Hi kagglers. I'm struggling to catch up kkiller baseline performance with tensorflow. I generated mels for every 5 seconds clip and trained EfficientNetB0 (and other models) but i could'nt break past 0.61. So is somebody using tensorflow successfully in this competiton? I will be happy if you can share your best score achieved with a tensorflow pipeline.
", "messages": [ { "id": "1317562", "postDate": "05/21/2021 13:49:12", "content": "Hi kagglers. I'm struggling to catch up kkiller baseline performance with tensorflow. I generated mels for every 5 seconds clip and trained EfficientNetB0 (and other models) but i could'nt break past 0.61. So is somebody using tensorflow successfully in this competiton? I will be happy if you can share your best score achieved with a tensorflow pipeline.
", "rawMarkdown": "Hi kagglers. I'm struggling to catch up [kkiller](https://www.kaggle.com/kneroma) [baseline](https://www.kaggle.com/kneroma/clean-fast-simple-bird-identifier-inference) performance with tensorflow. I generated mels for every 5 seconds clip and trained EfficientNetB0 (and other models) but i could'nt break past 0.61. So is somebody using tensorflow successfully in this competiton? I will be happy if you can share your best score achieved with a tensorflow pipeline.", "votes": null }, { "id": "1318057", "postDate": "05/22/2021 00:14:18", "content": "Glad to see a fellow TF user! :) I just joined the competition, and am building TF/Keras pipelines at the moment. Will keep you posted soon.
", "rawMarkdown": "Glad to see a fellow TF user! :) I just joined the competition, and am building TF/Keras pipelines at the moment. Will keep you posted soon.", "votes": null }, { "id": "1318155", "postDate": "05/22/2021 04:22:11", "content": "wow surprised to see you back here! I am a big fan of your youtube channel - really helpful resources👍
", "rawMarkdown": "wow surprised to see you back here! I am a big fan of your youtube channel - really helpful resources👍", "votes": null }, { "id": "1318599", "postDate": "05/22/2021 12:30:03", "content": "I feel there are only few tensorflow users, so lonely…
\nAnyway i stick to use Keras because I want to use TPU.
I got 0.62 so far.
\nIt's code base is all most the same as I wrote in last year's birdcall competition.
\n108th Place Solution Birdcall Keras TPU
But you seem to reach 0.66.
\nDid you use Pytorch? 😂
Yes my pytorch pipeline scores 0.66 but the TF counterpart scores 0.60
", "rawMarkdown": "Yes my pytorch pipeline scores 0.66 but the TF counterpart scores 0.60", "votes": null }, { "id": "1318815", "postDate": "05/22/2021 15:30:24", "content": "You are doing well with pytorch but still also trying with TF. Great😉
", "rawMarkdown": "You are doing well with pytorch but still also trying with TF. Great😉", "votes": null }, { "id": "1323051", "postDate": "05/25/2021 23:10:41", "content": "Hi @jeongyoonlee can you release a proper TF Pipeline for object detection with custom dataset ? Most of the time we found pytorch pipelines and the few TF ones we find are not clear enough (at least for me). By the way your youtube channel is 👍👍👍
", "rawMarkdown": "Hi @jeongyoonlee can you release a proper TF Pipeline for object detection with custom dataset ? Most of the time we found pytorch pipelines and the few TF ones we find are not clear enough (at least for me). By the way your youtube channel is 👍👍👍", "votes": null }, { "id": "1323068", "postDate": "05/26/2021 00:08:05", "content": "Thanks @qypeng12 and @ulrich07 for kind words. :)
\nRegarding the TF pipeline, I combined two notebooks, @kneroma's pretrained model + MelSpectogram and @drcapa's 1d convnet + FFT in TF/Keras, but haven't gotten good performance yet. :(
\nI can share it just as a reference as is, but I thought that it'd be better if I share it after fixing the issue.
\nDo you think it'd be still useful? Let me know.
", "rawMarkdown": "Thanks @qypeng12 and @ulrich07 for kind words. :)\n\nRegarding the TF pipeline, I combined two notebooks, @kneroma's pretrained model + MelSpectogram and @drcapa's 1d convnet + FFT in TF/Keras, but haven't gotten good performance yet. :(\n\nI can share it just as a reference as is, but I thought that it'd be better if I share it after fixing the issue.\n\nDo you think it'd be still useful? Let me know.", "votes": null }, { "id": "1323106", "postDate": "05/26/2021 02:14:48", "content": "Oh man, i wish i could see more TF models around, but for the past 3 months i've been working on competitions and pytorch seems to be the king around here. TF is so user friendly, i wonder why torch is the favorite…
", "rawMarkdown": "Oh man, i wish i could see more TF models around, but for the past 3 months i've been working on competitions and pytorch seems to be the king around here. TF is so user friendly, i wonder why torch is the favorite...", "votes": null }, { "id": "1323107", "postDate": "05/26/2021 02:18:25", "content": "Please share, i am much more confortable with TF, i want to see what you did to convert those kernels to TF and (maybe) i can even help to find the issue
", "rawMarkdown": "Please share, i am much more confortable with TF, i want to see what you did to convert those kernels to TF and (maybe) i can even help to find the issue", "votes": null }, { "id": "1323134", "postDate": "05/26/2021 03:04:54", "content": "I think it may be because the only few baselines are written by pytorch, and everyone likes to change the existing ones instead of writing a new tf baseline.
", "rawMarkdown": "I think it may be because the only few baselines are written by pytorch, and everyone likes to change the existing ones instead of writing a new tf baseline.", "votes": null }, { "id": "1323388", "postDate": "05/26/2021 07:32:45", "content": "I concur. It depends on the competition. When the first good baseline was TF most used TF.
", "rawMarkdown": "I concur. It depends on the competition. When the first good baseline was TF most used TF.", "votes": null }, { "id": "1323411", "postDate": "05/26/2021 08:03:46", "content": "Yes of course. it would be useful.
", "rawMarkdown": "Yes of course. it would be useful.", "votes": null }, { "id": "1323418", "postDate": "05/26/2021 08:06:40", "content": "Nice observation @cpmpml
", "rawMarkdown": "Nice observation @cpmpml", "votes": null }, { "id": "1324962", "postDate": "05/27/2021 11:44:25", "content": "Hi Ulrich, I have had some success (0.70 LB) using TF with the tf.keras.applications.ResNet50V2 base model and a PANN head, using 11-fold CV. Good to see other TF users in this thread!
Nice to hear it and congrats @brentspell . I dashed in this competition too late. This encouraging performance leads me to stick with TF for the next Bird Challenge.
", "rawMarkdown": "Nice to hear it and congrats @brentspell . I dashed in this competition too late. This encouraging performance leads me to stick with TF for the next Bird Challenge.", "votes": null }, { "id": "1325659", "postDate": "05/27/2021 23:03:39", "content": "Hi @brentspell, nice to hear that, if you don't mind i would like you to share your pipeline after the deadline, i'm really curious about it
", "rawMarkdown": "Hi @brentspell, nice to hear that, if you don't mind i would like you to share your pipeline after the deadline, i'm really curious about it", "votes": null }, { "id": "1325698", "postDate": "05/28/2021 00:31:43", "content": "Thanks for the interest @victorasso, I'll see if I can put something together then.
", "rawMarkdown": "Thanks for the interest @victorasso, I'll see if I can put something together then.", "votes": null }, { "id": "1325716", "postDate": "05/28/2021 00:53:41", "content": "You can just post \"as is\" please don't put too much effort on it, i just want to see the pipeline, i have one in my mind that i never coded, i want to compare
", "rawMarkdown": "You can just post \"as is\" please don't put too much effort on it, i just want to see the pipeline, i have one in my mind that i never coded, i want to compare", "votes": null }, { "id": "1326161", "postDate": "05/28/2021 09:09:41", "content": "Yes sir @brentspell as @victorasso I am still interested in your TF Pipeline !
", "rawMarkdown": "Yes sir @brentspell as @victorasso I am still interested in your TF Pipeline !", "votes": null }, { "id": "1330635", "postDate": "06/01/2021 03:02:45", "content": "I ended up switching over to the PyTorch pipeline. Somehow, my equivalent TF pipeline produces all nocall predictions with whichever low thresholds. I'm also very curious who were able to use TF successfully in this competition.
Well thank you for the feedback, good luck! 👍
", "rawMarkdown": "Well thank you for the feedback, good luck! 👍", "votes": null } ], "comments": [ { "id": 1318057, "author_name": "jeongyoonlee", "author_url": "", "post_date": "05/22/2021 00:14:18", "content": "Glad to see a fellow TF user! :) I just joined the competition, and am building TF/Keras pipelines at the moment. Will keep you posted soon.
", "votes": null, "replies": [ { "id": 1318155, "author_name": "qypeng12", "author_url": "", "post_date": "05/22/2021 04:22:11", "content": "wow surprised to see you back here! I am a big fan of your youtube channel - really helpful resources👍
", "votes": null, "replies": [] }, { "id": 1323051, "author_name": "ulrich07", "author_url": "", "post_date": "05/25/2021 23:10:41", "content": "Hi @jeongyoonlee can you release a proper TF Pipeline for object detection with custom dataset ? Most of the time we found pytorch pipelines and the few TF ones we find are not clear enough (at least for me). By the way your youtube channel is 👍👍👍
", "votes": null, "replies": [] }, { "id": 1323068, "author_name": "jeongyoonlee", "author_url": "", "post_date": "05/26/2021 00:08:05", "content": "Thanks @qypeng12 and @ulrich07 for kind words. :)
\nRegarding the TF pipeline, I combined two notebooks, @kneroma's pretrained model + MelSpectogram and @drcapa's 1d convnet + FFT in TF/Keras, but haven't gotten good performance yet. :(
\nI can share it just as a reference as is, but I thought that it'd be better if I share it after fixing the issue.
\nDo you think it'd be still useful? Let me know.
", "votes": null, "replies": [] }, { "id": 1323107, "author_name": "victorasso", "author_url": "", "post_date": "05/26/2021 02:18:25", "content": "Please share, i am much more confortable with TF, i want to see what you did to convert those kernels to TF and (maybe) i can even help to find the issue
", "votes": null, "replies": [] }, { "id": 1323411, "author_name": "ulrich07", "author_url": "", "post_date": "05/26/2021 08:03:46", "content": "Yes of course. it would be useful.
", "votes": null, "replies": [] }, { "id": 1330635, "author_name": "jeongyoonlee", "author_url": "", "post_date": "06/01/2021 03:02:45", "content": "I ended up switching over to the PyTorch pipeline. Somehow, my equivalent TF pipeline produces all nocall predictions with whichever low thresholds. I'm also very curious who were able to use TF successfully in this competition.
Well thank you for the feedback, good luck! 👍
", "votes": null, "replies": [] } ] }, { "id": 1318599, "author_name": "enukuro", "author_url": "", "post_date": "05/22/2021 12:30:03", "content": "I feel there are only few tensorflow users, so lonely…
\nAnyway i stick to use Keras because I want to use TPU.
I got 0.62 so far.
\nIt's code base is all most the same as I wrote in last year's birdcall competition.
\n108th Place Solution Birdcall Keras TPU
But you seem to reach 0.66.
\nDid you use Pytorch? 😂
Yes my pytorch pipeline scores 0.66 but the TF counterpart scores 0.60
", "votes": null, "replies": [] }, { "id": 1318815, "author_name": "enukuro", "author_url": "", "post_date": "05/22/2021 15:30:24", "content": "You are doing well with pytorch but still also trying with TF. Great😉
", "votes": null, "replies": [] } ] }, { "id": 1323106, "author_name": "victorasso", "author_url": "", "post_date": "05/26/2021 02:14:48", "content": "Oh man, i wish i could see more TF models around, but for the past 3 months i've been working on competitions and pytorch seems to be the king around here. TF is so user friendly, i wonder why torch is the favorite…
", "votes": null, "replies": [ { "id": 1323134, "author_name": "zzy990106", "author_url": "", "post_date": "05/26/2021 03:04:54", "content": "I think it may be because the only few baselines are written by pytorch, and everyone likes to change the existing ones instead of writing a new tf baseline.
", "votes": null, "replies": [] }, { "id": 1323388, "author_name": "cpmpml", "author_url": "", "post_date": "05/26/2021 07:32:45", "content": "I concur. It depends on the competition. When the first good baseline was TF most used TF.
", "votes": null, "replies": [] }, { "id": 1323418, "author_name": "ulrich07", "author_url": "", "post_date": "05/26/2021 08:06:40", "content": "Nice observation @cpmpml
", "votes": null, "replies": [] } ] }, { "id": 1324962, "author_name": "brentspell", "author_url": "", "post_date": "05/27/2021 11:44:25", "content": "Hi Ulrich, I have had some success (0.70 LB) using TF with the tf.keras.applications.ResNet50V2 base model and a PANN head, using 11-fold CV. Good to see other TF users in this thread!
Nice to hear it and congrats @brentspell . I dashed in this competition too late. This encouraging performance leads me to stick with TF for the next Bird Challenge.
", "votes": null, "replies": [] }, { "id": 1325659, "author_name": "victorasso", "author_url": "", "post_date": "05/27/2021 23:03:39", "content": "Hi @brentspell, nice to hear that, if you don't mind i would like you to share your pipeline after the deadline, i'm really curious about it
", "votes": null, "replies": [] }, { "id": 1325698, "author_name": "brentspell", "author_url": "", "post_date": "05/28/2021 00:31:43", "content": "Thanks for the interest @victorasso, I'll see if I can put something together then.
", "votes": null, "replies": [] }, { "id": 1325716, "author_name": "victorasso", "author_url": "", "post_date": "05/28/2021 00:53:41", "content": "You can just post \"as is\" please don't put too much effort on it, i just want to see the pipeline, i have one in my mind that i never coded, i want to compare
", "votes": null, "replies": [] }, { "id": 1326161, "author_name": "ulrich07", "author_url": "", "post_date": "05/28/2021 09:09:41", "content": "Yes sir @brentspell as @victorasso I am still interested in your TF Pipeline !
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1317562": "Hi kagglers. I'm struggling to catch up [kkiller](https://www.kaggle.com/kneroma) [baseline](https://www.kaggle.com/kneroma/clean-fast-simple-bird-identifier-inference) performance with tensorflow. I generated mels for every 5 seconds clip and trained EfficientNetB0 (and other models) but i could'nt break past 0.61. So is somebody using tensorflow successfully in this competiton? I will be happy if you can share your best score achieved with a tensorflow pipeline.", "1318057": "Glad to see a fellow TF user! :) I just joined the competition, and am building TF/Keras pipelines at the moment. Will keep you posted soon.", "1318155": "wow surprised to see you back here! I am a big fan of your youtube channel - really helpful resources👍", "1318599": "I feel there are only few tensorflow users, so lonely... \nAnyway i stick to use Keras because I want to use TPU.\n\nI got 0.62 so far.\nIt's code base is all most the same as I wrote in last year's birdcall competition.\n[108th Place Solution Birdcall Keras TPU](https://www.kaggle.com/enukuro/108th-place-solution-birdcall-keras-tpu)", "1318609": "But you seem to reach 0.66.\nDid you use Pytorch? 😂", "1318782": "Yes my pytorch pipeline scores 0.66 but the TF counterpart scores 0.60", "1318815": "You are doing well with pytorch but still also trying with TF. Great😉", "1323051": "Hi @jeongyoonlee can you release a proper TF Pipeline for object detection with custom dataset ? Most of the time we found pytorch pipelines and the few TF ones we find are not clear enough (at least for me). By the way your youtube channel is 👍👍👍", "1323068": "Thanks @qypeng12 and @ulrich07 for kind words. :)\n\nRegarding the TF pipeline, I combined two notebooks, @kneroma's pretrained model + MelSpectogram and @drcapa's 1d convnet + FFT in TF/Keras, but haven't gotten good performance yet. :(\n\nI can share it just as a reference as is, but I thought that it'd be better if I share it after fixing the issue.\n\nDo you think it'd be still useful? Let me know.", "1323106": "Oh man, i wish i could see more TF models around, but for the past 3 months i've been working on competitions and pytorch seems to be the king around here. TF is so user friendly, i wonder why torch is the favorite...", "1323107": "Please share, i am much more confortable with TF, i want to see what you did to convert those kernels to TF and (maybe) i can even help to find the issue", "1323134": "I think it may be because the only few baselines are written by pytorch, and everyone likes to change the existing ones instead of writing a new tf baseline.", "1323388": "I concur. It depends on the competition. When the first good baseline was TF most used TF.", "1323411": "Yes of course. it would be useful.", "1323418": "Nice observation @cpmpml", "1324962": "Hi Ulrich, I have had some success (0.70 LB) using TF with the `tf.keras.applications.ResNet50V2` base model and a PANN head, using 11-fold CV. Good to see other TF users in this thread!", "1324976": "Nice to hear it and congrats @brentspell . I dashed in this competition too late. This encouraging performance leads me to stick with TF for the next Bird Challenge.", "1325659": "Hi @brentspell, nice to hear that, if you don't mind i would like you to share your pipeline after the deadline, i'm really curious about it", "1325698": "Thanks for the interest @victorasso, I'll see if I can put something together then.", "1325716": "You can just post \"as is\" please don't put too much effort on it, i just want to see the pipeline, i have one in my mind that i never coded, i want to compare", "1326161": "Yes sir @brentspell as @victorasso I am still interested in your TF Pipeline !", "1330635": "I ended up switching over to the PyTorch pipeline. Somehow, my equivalent TF pipeline produces all `nocall` predictions with whichever low thresholds. I'm also very curious who were able to use TF successfully in this competition.", "1332098": "Well thank you for the feedback, good luck! 👍" }, "source": "meta" }