I was trying to add TPU capabilities to the current highest-scored public notebook. I am running into some problems, particularly an error that says \"A driver for /dev/accel0 was already created at…\" (similar to what we can find here and here).
\nIf someone could point me in the right direction, I would really appreciate it. I have shared the notebook with my changes: https://www.kaggle.com/code/fpeccia/keras-baseline-seq2seq-tpu
", "messages": [ { "id": 2872007, "postDate": "2024-06-14T14:14:22.740Z", "content": "I managed to make everything work by switching to tensorflow backend and updating keras and tensorflow to the latest version.
\nTo save the model do something like that
\n strategy.scope():\n tpu_model = build_model()\ntpu_model.fit(...)\n\nlocal_model = build_model()\nlocal_model.set_weights(tpu_model.get_weights())\nlocal_model.save()\nAs I suspected hahaha. Good job.
", "rawMarkdown": "As I suspected hahaha. Good job.", "replies": [ { "id": 2872028, "postDate": "2024-06-14T14:24:03.937Z", "content": "Thanks both, I had just tested changing to tensorflow backend and it worked :D
", "rawMarkdown": "Thanks both, I had just tested changing to tensorflow backend and it worked :D", "votes": 1 } ] } ] }, { "id": 2871957, "postDate": "2024-06-14T13:42:48.363Z", "content": "My gut feeling is that Jax is the culprit. However, I have never used Jax, so I won't dive into this. If you don't have a lot of experience with TPU, I STRONGLY recommend building on a working TPU notebook. TPU is hard enough when building on a working pipeline. Creating a new working TPU pipeline without a lot of experience will be a hard path full of tears and frustration…haha
", "rawMarkdown": "My gut feeling is that Jax is the culprit. However, I have never used Jax, so I won't dive into this. If you don't have a lot of experience with TPU, I STRONGLY recommend building on a working TPU notebook. TPU is hard enough when building on a working pipeline. Creating a new working TPU pipeline without a lot of experience will be a hard path full of tears and frustration...haha", "votes": 2 }, { "id": 2871767, "postDate": "2024-06-14T11:49:10.457Z", "content": "I was trying to add TPU capabilities to the current highest-scored public notebook. I am running into some problems, particularly an error that says \"A driver for /dev/accel0 was already created at…\" (similar to what we can find here and here).
\nIf someone could point me in the right direction, I would really appreciate it. I have shared the notebook with my changes: https://www.kaggle.com/code/fpeccia/keras-baseline-seq2seq-tpu
", "rawMarkdown": "I was trying to add TPU capabilities to the current highest-scored public notebook. I am running into some problems, particularly an error that says \"A driver for /dev/accel0 was already created at...\" (similar to what we can find [here ](https://github.com/google/jax/issues/13260) and [here](https://github.com/huggingface/accelerate/issues/525)).\n\nIf someone could point me in the right direction, I would really appreciate it. I have shared the notebook with my changes: https://www.kaggle.com/code/fpeccia/keras-baseline-seq2seq-tpu\n\n" }, { "id": 2881671, "postDate": "2024-06-20T22:32:27.313Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 2872007, "author_name": "Amedeo Biolatti", "author_url": "", "post_date": "2024-06-14T14:14:22.740000", "content": "I managed to make everything work by switching to tensorflow backend and updating keras and tensorflow to the latest version.
\nTo save the model do something like that
\n strategy.scope():\n tpu_model = build_model()\ntpu_model.fit(...)\n\nlocal_model = build_model()\nlocal_model.set_weights(tpu_model.get_weights())\nlocal_model.save()\nAs I suspected hahaha. Good job.
", "votes": 0, "replies": [ { "id": 2872028, "author_name": "Federico Peccia", "author_url": "", "post_date": "2024-06-14T14:24:03.937000", "content": "Thanks both, I had just tested changing to tensorflow backend and it worked :D
", "votes": 1, "replies": [] } ] } ] }, { "id": 2871957, "author_name": "greySnow", "author_url": "", "post_date": "2024-06-14T13:42:48.363000", "content": "My gut feeling is that Jax is the culprit. However, I have never used Jax, so I won't dive into this. If you don't have a lot of experience with TPU, I STRONGLY recommend building on a working TPU notebook. TPU is hard enough when building on a working pipeline. Creating a new working TPU pipeline without a lot of experience will be a hard path full of tears and frustration…haha
", "votes": 2, "replies": [] }, { "id": 2881671, "author_name": "", "author_url": "", "post_date": "2024-06-20T22:32:27.313000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2872007": "I managed to make everything work by switching to tensorflow backend and updating keras and tensorflow to the latest version.\n\nTo save the model do something like that\n\n```python\nwith strategy.scope():\n tpu_model = build_model()\ntpu_model.fit(...)\n\nlocal_model = build_model()\nlocal_model.set_weights(tpu_model.get_weights())\nlocal_model.save(\"model.keras\")\n```", "2871957": "My gut feeling is that Jax is the culprit. However, I have never used Jax, so I won't dive into this. If you don't have a lot of experience with TPU, I STRONGLY recommend building on a working TPU notebook. TPU is hard enough when building on a working pipeline. Creating a new working TPU pipeline without a lot of experience will be a hard path full of tears and frustration...haha", "2871767": "I was trying to add TPU capabilities to the current highest-scored public notebook. I am running into some problems, particularly an error that says \"A driver for /dev/accel0 was already created at...\" (similar to what we can find [here ](https://github.com/google/jax/issues/13260) and [here](https://github.com/huggingface/accelerate/issues/525)).\n\nIf someone could point me in the right direction, I would really appreciate it. I have shared the notebook with my changes: https://www.kaggle.com/code/fpeccia/keras-baseline-seq2seq-tpu\n\n", "2881671": "" } }