Hi! This is somewhat of a noobish question when it comes to Kaggle kernels, but I was wondering how do you usually monitor the training process on Kaggle kernels?
\nDo you just log the loss in a text file and use it to plot later? Or is some kind of tensorboard-ish workaround available with Kaggle kernels as well?
\nI am using Pytorch for this competition.
\nThanks!
", "messages": [ { "id": "979324", "postDate": "08/20/2020 19:02:11", "content": "Hi! This is somewhat of a noobish question when it comes to Kaggle kernels, but I was wondering how do you usually monitor the training process on Kaggle kernels?
\nDo you just log the loss in a text file and use it to plot later? Or is some kind of tensorboard-ish workaround available with Kaggle kernels as well?
\nI am using Pytorch for this competition.
\nThanks!
", "rawMarkdown": "Hi! This is somewhat of a noobish question when it comes to Kaggle kernels, but I was wondering how do you usually monitor the training process on Kaggle kernels? \n\nDo you just log the loss in a text file and use it to plot later? Or is some kind of tensorboard-ish workaround available with Kaggle kernels as well?\n\nI am using Pytorch for this competition.\n\nThanks!", "votes": null }, { "id": "979406", "postDate": "08/20/2020 20:04:39", "content": "This should be helpful:
\nhttps://pytorch.org/tutorials/intermediate/tensorboard_tutorial.html
", "rawMarkdown": "This should be helpful:\n\nhttps://pytorch.org/tutorials/intermediate/tensorboard_tutorial.html", "votes": null }, { "id": "979848", "postDate": "08/21/2020 06:40:17", "content": "Thanks for the reply, yeah I know about the summary writer, but I wanted to know whether there is something real-time available that can be hooked onto the kaggle kernels.
\nI guess it is easier to write the logs in a text file, parse them and analyze later my plotting them.
Thanks!
", "rawMarkdown": "Thanks for the reply, yeah I know about the summary writer, but I wanted to know whether there is something real-time available that can be hooked onto the kaggle kernels. \nI guess it is easier to write the logs in a text file, parse them and analyze later my plotting them.\n\nThanks!", "votes": null }, { "id": "979896", "postDate": "08/21/2020 07:29:20", "content": "There are couple of experiment tracking tools e.g. Neptune.ai, MLFlow, Comet and Tensorboard as well. Some of them are free to use and can be easily integrated into your code. I know that PyTorch Lightning provides an interface* for all of these, so you can use e.g. Neptune.ai in your experiment and track all metrics and artifacts in real-time through your web browser. I think that running the code with Kaggle notebok shouldn't complicate anything here.
\n[*] - you can also intergate them with other tools like TF or Torch I believe
", "rawMarkdown": "There are couple of experiment tracking tools e.g. Neptune.ai, MLFlow, Comet and Tensorboard as well. Some of them are free to use and can be easily integrated into your code. I know that PyTorch Lightning provides an interface* for all of these, so you can use e.g. Neptune.ai in your experiment and track all metrics and artifacts in real-time through your web browser. I think that running the code with Kaggle notebok shouldn't complicate anything here.\n\n[*] - you can also intergate them with other tools like TF or Torch I believe", "votes": null }, { "id": "979899", "postDate": "08/21/2020 07:31:48", "content": "Here's my project using Lightning and Neptune.ai integration: https://github.com/mtszkw/surface-crack
\nYou can see how it's visualized here: https://ui.neptune.ai/mtszkw/surface-crack-detect/experiments?viewId=standard-view
checkout wandb. I have used it in kaggle kernels. it's realtime and you can compare multiple runs at one place without any hassle. definitely recomended.
", "rawMarkdown": "checkout [wandb](https://www.wandb.com/). I have used it in kaggle kernels. it's realtime and you can compare multiple runs at one place without any hassle. definitely recomended.", "votes": null } ], "comments": [ { "id": 979406, "author_name": "pukkinming", "author_url": "", "post_date": "08/20/2020 20:04:39", "content": "This should be helpful:
\nhttps://pytorch.org/tutorials/intermediate/tensorboard_tutorial.html
", "votes": null, "replies": [ { "id": 979848, "author_name": "saran95", "author_url": "", "post_date": "08/21/2020 06:40:17", "content": "Thanks for the reply, yeah I know about the summary writer, but I wanted to know whether there is something real-time available that can be hooked onto the kaggle kernels.
\nI guess it is easier to write the logs in a text file, parse them and analyze later my plotting them.
Thanks!
", "votes": null, "replies": [] } ] }, { "id": 979896, "author_name": "mtszkw", "author_url": "", "post_date": "08/21/2020 07:29:20", "content": "There are couple of experiment tracking tools e.g. Neptune.ai, MLFlow, Comet and Tensorboard as well. Some of them are free to use and can be easily integrated into your code. I know that PyTorch Lightning provides an interface* for all of these, so you can use e.g. Neptune.ai in your experiment and track all metrics and artifacts in real-time through your web browser. I think that running the code with Kaggle notebok shouldn't complicate anything here.
\n[*] - you can also intergate them with other tools like TF or Torch I believe
", "votes": null, "replies": [ { "id": 979899, "author_name": "mtszkw", "author_url": "", "post_date": "08/21/2020 07:31:48", "content": "Here's my project using Lightning and Neptune.ai integration: https://github.com/mtszkw/surface-crack
\nYou can see how it's visualized here: https://ui.neptune.ai/mtszkw/surface-crack-detect/experiments?viewId=standard-view
checkout wandb. I have used it in kaggle kernels. it's realtime and you can compare multiple runs at one place without any hassle. definitely recomended.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "979324": "Hi! This is somewhat of a noobish question when it comes to Kaggle kernels, but I was wondering how do you usually monitor the training process on Kaggle kernels? \n\nDo you just log the loss in a text file and use it to plot later? Or is some kind of tensorboard-ish workaround available with Kaggle kernels as well?\n\nI am using Pytorch for this competition.\n\nThanks!", "979406": "This should be helpful:\n\nhttps://pytorch.org/tutorials/intermediate/tensorboard_tutorial.html", "979848": "Thanks for the reply, yeah I know about the summary writer, but I wanted to know whether there is something real-time available that can be hooked onto the kaggle kernels. \nI guess it is easier to write the logs in a text file, parse them and analyze later my plotting them.\n\nThanks!", "979896": "There are couple of experiment tracking tools e.g. Neptune.ai, MLFlow, Comet and Tensorboard as well. Some of them are free to use and can be easily integrated into your code. I know that PyTorch Lightning provides an interface* for all of these, so you can use e.g. Neptune.ai in your experiment and track all metrics and artifacts in real-time through your web browser. I think that running the code with Kaggle notebok shouldn't complicate anything here.\n\n[*] - you can also intergate them with other tools like TF or Torch I believe", "979899": "Here's my project using Lightning and Neptune.ai integration: https://github.com/mtszkw/surface-crack\nYou can see how it's visualized here: https://ui.neptune.ai/mtszkw/surface-crack-detect/experiments?viewId=standard-view", "980983": "checkout [wandb](https://www.wandb.com/). I have used it in kaggle kernels. it's realtime and you can compare multiple runs at one place without any hassle. definitely recomended." }, "source": "meta" }