Hi to all!
\nI got the accuracy of the model on the public dataset (val_lwlrap also increased) at 0.856 with a change of one parameter: sample_time from 6 to 10
\nAll changes
\n 'parse_params': {
\n 'cut_time': 11,
\n },
\n 'data_params': {
\n 'sample_time': 10.
\nAll other parameters of the model remained unchanged as in this notebook: https://www.kaggle.com/aikhmelnytskyy/resnet-tpu-on-colab-and-kaggle .
\nThis notebook does not work after the last changes on kaggle (Given the specifics of tf, I do not see anything strange here), but for Google colab still works well.
\nWho found a solution to this problem, please write.
\nBy the way, what sample_time parameters (or essentially similar parameters) do you use?
\nGood luck to all!
Hi to all!
\nI got the accuracy of the model on the public dataset (val_lwlrap also increased) at 0.856 with a change of one parameter: sample_time from 6 to 10
\nAll changes
\n 'parse_params': {
\n 'cut_time': 11,
\n },
\n 'data_params': {
\n 'sample_time': 10.
\nAll other parameters of the model remained unchanged as in this notebook: https://www.kaggle.com/aikhmelnytskyy/resnet-tpu-on-colab-and-kaggle .
\nThis notebook does not work after the last changes on kaggle (Given the specifics of tf, I do not see anything strange here), but for Google colab still works well.
\nWho found a solution to this problem, please write.
\nBy the way, what sample_time parameters (or essentially similar parameters) do you use?
\nGood luck to all!
Thanks @aikhmelnytskyy . As always this is helpful. I was using default 10 sec clips for each label out of which I was carving out 6 sec samples randomly. @barnwellguy also suggested this and I was planning to try some different combinations. I will keep you posted
\nNote you can get TPU working by commenting out the @tf function decorator at the top of the augmentations module (the one where def _preprocess_img(x, training=False, test=False) is defined). Also just make sure you turn off specaugment
", "rawMarkdown": "Thanks @aikhmelnytskyy . As always this is helpful. I was using default 10 sec clips for each label out of which I was carving out 6 sec samples randomly. @barnwellguy also suggested this and I was planning to try some different combinations. I will keep you posted\n\nNote you can get TPU working by commenting out the @tf function decorator at the top of the augmentations module (the one where def _preprocess_img(x, training=False, test=False) is defined). Also just make sure you turn off specaugment", "votes": null }, { "id": "1190370", "postDate": "02/07/2021 17:06:04", "content": "great work
", "rawMarkdown": "great work", "votes": null }, { "id": "1190373", "postDate": "02/07/2021 17:08:29", "content": "Thank you very much! I have already made the necessary changes to the public notebook.
", "rawMarkdown": "Thank you very much! I have already made the necessary changes to the public notebook.", "votes": null }, { "id": "1190374", "postDate": "02/07/2021 17:09:28", "content": "Thanks! Good luck!
", "rawMarkdown": "Thanks! Good luck!", "votes": null }, { "id": "1192106", "postDate": "02/09/2021 00:12:54", "content": "Hi,I run this notebook have some bug.Is it must run on colab?
", "rawMarkdown": "Hi,I run this notebook have some bug.Is it must run on colab?", "votes": null }, { "id": "1192592", "postDate": "02/09/2021 07:49:35", "content": "Hi there! After changing the environment kaggle notebook stopped working, in version 4 I made all the necessary changes to make it work again, on COLAB everything works without changes
", "rawMarkdown": "Hi there! After changing the environment kaggle notebook stopped working, in version 4 I made all the necessary changes to make it work again, on COLAB everything works without changes", "votes": null }, { "id": "1194405", "postDate": "02/10/2021 07:24:39", "content": "@aikhmelnytskyy can u check once? I think the tf env has broken once again? even with the tf.function and specaug commenting
", "rawMarkdown": "aikhmelnytskyy can u check once? I think the tf env has broken once again? even with the tf.function and specaug commenting", "votes": null }, { "id": "1194436", "postDate": "02/10/2021 07:43:28", "content": "Yes,I met that.
", "rawMarkdown": "Yes,I met that.", "votes": null }, { "id": "1194437", "postDate": "02/10/2021 07:43:53", "content": "Thanks you!:)
", "rawMarkdown": "Thanks you!:)", "votes": null }, { "id": "1194455", "postDate": "02/10/2021 07:51:39", "content": "I just ran this notebook https://www.kaggle.com/aikhmelnytskyy/resnet-wavenet-my-best-single-model-ensemble, the model works (at least 1 first epoch, run training again)
", "rawMarkdown": "I just ran this notebook https://www.kaggle.com/aikhmelnytskyy/resnet-wavenet-my-best-single-model-ensemble, the model works (at least 1 first epoch, run training again)", "votes": null }, { "id": "1194731", "postDate": "02/10/2021 10:33:31", "content": "Hmm..seems to be working now…shady issue :)
\nAnyway thanks for your quick support…really appreciate
Hi @aikhmelnytskyy,
\nI have run with above changes and got 0.877 score, still missing the bronze range :(
\nThanks and Regards,
\nSaurabh Bagchi
Hi there! Did you get a public accuracy of 0.877 with one model? If so it is a very, very good result.
", "rawMarkdown": "Hi there! Did you get a public accuracy of 0.877 with one model? If so it is a very, very good result.", "votes": null }, { "id": "1195276", "postDate": "02/10/2021 16:44:54", "content": "Yes, thanks! Just followed your guidance, learnt a lot from you!
", "rawMarkdown": "Yes, thanks! Just followed your guidance, learnt a lot from you!", "votes": null }, { "id": "1195340", "postDate": "02/10/2021 17:52:58", "content": "Just make an ensemble with a model with a different pipeline) Good luck!
", "rawMarkdown": "Just make an ensemble with a model with a different pipeline) Good luck!", "votes": null }, { "id": "1196726", "postDate": "02/11/2021 15:38:58", "content": "The model already has an ensemble. I am assuming u got 87.7 with that? otherwise 87.7 for a single model is definitely in medal zone (after ensembling)…maybe even silver if the right ensembling is done
", "rawMarkdown": "The model already has an ensemble. I am assuming u got 87.7 with that? otherwise 87.7 for a single model is definitely in medal zone (after ensembling)...maybe even silver if the right ensembling is done", "votes": null } ], "comments": [ { "id": 1188511, "author_name": "allohvk", "author_url": "", "post_date": "02/06/2021 09:51:04", "content": "Thanks @aikhmelnytskyy . As always this is helpful. I was using default 10 sec clips for each label out of which I was carving out 6 sec samples randomly. @barnwellguy also suggested this and I was planning to try some different combinations. I will keep you posted
\nNote you can get TPU working by commenting out the @tf function decorator at the top of the augmentations module (the one where def _preprocess_img(x, training=False, test=False) is defined). Also just make sure you turn off specaugment
", "votes": null, "replies": [ { "id": 1190373, "author_name": "aikhmelnytskyy", "author_url": "", "post_date": "02/07/2021 17:08:29", "content": "Thank you very much! I have already made the necessary changes to the public notebook.
", "votes": null, "replies": [] }, { "id": 1194405, "author_name": "allohvk", "author_url": "", "post_date": "02/10/2021 07:24:39", "content": "@aikhmelnytskyy can u check once? I think the tf env has broken once again? even with the tf.function and specaug commenting
", "votes": null, "replies": [] }, { "id": 1194436, "author_name": "zekunn", "author_url": "", "post_date": "02/10/2021 07:43:28", "content": "Yes,I met that.
", "votes": null, "replies": [] }, { "id": 1194455, "author_name": "aikhmelnytskyy", "author_url": "", "post_date": "02/10/2021 07:51:39", "content": "I just ran this notebook https://www.kaggle.com/aikhmelnytskyy/resnet-wavenet-my-best-single-model-ensemble, the model works (at least 1 first epoch, run training again)
", "votes": null, "replies": [] }, { "id": 1194731, "author_name": "allohvk", "author_url": "", "post_date": "02/10/2021 10:33:31", "content": "Hmm..seems to be working now…shady issue :)
\nAnyway thanks for your quick support…really appreciate
Hi @aikhmelnytskyy,
\nI have run with above changes and got 0.877 score, still missing the bronze range :(
\nThanks and Regards,
\nSaurabh Bagchi
Hi there! Did you get a public accuracy of 0.877 with one model? If so it is a very, very good result.
", "votes": null, "replies": [] }, { "id": 1195276, "author_name": "saurabhbagchi", "author_url": "", "post_date": "02/10/2021 16:44:54", "content": "Yes, thanks! Just followed your guidance, learnt a lot from you!
", "votes": null, "replies": [] }, { "id": 1195340, "author_name": "aikhmelnytskyy", "author_url": "", "post_date": "02/10/2021 17:52:58", "content": "Just make an ensemble with a model with a different pipeline) Good luck!
", "votes": null, "replies": [] }, { "id": 1196726, "author_name": "allohvk", "author_url": "", "post_date": "02/11/2021 15:38:58", "content": "The model already has an ensemble. I am assuming u got 87.7 with that? otherwise 87.7 for a single model is definitely in medal zone (after ensembling)…maybe even silver if the right ensembling is done
", "votes": null, "replies": [] } ] }, { "id": 1190370, "author_name": "riadalmadani", "author_url": "", "post_date": "02/07/2021 17:06:04", "content": "great work
", "votes": null, "replies": [ { "id": 1190374, "author_name": "aikhmelnytskyy", "author_url": "", "post_date": "02/07/2021 17:09:28", "content": "Thanks! Good luck!
", "votes": null, "replies": [] } ] }, { "id": 1192106, "author_name": "zekunn", "author_url": "", "post_date": "02/09/2021 00:12:54", "content": "Hi,I run this notebook have some bug.Is it must run on colab?
", "votes": null, "replies": [ { "id": 1192592, "author_name": "aikhmelnytskyy", "author_url": "", "post_date": "02/09/2021 07:49:35", "content": "Hi there! After changing the environment kaggle notebook stopped working, in version 4 I made all the necessary changes to make it work again, on COLAB everything works without changes
", "votes": null, "replies": [] }, { "id": 1194437, "author_name": "zekunn", "author_url": "", "post_date": "02/10/2021 07:43:53", "content": "Thanks you!:)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1188435": "Hi to all!\nI got the accuracy of the model on the public dataset (val_lwlrap also increased) at 0.856 with a change of one parameter: sample_time from 6 to 10\nAll changes\n 'parse_params': {\n 'cut_time': 11,\n },\n 'data_params': {\n 'sample_time': 10.\nAll other parameters of the model remained unchanged as in this notebook: https://www.kaggle.com/aikhmelnytskyy/resnet-tpu-on-colab-and-kaggle .\nThis notebook does not work after the last changes on kaggle (Given the specifics of tf, I do not see anything strange here), but for Google colab still works well. \nWho found a solution to this problem, please write.\nBy the way, what sample_time parameters (or essentially similar parameters) do you use?\nGood luck to all!", "1188511": "Thanks @aikhmelnytskyy . As always this is helpful. I was using default 10 sec clips for each label out of which I was carving out 6 sec samples randomly. @barnwellguy also suggested this and I was planning to try some different combinations. I will keep you posted\n\nNote you can get TPU working by commenting out the @tf function decorator at the top of the augmentations module (the one where def _preprocess_img(x, training=False, test=False) is defined). Also just make sure you turn off specaugment", "1190370": "great work", "1190373": "Thank you very much! I have already made the necessary changes to the public notebook.", "1190374": "Thanks! Good luck!", "1192106": "Hi,I run this notebook have some bug.Is it must run on colab?", "1192592": "Hi there! After changing the environment kaggle notebook stopped working, in version 4 I made all the necessary changes to make it work again, on COLAB everything works without changes", "1194405": "aikhmelnytskyy can u check once? I think the tf env has broken once again? even with the tf.function and specaug commenting", "1194436": "Yes,I met that.", "1194437": "Thanks you!:)", "1194455": "I just ran this notebook https://www.kaggle.com/aikhmelnytskyy/resnet-wavenet-my-best-single-model-ensemble, the model works (at least 1 first epoch, run training again)", "1194731": "Hmm..seems to be working now...shady issue :)\nAnyway thanks for your quick support...really appreciate", "1195243": "Hi @aikhmelnytskyy, \n\nI have run with above changes and got 0.877 score, still missing the bronze range :(\n\nThanks and Regards,\nSaurabh Bagchi", "1195267": "Hi there! Did you get a public accuracy of 0.877 with one model? If so it is a very, very good result.", "1195276": "Yes, thanks! Just followed your guidance, learnt a lot from you!", "1195340": "Just make an ensemble with a model with a different pipeline) Good luck!", "1196726": "The model already has an ensemble. I am assuming u got 87.7 with that? otherwise 87.7 for a single model is definitely in medal zone (after ensembling)...maybe even silver if the right ensembling is done" }, "source": "meta" }