{ "id": 150376, "title": "how much Pseudo Labeling contributes to your LB?", "url": "/competitions/global-wheat-detection/discussion/150376", "author_name": "", "post_date": "2020-05-12T04:26:20.717136400Z", "votes": 23, "comment_count": 20, "views": 0, "content": "

I think this thread is useful for people who use Psuedo technique. \nFor me, Psuedo improves my score from 0.7265 to 0.7423

", "messages": [ { "id": "843489", "postDate": "05/12/2020 04:26:20", "content": "

I think this thread is useful for people who use Psuedo technique. \nFor me, Psuedo improves my score from 0.7265 to 0.7423

", "rawMarkdown": "I think this thread is useful for people who use Psuedo technique. \nFor me, Psuedo improves my score from 0.7265 to 0.7423", "votes": null }, { "id": "843532", "postDate": "05/12/2020 05:07:05", "content": "

0.6913 to 0.6984

", "rawMarkdown": "0.6913 to 0.6984", "votes": null }, { "id": "845264", "postDate": "05/13/2020 06:00:44", "content": "

Hi, nvnn.\nThanks for sharing your kernel!

\n\n

I just started this competition and I'm going to improve my skills for object detection.

\n\n

' Psuedo improves my score from 0.7265 to 0.7423' means \nthat you get the score 0.72 not using psuedo labeling?

\n\n

And for inference using psuedo labeling,\nDo you continue to use updated weights(mean new retrained model)?

\n\n

Thanks.\np.s. congratulations on your first place :)

", "rawMarkdown": "Hi, nvnn.\nThanks for sharing your kernel!\n\nI just started this competition and I'm going to improve my skills for object detection.\n\n' Psuedo improves my score from 0.7265 to 0.7423' means \nthat you get the score 0.72 not using psuedo labeling?\n\nAnd for inference using psuedo labeling,\nDo you continue to use updated weights(mean new retrained model)?\n\nThanks.\np.s. congratulations on your first place :)", "votes": null }, { "id": "845397", "postDate": "05/13/2020 07:32:45", "content": "

Hi @piantic \nget the score 0.72 not using psuedo labeling then use the same model for psuedo labeling \np.s. i will be happy if I finish this competition in medal zone.

", "rawMarkdown": "Hi @piantic \nget the score 0.72 not using psuedo labeling then use the same model for psuedo labeling \np.s. i will be happy if I finish this competition in medal zone.", "votes": null }, { "id": "845658", "postDate": "05/13/2020 10:59:02", "content": "

Thanks :)

\n\n

I think you will get medals!

", "rawMarkdown": "Thanks :)\n\nI think you will get medals!", "votes": null }, { "id": "845725", "postDate": "05/13/2020 11:52:12", "content": "

Nice. I got managed to get 63.06 -> 65.76 with only 2 epochs for the PL generation.

\n\n

I think it is worth noting that the final evaluation will be done on 38% of the test data, so it's probable that the PL technique won't be as fruitful then.

", "rawMarkdown": "Nice. I got managed to get 63.06 -> 65.76 with only 2 epochs for the PL generation.\n\nI think it is worth noting that the final evaluation will be done on 38% of the test data, so it's probable that the PL technique won't be as fruitful then.", "votes": null }, { "id": "846586", "postDate": "05/13/2020 21:26:14", "content": "

0.65→0.70 with 3 rounds of labeling

", "rawMarkdown": "0.65→0.70 with 3 rounds of labeling", "votes": null }, { "id": "846705", "postDate": "05/14/2020 00:10:12", "content": "

thanks for sharing

", "rawMarkdown": "thanks for sharing", "votes": null }, { "id": "846706", "postDate": "05/14/2020 00:10:59", "content": "

does your model has enough time to do 4 or 5 rounds?

", "rawMarkdown": "does your model has enough time to do 4 or 5 rounds?", "votes": null }, { "id": "846745", "postDate": "05/14/2020 01:26:18", "content": "

unfortunatly, 3 rounds were the limit.\nI might make it with fewer epochs/rounds but I'm not sure if it will improve the LB

", "rawMarkdown": "unfortunatly, 3 rounds were the limit.\nI might make it with fewer epochs/rounds but I'm not sure if it will improve the LB", "votes": null }, { "id": "847531", "postDate": "05/14/2020 13:23:00", "content": "

and spend some time to do ensemble too.

", "rawMarkdown": "and spend some time to do ensemble too.", "votes": null }, { "id": "850129", "postDate": "05/16/2020 11:05:09", "content": "

0.6709 to 0.6963

", "rawMarkdown": "0.6709 to 0.6963", "votes": null }, { "id": "850943", "postDate": "05/17/2020 07:33:31", "content": "

I train for one round but train six epochs .0.6559 -> 6929 :) amazing technique

", "rawMarkdown": "I train for one round but train six epochs .0.6559 -> 6929 :) amazing technique", "votes": null }, { "id": "850994", "postDate": "05/17/2020 08:40:35", "content": "

good to hear. Use it wisely, because we might be overfitted.

", "rawMarkdown": "good to hear. Use it wisely, because we might be overfitted.", "votes": null }, { "id": "856221", "postDate": "05/21/2020 15:29:55", "content": "

I'm a little worried about over fitting

", "rawMarkdown": "I'm a little worried about over fitting", "votes": null }, { "id": "856643", "postDate": "05/22/2020 00:14:31", "content": "

you are right, use it as your own risk

", "rawMarkdown": "you are right, use it as your own risk", "votes": null }, { "id": "891399", "postDate": "06/18/2020 06:44:19", "content": "

I'm pretty confused because it did not improve my score so far. I've tried various configs (1,4, 10 epochs) with 1 or 3 rounds but I've got each time a smaller LB than my original model. Also tried to motdify the learning rate without success. Do you guys have any things/hints that worked for you ?

", "rawMarkdown": "I'm pretty confused because it did not improve my score so far. I've tried various configs (1,4, 10 epochs) with 1 or 3 rounds but I've got each time a smaller LB than my original model. Also tried to motdify the learning rate without success. Do you guys have any things/hints that worked for you ?", "votes": null }, { "id": "891603", "postDate": "06/18/2020 09:43:15", "content": "

If you're trying it with the EffDet implementation, I can tell you that for me it was far less successful than with FRCNN. My best guess as to why is because EffDet seems to bring us closer to some performance limit given our constraints

", "rawMarkdown": "If you're trying it with the EffDet implementation, I can tell you that for me it was far less successful than with FRCNN. My best guess as to why is because EffDet seems to bring us closer to some performance limit given our constraints", "votes": null }, { "id": "891742", "postDate": "06/18/2020 12:31:41", "content": "

Yes with efficientDet. Oh ok good to know you've experienced that too. \nIt seems generalizing well enough only with the training set\nAre you still using effDet ?

", "rawMarkdown": "Yes with efficientDet. Oh ok good to know you've experienced that too. \nIt seems generalizing well enough only with the training set\nAre you still using effDet ?", "votes": null }, { "id": "891927", "postDate": "06/18/2020 14:39:22", "content": "

\"It seems generalizing well enough only with the training set\", that's what's going on in my head too. But it feels wrong for some reason... Anyway I'd somehow be glad to see PL become unpopular in this comp - it's a pain to run, and adds lots of room for a final evaluation to fail.

", "rawMarkdown": "\"It seems generalizing well enough only with the training set\", that's what's going on in my head too. But it feels wrong for some reason... Anyway I'd somehow be glad to see PL become unpopular in this comp - it's a pain to run, and adds lots of room for a final evaluation to fail.", "votes": null }, { "id": "891981", "postDate": "06/18/2020 15:18:25", "content": "

Agree, that's the first time I use PL on a hidden test set and I was quite long to implement and run, it's also too blind. I'll first focus on developing a model that is strong enough and take a look on PL at the near end of the competition.

", "rawMarkdown": "Agree, that's the first time I use PL on a hidden test set and I was quite long to implement and run, it's also too blind. I'll first focus on developing a model that is strong enough and take a look on PL at the near end of the competition.", "votes": null } ], "comments": [ { "id": 843532, "author_name": "ma7moud", "author_url": "", "post_date": "05/12/2020 05:07:05", "content": "

0.6913 to 0.6984

", "votes": null, "replies": [] }, { "id": 845264, "author_name": "piantic", "author_url": "", "post_date": "05/13/2020 06:00:44", "content": "

Hi, nvnn.\nThanks for sharing your kernel!

\n\n

I just started this competition and I'm going to improve my skills for object detection.

\n\n

' Psuedo improves my score from 0.7265 to 0.7423' means \nthat you get the score 0.72 not using psuedo labeling?

\n\n

And for inference using psuedo labeling,\nDo you continue to use updated weights(mean new retrained model)?

\n\n

Thanks.\np.s. congratulations on your first place :)

", "votes": null, "replies": [ { "id": 845397, "author_name": "nvnnghia", "author_url": "", "post_date": "05/13/2020 07:32:45", "content": "

Hi @piantic \nget the score 0.72 not using psuedo labeling then use the same model for psuedo labeling \np.s. i will be happy if I finish this competition in medal zone.

", "votes": null, "replies": [] }, { "id": 845658, "author_name": "piantic", "author_url": "", "post_date": "05/13/2020 10:59:02", "content": "

Thanks :)

\n\n

I think you will get medals!

", "votes": null, "replies": [] } ] }, { "id": 845725, "author_name": "alexandersoare", "author_url": "", "post_date": "05/13/2020 11:52:12", "content": "

Nice. I got managed to get 63.06 -> 65.76 with only 2 epochs for the PL generation.

\n\n

I think it is worth noting that the final evaluation will be done on 38% of the test data, so it's probable that the PL technique won't be as fruitful then.

", "votes": null, "replies": [ { "id": 846705, "author_name": "nvnnghia", "author_url": "", "post_date": "05/14/2020 00:10:12", "content": "

thanks for sharing

", "votes": null, "replies": [] } ] }, { "id": 846586, "author_name": "kyoshioka47", "author_url": "", "post_date": "05/13/2020 21:26:14", "content": "

0.65→0.70 with 3 rounds of labeling

", "votes": null, "replies": [ { "id": 846706, "author_name": "nvnnghia", "author_url": "", "post_date": "05/14/2020 00:10:59", "content": "

does your model has enough time to do 4 or 5 rounds?

", "votes": null, "replies": [] }, { "id": 846745, "author_name": "kyoshioka47", "author_url": "", "post_date": "05/14/2020 01:26:18", "content": "

unfortunatly, 3 rounds were the limit.\nI might make it with fewer epochs/rounds but I'm not sure if it will improve the LB

", "votes": null, "replies": [] }, { "id": 847531, "author_name": "nvnnghia", "author_url": "", "post_date": "05/14/2020 13:23:00", "content": "

and spend some time to do ensemble too.

", "votes": null, "replies": [] } ] }, { "id": 850129, "author_name": "ozdata", "author_url": "", "post_date": "05/16/2020 11:05:09", "content": "

0.6709 to 0.6963

", "votes": null, "replies": [] }, { "id": 850943, "author_name": "raininbox", "author_url": "", "post_date": "05/17/2020 07:33:31", "content": "

I train for one round but train six epochs .0.6559 -> 6929 :) amazing technique

", "votes": null, "replies": [ { "id": 850994, "author_name": "nvnnghia", "author_url": "", "post_date": "05/17/2020 08:40:35", "content": "

good to hear. Use it wisely, because we might be overfitted.

", "votes": null, "replies": [] } ] }, { "id": 856221, "author_name": "yangxiao945", "author_url": "", "post_date": "05/21/2020 15:29:55", "content": "

I'm a little worried about over fitting

", "votes": null, "replies": [ { "id": 856643, "author_name": "nvnnghia", "author_url": "", "post_date": "05/22/2020 00:14:31", "content": "

you are right, use it as your own risk

", "votes": null, "replies": [] } ] }, { "id": 891399, "author_name": "alexj21", "author_url": "", "post_date": "06/18/2020 06:44:19", "content": "

I'm pretty confused because it did not improve my score so far. I've tried various configs (1,4, 10 epochs) with 1 or 3 rounds but I've got each time a smaller LB than my original model. Also tried to motdify the learning rate without success. Do you guys have any things/hints that worked for you ?

", "votes": null, "replies": [ { "id": 891603, "author_name": "alexandersoare", "author_url": "", "post_date": "06/18/2020 09:43:15", "content": "

If you're trying it with the EffDet implementation, I can tell you that for me it was far less successful than with FRCNN. My best guess as to why is because EffDet seems to bring us closer to some performance limit given our constraints

", "votes": null, "replies": [] }, { "id": 891742, "author_name": "alexj21", "author_url": "", "post_date": "06/18/2020 12:31:41", "content": "

Yes with efficientDet. Oh ok good to know you've experienced that too. \nIt seems generalizing well enough only with the training set\nAre you still using effDet ?

", "votes": null, "replies": [] }, { "id": 891927, "author_name": "alexandersoare", "author_url": "", "post_date": "06/18/2020 14:39:22", "content": "

\"It seems generalizing well enough only with the training set\", that's what's going on in my head too. But it feels wrong for some reason... Anyway I'd somehow be glad to see PL become unpopular in this comp - it's a pain to run, and adds lots of room for a final evaluation to fail.

", "votes": null, "replies": [] }, { "id": 891981, "author_name": "alexj21", "author_url": "", "post_date": "06/18/2020 15:18:25", "content": "

Agree, that's the first time I use PL on a hidden test set and I was quite long to implement and run, it's also too blind. I'll first focus on developing a model that is strong enough and take a look on PL at the near end of the competition.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "843489": "I think this thread is useful for people who use Psuedo technique. \nFor me, Psuedo improves my score from 0.7265 to 0.7423", "843532": "0.6913 to 0.6984", "845264": "Hi, nvnn.\nThanks for sharing your kernel!\n\nI just started this competition and I'm going to improve my skills for object detection.\n\n' Psuedo improves my score from 0.7265 to 0.7423' means \nthat you get the score 0.72 not using psuedo labeling?\n\nAnd for inference using psuedo labeling,\nDo you continue to use updated weights(mean new retrained model)?\n\nThanks.\np.s. congratulations on your first place :)", "845397": "Hi @piantic \nget the score 0.72 not using psuedo labeling then use the same model for psuedo labeling \np.s. i will be happy if I finish this competition in medal zone.", "845658": "Thanks :)\n\nI think you will get medals!", "845725": "Nice. I got managed to get 63.06 -> 65.76 with only 2 epochs for the PL generation.\n\nI think it is worth noting that the final evaluation will be done on 38% of the test data, so it's probable that the PL technique won't be as fruitful then.", "846586": "0.65→0.70 with 3 rounds of labeling", "846705": "thanks for sharing", "846706": "does your model has enough time to do 4 or 5 rounds?", "846745": "unfortunatly, 3 rounds were the limit.\nI might make it with fewer epochs/rounds but I'm not sure if it will improve the LB", "847531": "and spend some time to do ensemble too.", "850129": "0.6709 to 0.6963", "850943": "I train for one round but train six epochs .0.6559 -> 6929 :) amazing technique", "850994": "good to hear. Use it wisely, because we might be overfitted.", "856221": "I'm a little worried about over fitting", "856643": "you are right, use it as your own risk", "891399": "I'm pretty confused because it did not improve my score so far. I've tried various configs (1,4, 10 epochs) with 1 or 3 rounds but I've got each time a smaller LB than my original model. Also tried to motdify the learning rate without success. Do you guys have any things/hints that worked for you ?", "891603": "If you're trying it with the EffDet implementation, I can tell you that for me it was far less successful than with FRCNN. My best guess as to why is because EffDet seems to bring us closer to some performance limit given our constraints", "891742": "Yes with efficientDet. Oh ok good to know you've experienced that too. \nIt seems generalizing well enough only with the training set\nAre you still using effDet ?", "891927": "\"It seems generalizing well enough only with the training set\", that's what's going on in my head too. But it feels wrong for some reason... Anyway I'd somehow be glad to see PL become unpopular in this comp - it's a pain to run, and adds lots of room for a final evaluation to fail.", "891981": "Agree, that's the first time I use PL on a hidden test set and I was quite long to implement and run, it's also too blind. I'll first focus on developing a model that is strong enough and take a look on PL at the near end of the competition." }, "source": "meta" }