GPU then installing tensorflow-gpu to your python environment might be a very good option https://www.tensorflow.org/install/ and setting up Keras on top of it, which is really easy to use if you have used scikit-learn before.\nhttps://keras.io/#installation
Here is a very good guide on building a system for Deep Learning: http://timdettmers.com/2015/03/09/deep-learning-hardware-guide/ \n
And a guide on selecting GPUs: http://timdettmers.com/2017/04/09/which-gpu-for-deep-learning/
Also getting a nice system and training NNs by yourself will probably be cheaper than using Cloud services, if you plan to do it for a while.
", "votes": null, "replies": [] }, { "id": 223832, "author_name": "wappyotoole", "author_url": "", "post_date": "09/23/2017 19:12:12", "content": "You can also try retraining an existing suitable Tensorflow model.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "207829": "Hi, sorry for the silly question. I have never worked with image datasets before. \n\nMy question to those who have worked with this kind of data... **will it be useful (good learning experience) to participate in this competition with 16GB system?**\n\nI am assuming that deep learning will play a big role in this competition, and with my current system specification it will be difficult to train deep learning models on this data. Any thoughts?", "207831": "16GB RAM will be fine, but i guess to use deep learning you should probably make use of GPU computing.", "207836": "Hi Anubhav,\n\nI am no expert by any means but most vision related problems today including image segmentation use convolutional neural nets to get state of the art results. I share your frustration in the inability to train such models on large datasets without having to chose a very small batch size (which in turns increases my training time) on my personal machine (which is 32GB mind you). \n\nI have found [Google Compute Engine][1] very handy for such tasks. You get $300 worth of credits on free signup and can get access to Tesla K80 cards on your instances for free by filling up a simple request to increase quota.\n\nIf you would like some help in setting up your work environment, the people organizing the [CS231n: Convolutional Neural Networks for Visual Recognition][2] over at Stanford have got your back. Hope this helps.\n\nBest of luck for the competition.\n\nBest Regards, \nShashank\n\n\n [1]: https://cloud.google.com/compute/\n [2]: http://cs231n.github.io/", "207855": "Hi Shashank,\n\nThanks for the suggestion. I've already set up Google cloud using those instructions. You mentioned above that we can get access to Tesla K80 cards. Can you provide some details?", "207870": "Once you have a Google Cloud account, you can access the [IAM & Admin -> Quotas][1] from the left panel. You can then increase the quota for GPU for a particular project from zero to as much as 4.\n\nPlease keep in mind thought that Google will ask for a small fee ($35) for increasing beyond 2 GPUs just to make sure you are genuine in your request. This fee will be added to your credits. You can increase upto 2 GPUs without having to pay anything.\n\n\n [1]: https://console.cloud.google.com/projectselector/iam-admin/quotas", "208021": "There is also Google Cloud ML, which is specifically for running TensorFlow jobs.", "208091": "If you have a decent Nvidia GPU then installing `tensorflow-gpu` to your python environment might be a very good option https://www.tensorflow.org/install/ and setting up Keras on top of it, which is really easy to use if you have used scikit-learn before.\nhttps://keras.io/#installation\n\n\nHere is a very good guide on building a system for Deep Learning: http://timdettmers.com/2015/03/09/deep-learning-hardware-guide/ \n