i am using pytorch and make custome dataset and using jpeg folder to laod image and resize it to 256 *256 and using pretrained resnset32 but still each epoch time is 1 hrs 40 mins so what is problem let me know it stops me to even submit my first submission
", "messages": [ { "id": "953002", "postDate": "07/31/2020 12:23:03", "content": "i am using pytorch and make custome dataset and using jpeg folder to laod image and resize it to 256 *256 and using pretrained resnset32 but still each epoch time is 1 hrs 40 mins so what is problem let me know it stops me to even submit my first submission
", "rawMarkdown": "i am using pytorch and make custome dataset and using jpeg folder to laod image and resize it to 256 *256 and using pretrained resnset32 but still each epoch time is 1 hrs 40 mins so what is problem let me know it stops me to even submit my first submission", "votes": null }, { "id": "953070", "postDate": "07/31/2020 13:54:59", "content": "If you are resizing in your training loop, you need to resize ahead of time and store the 256x256 images.
", "rawMarkdown": "If you are resizing in your training loop, you need to resize ahead of time and store the 256x256 images.", "votes": null }, { "id": "953175", "postDate": "07/31/2020 15:25:44", "content": "Yes. Some original images are 4000x6000. It takes our computers a lot of time to read these large files from disk and then resize them to 256x256. Try training with a Kaggle dataset where images have already been resized to 256x256 here.
If you want to cut a lot of training time, I can only advice you to go for TFRecords + TPU. There is a giant gain!\nThe best starting point:\nhttps://www.kaggle.com/cdeotte/triple-stratified-kfold-with-tfrecords (huge thanks to @cdeotte)\nIf you are not familiar with TPU and TFRecords, I recommend you take a look at this introduction, you will better understand that's going on in the notebook:\nhttps://codelabs.developers.google.com/codelabs/keras-flowers-data/#0\nLastly, if you want to use resnet based models with tf, take a look at this repository:\nhttps://github.com/qubvel/classification_models
", "rawMarkdown": "If you want to cut a lot of training time, I can only advice you to go for TFRecords + TPU. There is a giant gain!\nThe best starting point:\nhttps://www.kaggle.com/cdeotte/triple-stratified-kfold-with-tfrecords (huge thanks to @cdeotte)\nIf you are not familiar with TPU and TFRecords, I recommend you take a look at this introduction, you will better understand that's going on in the notebook:\nhttps://codelabs.developers.google.com/codelabs/keras-flowers-data/#0\nLastly, if you want to use resnet based models with tf, take a look at this repository:\nhttps://github.com/qubvel/classification_models", "votes": null }, { "id": "954462", "postDate": "08/01/2020 18:40:57", "content": "One thing you should make sure of is that you're training with an accelerator, like a GPU or TPU. You can do so by going into your editor, clicking the three dots on the menu bar on the right, and selecting your accelerator.
", "rawMarkdown": "One thing you should make sure of is that you're training with an accelerator, like a GPU or TPU. You can do so by going into your editor, clicking the three dots on the menu bar on the right, and selecting your accelerator.", "votes": null }, { "id": "954678", "postDate": "08/02/2020 01:57:02", "content": "Great suggestion
", "rawMarkdown": "Great suggestion", "votes": null }, { "id": "954857", "postDate": "08/02/2020 06:34:40", "content": "Use TPU. It will take some minutes to train on 256 X 256 images.
", "rawMarkdown": "Use TPU. It will take some minutes to train on 256 X 256 images.", "votes": null }, { "id": "955187", "postDate": "08/02/2020 12:26:25", "content": "thanks to everyone i do the same i used the kernel haveing 256 image size. but why its really slow when i resize it ont go it take about 2 hours for just one epoch with gpu
", "rawMarkdown": "thanks to everyone i do the same i used the kernel haveing 256 image size. but why its really slow when i resize it ont go it take about 2 hours for just one epoch with gpu", "votes": null } ], "comments": [ { "id": 953070, "author_name": "richardepstein", "author_url": "", "post_date": "07/31/2020 13:54:59", "content": "If you are resizing in your training loop, you need to resize ahead of time and store the 256x256 images.
", "votes": null, "replies": [ { "id": 953175, "author_name": "cdeotte", "author_url": "", "post_date": "07/31/2020 15:25:44", "content": "Yes. Some original images are 4000x6000. It takes our computers a lot of time to read these large files from disk and then resize them to 256x256. Try training with a Kaggle dataset where images have already been resized to 256x256 here.
If you want to cut a lot of training time, I can only advice you to go for TFRecords + TPU. There is a giant gain!\nThe best starting point:\nhttps://www.kaggle.com/cdeotte/triple-stratified-kfold-with-tfrecords (huge thanks to @cdeotte)\nIf you are not familiar with TPU and TFRecords, I recommend you take a look at this introduction, you will better understand that's going on in the notebook:\nhttps://codelabs.developers.google.com/codelabs/keras-flowers-data/#0\nLastly, if you want to use resnet based models with tf, take a look at this repository:\nhttps://github.com/qubvel/classification_models
", "votes": null, "replies": [] }, { "id": 954462, "author_name": "abhaykatoch", "author_url": "", "post_date": "08/01/2020 18:40:57", "content": "One thing you should make sure of is that you're training with an accelerator, like a GPU or TPU. You can do so by going into your editor, clicking the three dots on the menu bar on the right, and selecting your accelerator.
", "votes": null, "replies": [ { "id": 954678, "author_name": "cdeotte", "author_url": "", "post_date": "08/02/2020 01:57:02", "content": "Great suggestion
", "votes": null, "replies": [] } ] }, { "id": 954857, "author_name": "fireheart7", "author_url": "", "post_date": "08/02/2020 06:34:40", "content": "Use TPU. It will take some minutes to train on 256 X 256 images.
", "votes": null, "replies": [] }, { "id": 955187, "author_name": "jainji", "author_url": "", "post_date": "08/02/2020 12:26:25", "content": "thanks to everyone i do the same i used the kernel haveing 256 image size. but why its really slow when i resize it ont go it take about 2 hours for just one epoch with gpu
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "953002": "i am using pytorch and make custome dataset and using jpeg folder to laod image and resize it to 256 *256 and using pretrained resnset32 but still each epoch time is 1 hrs 40 mins so what is problem let me know it stops me to even submit my first submission", "953070": "If you are resizing in your training loop, you need to resize ahead of time and store the 256x256 images.", "953175": "Yes. Some original images are `4000x6000`. It takes our computers a lot of time to read these large files from disk and then resize them to `256x256`. Try training with a Kaggle dataset where images have already been resized to 256x256 [here][1].\n\n[1]: https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/164092", "954175": "If you want to cut a lot of training time, I can only advice you to go for TFRecords + TPU. There is a giant gain!\nThe best starting point:\nhttps://www.kaggle.com/cdeotte/triple-stratified-kfold-with-tfrecords (huge thanks to @cdeotte)\nIf you are not familiar with TPU and TFRecords, I recommend you take a look at this introduction, you will better understand that's going on in the notebook:\nhttps://codelabs.developers.google.com/codelabs/keras-flowers-data/#0\nLastly, if you want to use resnet based models with tf, take a look at this repository:\nhttps://github.com/qubvel/classification_models", "954462": "One thing you should make sure of is that you're training with an accelerator, like a GPU or TPU. You can do so by going into your editor, clicking the three dots on the menu bar on the right, and selecting your accelerator.", "954678": "Great suggestion", "954857": "Use TPU. It will take some minutes to train on 256 X 256 images.", "955187": "thanks to everyone i do the same i used the kernel haveing 256 image size. but why its really slow when i resize it ont go it take about 2 hours for just one epoch with gpu" }, "source": "meta" }