Of the popular ML toolsets, Tensorflow is the hardest to learn. Most people start with Fastai or Scikit Learn, then move to Pytorch to get more customization, and then finally to Tensorflow when they need more performance or want to impress Google recruiters specifically. That and it's pretty hard to fit a whole epoch of this training set into the nine-hour cap for a free GPU batch run.
", "rawMarkdown": "Of the popular ML toolsets, Tensorflow is the hardest to learn. Most people start with Fastai or Scikit Learn, then move to Pytorch to get more customization, and then finally to Tensorflow when they need more performance or want to impress Google recruiters specifically. That and it's pretty hard to fit a whole epoch of this training set into the nine-hour cap for a free GPU batch run.", "votes": 4, "replies": [ { "id": 931631, "postDate": "2020-07-16T10:31:56.473Z", "content": "Cool insight. The last sentence, isn't the nine-hour cap only required for model submission and not for training?
", "rawMarkdown": "Cool insight. The last sentence, isn't the nine-hour cap only required for model submission and not for training?", "votes": 2 } ] }, { "id": 933244, "postDate": "2020-07-17T15:34:56.480Z", "content": "Hardware required to get a competitive score could be one of factors stopping people from competing. I personally avoided making any submission because I know that with my modest hardware (one 1070 Ti and RTX 2060) I would probably end up in the bottom half πβ¦and neither can I afford the expensive cloud bill to run this on multiple GPU. Made my peace with a few notebook medals πβπΌ
", "rawMarkdown": "Hardware required to get a competitive score could be one of factors stopping people from competing. I personally avoided making any submission because I know that with my modest hardware (one 1070 Ti and RTX 2060) I would probably end up in the bottom half π...and neither can I afford the expensive cloud bill to run this on multiple GPU. Made my peace with a few notebook medals πβπΌ", "votes": 2 }, { "id": 931627, "postDate": "2020-07-16T10:28:04.090Z", "content": "DELG, which is probably the baseline, was trained with 21 P100 GPUs for 1.5 days on this dataset. It is just impossible to have enough hardware to participate.
", "rawMarkdown": "DELG, which is probably the baseline, was trained with 21 P100 GPUs for 1.5 days on this dataset. It is just impossible to have enough hardware to participate. ", "votes": 2 }, { "id": 929754, "postDate": "2020-07-14T23:24:41.533Z", "content": "Most participants can not exceed the baseline model.\nWhen training from baseline model, we have some problems.\nhttps://www.kaggle.com/c/landmark-retrieval-2020/discussion/166149
", "rawMarkdown": "Most participants can not exceed the baseline model.\nWhen training from baseline model, we have some problems.\nhttps://www.kaggle.com/c/landmark-retrieval-2020/discussion/166149", "votes": 2, "replies": [ { "id": 929907, "postDate": "2020-07-15T04:14:55.273Z", "content": "I'm also having problem when trying to finetune the baseline model. I hope that the organizers can provide the original code that generated the SaveModel that they provided. Seems that the backbone resembles the resnet_v1_101 in tf-slim, but I may never be sure.
I came to reply with this :)
", "rawMarkdown": "I came to reply with this :)" } ] }, { "id": 931085, "postDate": "2020-07-16T00:49:33.823Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 930577, "author_name": "Abyszero", "author_url": "", "post_date": "2020-07-15T14:57:08.337000", "content": "Of the popular ML toolsets, Tensorflow is the hardest to learn. Most people start with Fastai or Scikit Learn, then move to Pytorch to get more customization, and then finally to Tensorflow when they need more performance or want to impress Google recruiters specifically. That and it's pretty hard to fit a whole epoch of this training set into the nine-hour cap for a free GPU batch run.
", "votes": 4, "replies": [ { "id": 931631, "author_name": "Amar Maruf", "author_url": "", "post_date": "2020-07-16T10:31:56.473000", "content": "Cool insight. The last sentence, isn't the nine-hour cap only required for model submission and not for training?
", "votes": 2, "replies": [] } ] }, { "id": 933244, "author_name": "sandy1112", "author_url": "", "post_date": "2020-07-17T15:34:56.480000", "content": "Hardware required to get a competitive score could be one of factors stopping people from competing. I personally avoided making any submission because I know that with my modest hardware (one 1070 Ti and RTX 2060) I would probably end up in the bottom half πβ¦and neither can I afford the expensive cloud bill to run this on multiple GPU. Made my peace with a few notebook medals πβπΌ
", "votes": 2, "replies": [] }, { "id": 931627, "author_name": "Surui Li", "author_url": "", "post_date": "2020-07-16T10:28:04.090000", "content": "DELG, which is probably the baseline, was trained with 21 P100 GPUs for 1.5 days on this dataset. It is just impossible to have enough hardware to participate.
", "votes": 2, "replies": [] }, { "id": 929754, "author_name": "toshi_k", "author_url": "", "post_date": "2020-07-14T23:24:41.533000", "content": "Most participants can not exceed the baseline model.\nWhen training from baseline model, we have some problems.\nhttps://www.kaggle.com/c/landmark-retrieval-2020/discussion/166149
", "votes": 2, "replies": [ { "id": 929907, "author_name": "Peiyuan Liao", "author_url": "", "post_date": "2020-07-15T04:14:55.273000", "content": "I'm also having problem when trying to finetune the baseline model. I hope that the organizers can provide the original code that generated the SaveModel that they provided. Seems that the backbone resembles the resnet_v1_101 in tf-slim, but I may never be sure.
I came to reply with this :)
", "votes": 0, "replies": [] } ] }, { "id": 931085, "author_name": "", "author_url": "", "post_date": "2020-07-16T00:49:33.823000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "929111": "", "930577": "Of the popular ML toolsets, Tensorflow is the hardest to learn. Most people start with Fastai or Scikit Learn, then move to Pytorch to get more customization, and then finally to Tensorflow when they need more performance or want to impress Google recruiters specifically. That and it's pretty hard to fit a whole epoch of this training set into the nine-hour cap for a free GPU batch run.", "933244": "Hardware required to get a competitive score could be one of factors stopping people from competing. I personally avoided making any submission because I know that with my modest hardware (one 1070 Ti and RTX 2060) I would probably end up in the bottom half π...and neither can I afford the expensive cloud bill to run this on multiple GPU. Made my peace with a few notebook medals πβπΌ", "931627": "DELG, which is probably the baseline, was trained with 21 P100 GPUs for 1.5 days on this dataset. It is just impossible to have enough hardware to participate. ", "929754": "Most participants can not exceed the baseline model.\nWhen training from baseline model, we have some problems.\nhttps://www.kaggle.com/c/landmark-retrieval-2020/discussion/166149", "931085": "" } }