{ "id": 121255, "title": "Anyone using PyTorch?", "url": "/competitions/tensorflow2-question-answering/discussion/121255", "author_name": "ilovescience", "post_date": "2019-12-12T03:05:40.954000", "votes": 2, "comment_count": 11, "views": 0, "content": "

Most people are using TF1.x or TF2.0. Has anybody made a submission successfully in PyTorch?\nI know there's a training kernel available, but there was no working inference to go along with it.

", "messages": [ { "id": 693147, "postDate": "2019-12-12T04:43:14.903Z", "content": "

I am using pytorch only

", "rawMarkdown": "I am using pytorch only", "votes": 2, "replies": [ { "id": 693169, "postDate": "2019-12-12T05:15:45.750Z", "content": "

Wow that's great. Is this based on the 0.48 TF baseline ported to PyTorch? I know there were some issues with inference with the pytorch kernel. Did you have similar issues?

", "rawMarkdown": "Wow that's great. Is this based on the 0.48 TF baseline ported to PyTorch? I know there were some issues with inference with the pytorch kernel. Did you have similar issues?" }, { "id": 693260, "postDate": "2019-12-12T07:33:33.427Z", "content": "

Haha I love how you rose from 108th to 17th place! Great work!

", "rawMarkdown": "Haha I love how you rose from 108th to 17th place! Great work!", "votes": 1 }, { "id": 693828, "postDate": "2019-12-12T20:33:26.373Z", "content": "

Would you mind writing a PyTorch starter? I don't even need the score to be high, just something I can start playing around with.

", "rawMarkdown": "Would you mind writing a PyTorch starter? I don't even need the score to be high, just something I can start playing around with. " } ] }, { "id": 693098, "postDate": "2019-12-12T03:05:40.953Z", "content": "

Most people are using TF1.x or TF2.0. Has anybody made a submission successfully in PyTorch?\nI know there's a training kernel available, but there was no working inference to go along with it.

", "rawMarkdown": "Most people are using TF1.x or TF2.0. Has anybody made a submission successfully in PyTorch?\nI know there's a training kernel available, but there was no working inference to go along with it. ", "votes": 2 }, { "id": 693306, "postDate": "2019-12-12T08:32:33.583Z", "content": "

PyTorch only. The only reason I see here for using TensorFlow is the monetary prize

", "rawMarkdown": "PyTorch only. The only reason I see here for using TensorFlow is the monetary prize ", "replies": [ { "id": 693368, "postDate": "2019-12-12T10:03:12.227Z", "content": "

But how about TPUs?

", "rawMarkdown": "But how about TPUs?" }, { "id": 693413, "postDate": "2019-12-12T10:43:48.613Z", "content": "

That is a valid point. However, I prefer to lose some extra money on GCP GPU or get a higher electricity bill than to keep losing a lot of time and nerves debugging the TensorFlow code :/

", "rawMarkdown": "That is a valid point. However, I prefer to lose some extra money on GCP GPU or get a higher electricity bill than to keep losing a lot of time and nerves debugging the TensorFlow code :/\n\n", "votes": 3 }, { "id": 693468, "postDate": "2019-12-12T12:00:21.857Z", "content": "

You can use TPUs with PyTorch as well - I have a pipeline which should work on TPUs (probably would need a fix here and there), but in practice training with GPUs is quick enough for now.

", "rawMarkdown": "You can use TPUs with PyTorch as well - I have a pipeline which should work on TPUs (probably would need a fix here and there), but in practice training with GPUs is quick enough for now.", "votes": 2 }, { "id": 693507, "postDate": "2019-12-12T13:09:40.387Z", "content": "

I think I’ll also rewrite the whole thing in PyTorch (maybe even just for the sake of experience- long time’s passed since I last time implemented any paper) but only if there’s some progress with the Metric :harold:

", "rawMarkdown": "I think I’ll also rewrite the whole thing in PyTorch (maybe even just for the sake of experience- long time’s passed since I last time implemented any paper) but only if there’s some progress with the Metric :harold:", "votes": 2 }, { "id": 693551, "postDate": "2019-12-12T13:55:49.730Z", "content": "

Yeah quite annoying, its still not clear if the metric is implemented correctly

", "rawMarkdown": "Yeah quite annoying, its still not clear if the metric is implemented correctly", "votes": 4 } ] }, { "id": 693107, "postDate": "2019-12-12T03:22:07.580Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 693147, "author_name": "Dieter", "author_url": "", "post_date": "2019-12-12T04:43:14.903000", "content": "

I am using pytorch only

", "votes": 2, "replies": [ { "id": 693169, "author_name": "ilovescience", "author_url": "", "post_date": "2019-12-12T05:15:45.750000", "content": "

Wow that's great. Is this based on the 0.48 TF baseline ported to PyTorch? I know there were some issues with inference with the pytorch kernel. Did you have similar issues?

", "votes": 0, "replies": [] }, { "id": 693260, "author_name": "ilovescience", "author_url": "", "post_date": "2019-12-12T07:33:33.427000", "content": "

Haha I love how you rose from 108th to 17th place! Great work!

", "votes": 1, "replies": [] }, { "id": 693828, "author_name": "ilovescience", "author_url": "", "post_date": "2019-12-12T20:33:26.373000", "content": "

Would you mind writing a PyTorch starter? I don't even need the score to be high, just something I can start playing around with.

", "votes": 0, "replies": [] } ] }, { "id": 693306, "author_name": "Dmytro Danevskyi", "author_url": "", "post_date": "2019-12-12T08:32:33.583000", "content": "

PyTorch only. The only reason I see here for using TensorFlow is the monetary prize

", "votes": 0, "replies": [ { "id": 693368, "author_name": "Yury Kashnitsky", "author_url": "", "post_date": "2019-12-12T10:03:12.227000", "content": "

But how about TPUs?

", "votes": 0, "replies": [] }, { "id": 693413, "author_name": "Dmytro Danevskyi", "author_url": "", "post_date": "2019-12-12T10:43:48.613000", "content": "

That is a valid point. However, I prefer to lose some extra money on GCP GPU or get a higher electricity bill than to keep losing a lot of time and nerves debugging the TensorFlow code :/

", "votes": 3, "replies": [] }, { "id": 693468, "author_name": "Konstantin Lopukhin", "author_url": "", "post_date": "2019-12-12T12:00:21.857000", "content": "

You can use TPUs with PyTorch as well - I have a pipeline which should work on TPUs (probably would need a fix here and there), but in practice training with GPUs is quick enough for now.

", "votes": 2, "replies": [] }, { "id": 693507, "author_name": "Yury Kashnitsky", "author_url": "", "post_date": "2019-12-12T13:09:40.387000", "content": "

I think I’ll also rewrite the whole thing in PyTorch (maybe even just for the sake of experience- long time’s passed since I last time implemented any paper) but only if there’s some progress with the Metric :harold:

", "votes": 2, "replies": [] }, { "id": 693551, "author_name": "Dieter", "author_url": "", "post_date": "2019-12-12T13:55:49.730000", "content": "

Yeah quite annoying, its still not clear if the metric is implemented correctly

", "votes": 4, "replies": [] } ] }, { "id": 693107, "author_name": "", "author_url": "", "post_date": "2019-12-12T03:22:07.580000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "693147": "I am using pytorch only", "693098": "Most people are using TF1.x or TF2.0. Has anybody made a submission successfully in PyTorch?\nI know there's a training kernel available, but there was no working inference to go along with it. ", "693306": "PyTorch only. The only reason I see here for using TensorFlow is the monetary prize ", "693107": "" } }