There is also a TPU on Google Colab. Can I use the TPU on Google Colab to execute the entry code?
", "messages": [ { "id": "764453", "postDate": "03/05/2020 13:32:16", "content": "There is also a TPU on Google Colab. Can I use the TPU on Google Colab to execute the entry code?
", "rawMarkdown": "There is also a TPU on Google Colab. Can I use the TPU on Google Colab to execute the entry code?", "votes": null }, { "id": "764482", "postDate": "03/05/2020 14:03:06", "content": "Yes, but colab can't use KaggleDatasets library(maybe..), so you need to save the GCS path by yourself.
\nAs following:\n
\nGCS_PATH = os.path.join('gs://kds-b2e6cdbc4af76dcf0363776c09c12fe46872cab211d1de9f60ec7aec', 'tfrecords-jpeg-512x512')\nTrain_filenames = tf.io.gfile.glob(os.path.join(GCS_PATH, 'train/*.tfrec'))\nValid_filenames = tf.io.gfile.glob(os.path.join(GCS_PATH, 'val/*.tfrec'))\nTest_filenames = tf.io.gfile.glob(os.path.join(GCS_PATH, 'test/*.tfrec'))\n
\nAnd rest of the parts are pretty much the same.\nBut I noticed that TPU in colab seems has smaller memory size, since I need to reduce the batch_size in colab with exactly the same setting in kaggle.
Yes, it should work fine with a couple of minor changes.
\n\n1) As @xiejialun said, you need to point to a GCS bucket with the data directly. The KaggleDatasets call is Kaggle specific.
2) Colab also defaults to Tensorflow 1.x (for now) and you need to add the following \"magic\" call to switch it to Tensorflow 2.1:\n\n%tensorflow_version 2.x\n
3) The TPU offered on Colab is indeed a TPU v2 (again @xiejialun got it right). Less powerful than the TPU v3 you get on Kaggle so yes please lower your batch size.
", "rawMarkdown": "Yes, it should work fine with a couple of minor changes.\n\n1) As @xiejialun said, you need to point to a GCS bucket with the data directly. The `KaggleDatasets` call is Kaggle specific.\n\n2) Colab also defaults to Tensorflow 1.x (for now) and you need to add the following \"magic\" call to switch it to Tensorflow 2.1:\n```\n%tensorflow_version 2.x\n```\n\n3) The TPU offered on Colab is indeed a TPU v2 (again @xiejialun got it right). Less powerful than the TPU v3 you get on Kaggle so yes please lower your batch size.", "votes": null }, { "id": "766698", "postDate": "03/08/2020 15:21:58", "content": "Is that GCS_PATH a valid google cloud storage link?
\n\nThanks for sharing your observations
", "rawMarkdown": "Is that GCS_PATH a valid google cloud storage link?\n\nThanks for sharing your observations", "votes": null }, { "id": "766912", "postDate": "03/08/2020 23:47:25", "content": "Seems to be, you can print the GCS_PATH from your kaggle kernel to check that, then use the print out GCS_PATH in your colab notebook.
", "rawMarkdown": "Seems to be, you can print the GCS_PATH from your kaggle kernel to check that, then use the print out GCS_PATH in your colab notebook.", "votes": null }, { "id": "1093672", "postDate": "11/27/2020 22:47:58", "content": "Is someone tried to use JAX with TPU on Kaggle?
", "rawMarkdown": "Is someone tried to use JAX with TPU on Kaggle?", "votes": null } ], "comments": [ { "id": 1093672, "author_name": "yurilla", "author_url": "", "post_date": "11/27/2020 22:47:58", "content": "Is someone tried to use JAX with TPU on Kaggle?
", "votes": null, "replies": [] }, { "id": 764482, "author_name": "xiejialun", "author_url": "", "post_date": "03/05/2020 14:03:06", "content": "Yes, but colab can't use KaggleDatasets library(maybe..), so you need to save the GCS path by yourself.
\nAs following:\n
\nGCS_PATH = os.path.join('gs://kds-b2e6cdbc4af76dcf0363776c09c12fe46872cab211d1de9f60ec7aec', 'tfrecords-jpeg-512x512')\nTrain_filenames = tf.io.gfile.glob(os.path.join(GCS_PATH, 'train/*.tfrec'))\nValid_filenames = tf.io.gfile.glob(os.path.join(GCS_PATH, 'val/*.tfrec'))\nTest_filenames = tf.io.gfile.glob(os.path.join(GCS_PATH, 'test/*.tfrec'))\n
\nAnd rest of the parts are pretty much the same.\nBut I noticed that TPU in colab seems has smaller memory size, since I need to reduce the batch_size in colab with exactly the same setting in kaggle.
Is that GCS_PATH a valid google cloud storage link?
\n\nThanks for sharing your observations
", "votes": null, "replies": [] }, { "id": 766912, "author_name": "xiejialun", "author_url": "", "post_date": "03/08/2020 23:47:25", "content": "Seems to be, you can print the GCS_PATH from your kaggle kernel to check that, then use the print out GCS_PATH in your colab notebook.
", "votes": null, "replies": [] } ] }, { "id": 764667, "author_name": "mgorner", "author_url": "", "post_date": "03/05/2020 17:55:57", "content": "Yes, it should work fine with a couple of minor changes.
\n\n1) As @xiejialun said, you need to point to a GCS bucket with the data directly. The KaggleDatasets call is Kaggle specific.
2) Colab also defaults to Tensorflow 1.x (for now) and you need to add the following \"magic\" call to switch it to Tensorflow 2.1:\n\n%tensorflow_version 2.x\n
3) The TPU offered on Colab is indeed a TPU v2 (again @xiejialun got it right). Less powerful than the TPU v3 you get on Kaggle so yes please lower your batch size.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "764453": "There is also a TPU on Google Colab. Can I use the TPU on Google Colab to execute the entry code?", "764482": "Yes, but colab can't use `KaggleDatasets` library(maybe..), so you need to save the GCS path by yourself. \nAs following:\n