Hi, I am new to this and just wondering do we need GPU to get a good result? Thx.
", "messages": [ { "id": 403761, "postDate": "2018-10-14T14:48:54.390Z", "content": "It's not mandatory but your model will need to run for days/weeks if you want good results with CPU-only ( Particularly when it involves Convnets on images )
\n\nYou can also use kernels ... You have GPU option on kernels and you can run up to 10 GPU-kernels in parallel (very useful for k-folds training) .
\n\nAs for me , I don't have GPU. So I use Kernels in all images based competitions I participate (including this one) ... Even though I know I will hit here a big wall very soon, due to the huge size of the data ;)
", "rawMarkdown": "It's not mandatory but your model will need to run for days/weeks if you want good results with CPU-only ( Particularly when it involves Convnets on images ) \n\nYou can also use kernels ... You have GPU option on kernels and you can run up to 10 GPU-kernels in parallel (very useful for k-folds training) . \n\nAs for me , I don't have GPU. So I use Kernels in all images based competitions I participate (including this one) ... Even though I know I will hit here a big wall very soon, due to the huge size of the data ;)", "votes": 6, "replies": [ { "id": 408343, "postDate": "2018-10-22T17:57:35.553Z", "content": "That is very cool insight but How do you run gpu's in parallel?
", "rawMarkdown": "That is very cool insight but How do you run gpu's in parallel?" }, { "id": 408409, "postDate": "2018-10-22T20:34:28.960Z", "content": "Not GPUs in parallel (on single kernel) but many kernels with GPU
", "rawMarkdown": "Not GPUs in parallel (on single kernel) but many kernels with GPU " }, { "id": 408583, "postDate": "2018-10-23T05:43:44.180Z", "content": "I mean how will you transfer training on multiple kernels. Do you train batches of images on multiple kernels?
", "rawMarkdown": "I mean how will you transfer training on multiple kernels. Do you train batches of images on multiple kernels?" }, { "id": 408780, "postDate": "2018-10-23T13:03:09.737Z", "content": "I don't transfer training on multiple kernels. I use multiple kernels when I try many (independent ) experiments or when I need to train k-folds ( each fold per kernel)
", "rawMarkdown": "I don't transfer training on multiple kernels. I use multiple kernels when I try many (independent ) experiments or when I need to train k-folds ( each fold per kernel)", "votes": 1 }, { "id": 408944, "postDate": "2018-10-23T16:18:49.090Z", "content": "okay that makes sense
", "rawMarkdown": "okay that makes sense" } ] }, { "id": 403569, "postDate": "2018-10-14T01:17:50.120Z", "content": "Hi, I am new to this and just wondering do we need GPU to get a good result? Thx.
", "rawMarkdown": "Hi, I am new to this and just wondering do we need GPU to get a good result? Thx.", "votes": 3 }, { "id": 404902, "postDate": "2018-10-16T14:43:15.213Z", "content": "You probably need GPU to reach Silver/Bronze medal.\nI don't have GPU either. Previously I rented from AWS. For this competition I decided to burn my 300 GCP credit. As Serigne mentioned you could try kaggle kernels for free (e.g. https://www.kaggle.com/gaborfodor/black-white-cnn-lb-0-75).
\n\nK80 is not the best card for DL but it should be enough to start with.
", "rawMarkdown": "You probably need GPU to reach Silver/Bronze medal.\nI don't have GPU either. Previously I rented from AWS. For this competition I decided to burn my 300 GCP credit. As Serigne mentioned you could try kaggle kernels for free (e.g. https://www.kaggle.com/gaborfodor/black-white-cnn-lb-0-75).\n\nK80 is not the best card for DL but it should be enough to start with.", "votes": 2 }, { "id": 403591, "postDate": "2018-10-14T02:58:41.653Z", "content": "No its not. you can train your model on CPU also but the difference will be only the computational speed. People prefer GPU for NN is because in NN we need multiple parallel computation. \nLets say if it takes a model to train on CPU takes 1 hr then on GPU it might take 30 min or even less based on the way you utilise the GPU.
\n\nThat's my little experience with GPU. There are even more advantages I think. hope some else can add their input to it
", "rawMarkdown": "No its not. you can train your model on CPU also but the difference will be only the computational speed. People prefer GPU for NN is because in NN we need multiple parallel computation. \nLets say if it takes a model to train on CPU takes 1 hr then on GPU it might take 30 min or even less based on the way you utilise the GPU.\n\nThat's my little experience with GPU. There are even more advantages I think. hope some else can add their input to it" } ], "comments": [ { "id": 403761, "author_name": "Serigne ", "author_url": "", "post_date": "2018-10-14T14:48:54.390000", "content": "It's not mandatory but your model will need to run for days/weeks if you want good results with CPU-only ( Particularly when it involves Convnets on images )
\n\nYou can also use kernels ... You have GPU option on kernels and you can run up to 10 GPU-kernels in parallel (very useful for k-folds training) .
\n\nAs for me , I don't have GPU. So I use Kernels in all images based competitions I participate (including this one) ... Even though I know I will hit here a big wall very soon, due to the huge size of the data ;)
", "votes": 6, "replies": [ { "id": 408343, "author_name": "Ankit Sati", "author_url": "", "post_date": "2018-10-22T17:57:35.553000", "content": "That is very cool insight but How do you run gpu's in parallel?
", "votes": 0, "replies": [] }, { "id": 408409, "author_name": "Serigne ", "author_url": "", "post_date": "2018-10-22T20:34:28.960000", "content": "Not GPUs in parallel (on single kernel) but many kernels with GPU
", "votes": 0, "replies": [] }, { "id": 408583, "author_name": "Ankit Sati", "author_url": "", "post_date": "2018-10-23T05:43:44.180000", "content": "I mean how will you transfer training on multiple kernels. Do you train batches of images on multiple kernels?
", "votes": 0, "replies": [] }, { "id": 408780, "author_name": "Serigne ", "author_url": "", "post_date": "2018-10-23T13:03:09.737000", "content": "I don't transfer training on multiple kernels. I use multiple kernels when I try many (independent ) experiments or when I need to train k-folds ( each fold per kernel)
", "votes": 1, "replies": [] }, { "id": 408944, "author_name": "Ankit Sati", "author_url": "", "post_date": "2018-10-23T16:18:49.090000", "content": "okay that makes sense
", "votes": 0, "replies": [] } ] }, { "id": 404902, "author_name": "beluga", "author_url": "", "post_date": "2018-10-16T14:43:15.213000", "content": "You probably need GPU to reach Silver/Bronze medal.\nI don't have GPU either. Previously I rented from AWS. For this competition I decided to burn my 300 GCP credit. As Serigne mentioned you could try kaggle kernels for free (e.g. https://www.kaggle.com/gaborfodor/black-white-cnn-lb-0-75).
\n\nK80 is not the best card for DL but it should be enough to start with.
", "votes": 2, "replies": [] }, { "id": 403591, "author_name": "Ranjeet Jain", "author_url": "", "post_date": "2018-10-14T02:58:41.653000", "content": "No its not. you can train your model on CPU also but the difference will be only the computational speed. People prefer GPU for NN is because in NN we need multiple parallel computation. \nLets say if it takes a model to train on CPU takes 1 hr then on GPU it might take 30 min or even less based on the way you utilise the GPU.
\n\nThat's my little experience with GPU. There are even more advantages I think. hope some else can add their input to it
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "403761": "It's not mandatory but your model will need to run for days/weeks if you want good results with CPU-only ( Particularly when it involves Convnets on images ) \n\nYou can also use kernels ... You have GPU option on kernels and you can run up to 10 GPU-kernels in parallel (very useful for k-folds training) . \n\nAs for me , I don't have GPU. So I use Kernels in all images based competitions I participate (including this one) ... Even though I know I will hit here a big wall very soon, due to the huge size of the data ;)", "403569": "Hi, I am new to this and just wondering do we need GPU to get a good result? Thx.", "404902": "You probably need GPU to reach Silver/Bronze medal.\nI don't have GPU either. Previously I rented from AWS. For this competition I decided to burn my 300 GCP credit. As Serigne mentioned you could try kaggle kernels for free (e.g. https://www.kaggle.com/gaborfodor/black-white-cnn-lb-0-75).\n\nK80 is not the best card for DL but it should be enough to start with.", "403591": "No its not. you can train your model on CPU also but the difference will be only the computational speed. People prefer GPU for NN is because in NN we need multiple parallel computation. \nLets say if it takes a model to train on CPU takes 1 hr then on GPU it might take 30 min or even less based on the way you utilise the GPU.\n\nThat's my little experience with GPU. There are even more advantages I think. hope some else can add their input to it" } }