I want to train efficient net for this competition, is it possible to train on kaggle? for 30-50 epochs? \nif not, is there anyway or any other equivalent algorithm running as good as efficient net for this competition?
", "messages": [ { "id": 920186, "postDate": "2020-07-08T12:20:17.127Z", "content": "I want to train efficient net for this competition, is it possible to train on kaggle? for 30-50 epochs? \nif not, is there anyway or any other equivalent algorithm running as good as efficient net for this competition?
", "rawMarkdown": "I want to train efficient net for this competition, is it possible to train on kaggle? for 30-50 epochs? \nif not, is there anyway or any other equivalent algorithm running as good as efficient net for this competition?", "votes": 3 }, { "id": 925523, "postDate": "2020-07-12T06:28:57.717Z", "content": "Yes. I use tensorflow to train my model (EfficientNet B0, 36 * 256 * 256 tiles) on TPU. It can train 52 epochs in three hours. Importantly, you need to use tfrecords.
", "rawMarkdown": "Yes. I use tensorflow to train my model (EfficientNet B0, 36 * 256 * 256 tiles) on TPU. It can train 52 epochs in three hours. Importantly, you need to use tfrecords.", "votes": 1 }, { "id": 921631, "postDate": "2020-07-09T12:40:56.373Z", "content": "hey, when I am using the GPU unit it needs about an hour to 1.5 hours to get about 25 epochs for avout 8000 images.\nI have considered trainnig on my latop but I cannot download the huge data.\nSo to get a feeling for the different cnns and hyperparameter, its okay for me to train directly on kaggle. Hope that helps.
", "rawMarkdown": "hey, when I am using the GPU unit it needs about an hour to 1.5 hours to get about 25 epochs for avout 8000 images.\nI have considered trainnig on my latop but I cannot download the huge data.\nSo to get a feeling for the different cnns and hyperparameter, its okay for me to train directly on kaggle. Hope that helps.", "votes": 1 }, { "id": 920204, "postDate": "2020-07-08T12:42:01.270Z", "content": "You can save your model from one notebook and use it another one increasing the number of epoch on which the model is being trained. Otherwise you are limited by the kaggle runtime for the notebooks in commit.
", "rawMarkdown": "You can save your model from one notebook and use it another one increasing the number of epoch on which the model is being trained. Otherwise you are limited by the kaggle runtime for the notebooks in commit.", "votes": 2, "replies": [ { "id": 920220, "postDate": "2020-07-08T12:52:26.640Z", "content": "I tried this. my cross val accuracy increased but my accuracy after submission on kaggle decreased very badly. so I think this way might not work and I have to continuously train it
", "rawMarkdown": "I tried this. my cross val accuracy increased but my accuracy after submission on kaggle decreased very badly. so I think this way might not work and I have to continuously train it " }, { "id": 920721, "postDate": "2020-07-08T19:05:43.747Z", "content": "Ah did you fix the seed for the train-valid split? Maybe for the epochs in the other notebook after the random split you introduce the train images of the first notebook as validation images in the second epoch. So you basically lose the validation set and get high performance. And The scoreboard doesn't use accuracy but the quadratic weighted kappa measure.
", "rawMarkdown": "Ah did you fix the seed for the train-valid split? Maybe for the epochs in the other notebook after the random split you introduce the train images of the first notebook as validation images in the second epoch. So you basically lose the validation set and get high performance. And The scoreboard doesn't use accuracy but the quadratic weighted kappa measure." }, { "id": 920932, "postDate": "2020-07-08T23:46:14.023Z", "content": "@pranavkasela Yes you're right that's a probable possible scenario. @muhammad4hmed Another possible scenario is that you didn't change the starting learning rate when continuing the training. Re-starting from a big LR will destroy the optimal weights rather than keep improving them which make the model requires further epochs to re-find an optimal state.
", "rawMarkdown": "@pranavkasela Yes you're right that's a probable possible scenario. @muhammad4hmed Another possible scenario is that you didn't change the starting learning rate when continuing the training. Re-starting from a big LR will destroy the optimal weights rather than keep improving them which make the model requires further epochs to re-find an optimal state.", "votes": 1 } ] } ], "comments": [ { "id": 925523, "author_name": "David", "author_url": "", "post_date": "2020-07-12T06:28:57.717000", "content": "Yes. I use tensorflow to train my model (EfficientNet B0, 36 * 256 * 256 tiles) on TPU. It can train 52 epochs in three hours. Importantly, you need to use tfrecords.
", "votes": 1, "replies": [] }, { "id": 921631, "author_name": "Andreas Horlbeck", "author_url": "", "post_date": "2020-07-09T12:40:56.373000", "content": "hey, when I am using the GPU unit it needs about an hour to 1.5 hours to get about 25 epochs for avout 8000 images.\nI have considered trainnig on my latop but I cannot download the huge data.\nSo to get a feeling for the different cnns and hyperparameter, its okay for me to train directly on kaggle. Hope that helps.
", "votes": 1, "replies": [] }, { "id": 920204, "author_name": "Pranav Kasela", "author_url": "", "post_date": "2020-07-08T12:42:01.270000", "content": "You can save your model from one notebook and use it another one increasing the number of epoch on which the model is being trained. Otherwise you are limited by the kaggle runtime for the notebooks in commit.
", "votes": 2, "replies": [ { "id": 920220, "author_name": "Muhammad Ahmed", "author_url": "", "post_date": "2020-07-08T12:52:26.640000", "content": "I tried this. my cross val accuracy increased but my accuracy after submission on kaggle decreased very badly. so I think this way might not work and I have to continuously train it
", "votes": 0, "replies": [] }, { "id": 920721, "author_name": "Pranav Kasela", "author_url": "", "post_date": "2020-07-08T19:05:43.747000", "content": "Ah did you fix the seed for the train-valid split? Maybe for the epochs in the other notebook after the random split you introduce the train images of the first notebook as validation images in the second epoch. So you basically lose the validation set and get high performance. And The scoreboard doesn't use accuracy but the quadratic weighted kappa measure.
", "votes": 0, "replies": [] }, { "id": 920932, "author_name": "Fares FOURATI", "author_url": "", "post_date": "2020-07-08T23:46:14.023000", "content": "@pranavkasela Yes you're right that's a probable possible scenario. @muhammad4hmed Another possible scenario is that you didn't change the starting learning rate when continuing the training. Re-starting from a big LR will destroy the optimal weights rather than keep improving them which make the model requires further epochs to re-find an optimal state.
", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "920186": "I want to train efficient net for this competition, is it possible to train on kaggle? for 30-50 epochs? \nif not, is there anyway or any other equivalent algorithm running as good as efficient net for this competition?", "925523": "Yes. I use tensorflow to train my model (EfficientNet B0, 36 * 256 * 256 tiles) on TPU. It can train 52 epochs in three hours. Importantly, you need to use tfrecords.", "921631": "hey, when I am using the GPU unit it needs about an hour to 1.5 hours to get about 25 epochs for avout 8000 images.\nI have considered trainnig on my latop but I cannot download the huge data.\nSo to get a feeling for the different cnns and hyperparameter, its okay for me to train directly on kaggle. Hope that helps.", "920204": "You can save your model from one notebook and use it another one increasing the number of epoch on which the model is being trained. Otherwise you are limited by the kaggle runtime for the notebooks in commit." } }