One epoch of training on 224x224 images using densenet161 takes ~1 hour on google colab's P-100.
", "messages": [ { "id": "753476", "postDate": "02/22/2020 08:33:59", "content": "One epoch of training on 224x224 images using densenet161 takes ~1 hour on google colab's P-100.
", "rawMarkdown": "One epoch of training on 224x224 images using densenet161 takes ~1 hour on google colab's P-100.", "votes": null }, { "id": "753482", "postDate": "02/22/2020 09:01:26", "content": "P100 are outdated, they are similar to 1080 Ti in performance. I wish Google and Kaggle were offering V100. Afterall, V100 are only 3 years old, maybe too recent for some ;)
", "rawMarkdown": "P100 are outdated, they are similar to 1080 Ti in performance. I wish Google and Kaggle were offering V100. Afterall, V100 are only 3 years old, maybe too recent for some ;)", "votes": null }, { "id": "753527", "postDate": "02/22/2020 10:11:45", "content": "I did not tried so high resolution, the maximum that I tested was 192x192 and cumulated with preprocessing techniques and let it train for 120-150 the added time was too big, so all my architectures in my test are on 128x128 now
", "rawMarkdown": "I did not tried so high resolution, the maximum that I tested was 192x192 and cumulated with preprocessing techniques and let it train for 120-150 the added time was too big, so all my architectures in my test are on 128x128 now", "votes": null }, { "id": "753567", "postDate": "02/22/2020 11:35:13", "content": "Nvidia should allow Quadro8000 for cloud computing ;)
", "rawMarkdown": "Nvidia should allow Quadro8000 for cloud computing ;)", "votes": null }, { "id": "753571", "postDate": "02/22/2020 11:40:19", "content": "You should try models with less parameters, smaller image size, and check your batch processing, because kaggle kernels only have 2 cores which is often the bottleneck when doing CV.
", "rawMarkdown": "You should try models with less parameters, smaller image size, and check your batch processing, because kaggle kernels only have 2 cores which is often the bottleneck when doing CV.", "votes": null }, { "id": "754111", "postDate": "02/23/2020 04:26:05", "content": "224x224 is almost twice the size of the original image. I don't think upscaling image will have any benefits at all.
", "rawMarkdown": "224x224 is almost twice the size of the original image. I don't think upscaling image will have any benefits at all.", "votes": null }, { "id": "754792", "postDate": "02/24/2020 03:51:59", "content": "V100 are much more expensive. Offering V100 could mean less training quota for every user. P100 is a good trade-off. Compared with GPUs, more CPUs are really needed.
", "rawMarkdown": "V100 are much more expensive. Offering V100 could mean less training quota for every user. P100 is a good trade-off. Compared with GPUs, more CPUs are really needed.", "votes": null } ], "comments": [ { "id": 753482, "author_name": "cpmpml", "author_url": "", "post_date": "02/22/2020 09:01:26", "content": "P100 are outdated, they are similar to 1080 Ti in performance. I wish Google and Kaggle were offering V100. Afterall, V100 are only 3 years old, maybe too recent for some ;)
", "votes": null, "replies": [ { "id": 753567, "author_name": "ilu000", "author_url": "", "post_date": "02/22/2020 11:35:13", "content": "Nvidia should allow Quadro8000 for cloud computing ;)
", "votes": null, "replies": [] }, { "id": 754792, "author_name": "syoya1997", "author_url": "", "post_date": "02/24/2020 03:51:59", "content": "V100 are much more expensive. Offering V100 could mean less training quota for every user. P100 is a good trade-off. Compared with GPUs, more CPUs are really needed.
", "votes": null, "replies": [] } ] }, { "id": 753527, "author_name": "vladvdv", "author_url": "", "post_date": "02/22/2020 10:11:45", "content": "I did not tried so high resolution, the maximum that I tested was 192x192 and cumulated with preprocessing techniques and let it train for 120-150 the added time was too big, so all my architectures in my test are on 128x128 now
", "votes": null, "replies": [] }, { "id": 753571, "author_name": "philippsinger", "author_url": "", "post_date": "02/22/2020 11:40:19", "content": "You should try models with less parameters, smaller image size, and check your batch processing, because kaggle kernels only have 2 cores which is often the bottleneck when doing CV.
", "votes": null, "replies": [] }, { "id": 754111, "author_name": "quandapro", "author_url": "", "post_date": "02/23/2020 04:26:05", "content": "224x224 is almost twice the size of the original image. I don't think upscaling image will have any benefits at all.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "753476": "One epoch of training on 224x224 images using densenet161 takes ~1 hour on google colab's P-100.", "753482": "P100 are outdated, they are similar to 1080 Ti in performance. I wish Google and Kaggle were offering V100. Afterall, V100 are only 3 years old, maybe too recent for some ;)", "753527": "I did not tried so high resolution, the maximum that I tested was 192x192 and cumulated with preprocessing techniques and let it train for 120-150 the added time was too big, so all my architectures in my test are on 128x128 now", "753567": "Nvidia should allow Quadro8000 for cloud computing ;)", "753571": "You should try models with less parameters, smaller image size, and check your batch processing, because kaggle kernels only have 2 cores which is often the bottleneck when doing CV.", "754111": "224x224 is almost twice the size of the original image. I don't think upscaling image will have any benefits at all.", "754792": "V100 are much more expensive. Offering V100 could mean less training quota for every user. P100 is a good trade-off. Compared with GPUs, more CPUs are really needed." }, "source": "meta" }