\nAny ideas ?
\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1856617%2Fcb67f065d7ab2f161f02eff8c305d9c1%2FGPU.JPG?generation=1659333429572123&alt=media\")
I am trying a to build a LSTM model but while training it is only using CPu but not using GPU.
\nAny ideas ?
\n
Update:
\nNotebook https://www.kaggle.com/vjagannath786/pytorch-lstm
idk if this can help but i got the same problem while training a CNN, found out that preprocessing was slowing down the entire process and GPU was idling most of the time, waiting for the data from CPU
", "rawMarkdown": "idk if this can help but i got the same problem while training a CNN, found out that **preprocessing** was slowing down the entire process and GPU was idling most of the time, waiting for the data from CPU", "votes": 1 }, { "id": 1879742, "postDate": "2022-08-01T08:20:13.373Z", "content": "Maybe you want to see this notebook to know how to train an RNN on GPU.
", "rawMarkdown": "Maybe you want to see [this notebook](https://www.kaggle.com/code/voix97/amex-pytorch-gru-training) to know how to train an RNN on GPU.", "votes": 1, "replies": [ { "id": 1880147, "postDate": "2022-08-01T13:48:00.650Z", "content": "I am just using output layer of LSTM and i think you are using hidden layers.
\nWill try to post the notebook in some time.
\nplease try to look and let me know if any issues.
I think the most time-consuming step is this:
\n\ntk0 = tqdm(data_loader, total=len(data_loader))\nfor data in tk0:\n
\nmaybe you could test the wall time consumption of each step in model training (i.e. check the function train_fn)
I removed tqdm and was just using enumerate or simple for dataloader,
\nlooke like dataloader is the bottleneck.
\nEven with small set like 20k samples loader is very slow.
did you send model and data to GPU? (.to('cuda'))
", "rawMarkdown": "did you send model and data to GPU? (.to('cuda'))", "votes": 2, "replies": [ { "id": 1880425, "postDate": "2022-08-01T17:56:24.117Z", "content": "@titericz , yes and updated the topic with notebook , any ideas ?
", "rawMarkdown": "@titericz , yes and updated the topic with notebook , any ideas ?" } ] }, { "id": 1879555, "postDate": "2022-08-01T05:57:24.310Z", "content": "I am trying a to build a LSTM model but while training it is only using CPu but not using GPU.
\nAny ideas ?
\n
Update:
\nNotebook https://www.kaggle.com/vjagannath786/pytorch-lstm
did you select GPU on on the right?
", "rawMarkdown": "did you select GPU on on the right?", "replies": [ { "id": 1879719, "postDate": "2022-08-01T08:06:32.007Z", "content": "Yes @raddar ,I have selected the GPU. But, still it uses CPU. As you can see it is taking around 35 mins to complete one fold.
", "rawMarkdown": "Yes @raddar ,I have selected the GPU. But, still it uses CPU. As you can see it is taking around 35 mins to complete one fold." } ] }, { "id": 1879736, "postDate": "2022-08-01T08:15:43.573Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1880600, "author_name": "Nick Dubenco", "author_url": "", "post_date": "2022-08-01T21:04:18.807000", "content": "idk if this can help but i got the same problem while training a CNN, found out that preprocessing was slowing down the entire process and GPU was idling most of the time, waiting for the data from CPU
", "votes": 1, "replies": [] }, { "id": 1879742, "author_name": "Xiang Sheng", "author_url": "", "post_date": "2022-08-01T08:20:13.373000", "content": "Maybe you want to see this notebook to know how to train an RNN on GPU.
", "votes": 1, "replies": [ { "id": 1880147, "author_name": "Varun Jagannath", "author_url": "", "post_date": "2022-08-01T13:48:00.650000", "content": "I am just using output layer of LSTM and i think you are using hidden layers.
\nWill try to post the notebook in some time.
\nplease try to look and let me know if any issues.
I think the most time-consuming step is this:
\n\ntk0 = tqdm(data_loader, total=len(data_loader))\nfor data in tk0:\n
\nmaybe you could test the wall time consumption of each step in model training (i.e. check the function train_fn)
I removed tqdm and was just using enumerate or simple for dataloader,
\nlooke like dataloader is the bottleneck.
\nEven with small set like 20k samples loader is very slow.
did you send model and data to GPU? (.to('cuda'))
", "votes": 2, "replies": [ { "id": 1880425, "author_name": "Varun Jagannath", "author_url": "", "post_date": "2022-08-01T17:56:24.117000", "content": "@titericz , yes and updated the topic with notebook , any ideas ?
", "votes": 0, "replies": [] } ] }, { "id": 1879659, "author_name": "raddar", "author_url": "", "post_date": "2022-08-01T07:29:48.543000", "content": "did you select GPU on on the right?
", "votes": 0, "replies": [ { "id": 1879719, "author_name": "Varun Jagannath", "author_url": "", "post_date": "2022-08-01T08:06:32.007000", "content": "Yes @raddar ,I have selected the GPU. But, still it uses CPU. As you can see it is taking around 35 mins to complete one fold.
", "votes": 0, "replies": [] } ] }, { "id": 1879736, "author_name": "", "author_url": "", "post_date": "2022-08-01T08:15:43.573000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1880600": "idk if this can help but i got the same problem while training a CNN, found out that **preprocessing** was slowing down the entire process and GPU was idling most of the time, waiting for the data from CPU", "1879742": "Maybe you want to see [this notebook](https://www.kaggle.com/code/voix97/amex-pytorch-gru-training) to know how to train an RNN on GPU.", "1880329": "did you send model and data to GPU? (.to('cuda'))", "1879555": "I am trying a to build a LSTM model but while training it is only using CPu but not using GPU.\nAny ideas ?\n\n\n\nUpdate:\nNotebook https://www.kaggle.com/vjagannath786/pytorch-lstm", "1879659": "did you select GPU on on the right?", "1879736": "" } }