{ "id": 540430, "title": "Regarding CPU, GPU and TPU runtime hours", "url": "/competitions/jane-street-real-time-market-data-forecasting/discussion/540430", "author_name": "", "post_date": "2024-10-14T14:38:54.887203600Z", "votes": 8, "comment_count": 2, "views": 0, "content": "

Here is an important note to take, which one might miss:

\n
    \n
  1. For the CPU and GPU the run-time for the training phase is <= 8 hours and not 9 hours (which is the typical case in other competitions).
    2. \n
  2. TPU is not allowed.
    3. \n
\n

For the forecasting purposes, the runtime is extended to 9 hours.

", "messages": [ { "id": "3017127", "postDate": "10/14/2024 14:38:54", "content": "

Here is an important note to take, which one might miss:

\n
    \n
  1. For the CPU and GPU the run-time for the training phase is <= 8 hours and not 9 hours (which is the typical case in other competitions).
    2. \n
  2. TPU is not allowed.
    3. \n
\n

For the forecasting purposes, the runtime is extended to 9 hours.

", "rawMarkdown": "Here is an important note to take, which one might miss:\n\n1. For the CPU and GPU the run-time for the **training phase is <= 8 hours** and not 9 hours (which is the typical case in other competitions). \n2. *TPU is not allowed*. \n\nFor the forecasting purposes, the runtime is extended to 9 hours.", "votes": null }, { "id": "3017180", "postDate": "10/14/2024 15:57:06", "content": "

Hi @ahsuna123, this is indeed an important point. The explanation lies in the data we use during the evaluation phase, detailed on the Data page. To help ensure continuity between the two phases, we'll be serving the entire test data (Public + Private) during the forecasting phase. You'll only need to make predictions on the private set, but processing the unscored portions will still take a bit of time. The extra hour is to give a bit of buffer to protect against time-out failures. It's still important to make your code robust to submission times, however.

", "rawMarkdown": "Hi @ahsuna123, this is indeed an important point. The explanation lies in the data we use during the evaluation phase, detailed on the [**Data page**](https://www.kaggle.com/competitions/jane-street-real-time-market-data-forecasting/data). To help ensure continuity between the two phases, we'll be serving the *entire* test data (Public + Private) during the forecasting phase. You'll only need to make predictions on the private set, but processing the unscored portions will still take a bit of time. The extra hour is to give a bit of buffer to protect against time-out failures. It's still important to make your code robust to submission times, however.", "votes": null }, { "id": "3017232", "postDate": "10/14/2024 16:53:53", "content": "

Hey, @ryanholbrook
\nThank you so much for taking the time and giving a detailed explanation. Much appreciated! :)
\nThis is my first time participating in Jane Street Group's competition and thought this might be a relevant information one might miss similar to my initial assumption of 9 hours.

", "rawMarkdown": "Hey, @ryanholbrook \nThank you so much for taking the time and giving a detailed explanation. Much appreciated! :)\nThis is my first time participating in Jane Street Group's competition and thought this might be a relevant information one might miss similar to my initial assumption of 9 hours.", "votes": null } ], "comments": [ { "id": 3017180, "author_name": "ryanholbrook", "author_url": "", "post_date": "10/14/2024 15:57:06", "content": "

Hi @ahsuna123, this is indeed an important point. The explanation lies in the data we use during the evaluation phase, detailed on the Data page. To help ensure continuity between the two phases, we'll be serving the entire test data (Public + Private) during the forecasting phase. You'll only need to make predictions on the private set, but processing the unscored portions will still take a bit of time. The extra hour is to give a bit of buffer to protect against time-out failures. It's still important to make your code robust to submission times, however.

", "votes": null, "replies": [ { "id": 3017232, "author_name": "ahsuna123", "author_url": "", "post_date": "10/14/2024 16:53:53", "content": "

Hey, @ryanholbrook
\nThank you so much for taking the time and giving a detailed explanation. Much appreciated! :)
\nThis is my first time participating in Jane Street Group's competition and thought this might be a relevant information one might miss similar to my initial assumption of 9 hours.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3017127": "Here is an important note to take, which one might miss:\n\n1. For the CPU and GPU the run-time for the **training phase is <= 8 hours** and not 9 hours (which is the typical case in other competitions). \n2. *TPU is not allowed*. \n\nFor the forecasting purposes, the runtime is extended to 9 hours.", "3017180": "Hi @ahsuna123, this is indeed an important point. The explanation lies in the data we use during the evaluation phase, detailed on the [**Data page**](https://www.kaggle.com/competitions/jane-street-real-time-market-data-forecasting/data). To help ensure continuity between the two phases, we'll be serving the *entire* test data (Public + Private) during the forecasting phase. You'll only need to make predictions on the private set, but processing the unscored portions will still take a bit of time. The extra hour is to give a bit of buffer to protect against time-out failures. It's still important to make your code robust to submission times, however.", "3017232": "Hey, @ryanholbrook \nThank you so much for taking the time and giving a detailed explanation. Much appreciated! :)\nThis is my first time participating in Jane Street Group's competition and thought this might be a relevant information one might miss similar to my initial assumption of 9 hours." }, "source": "meta" }