I tried to math this out and this is what I'm getting. The problem is that I've definitely submitted ensembles that take longer than that, but don't time out. I was hoping someone could spot my math error or tell me what the slowest per iteration my model(s) can be.
\n~700 Clips 1 minute long
\n60s / 5s = 12 segments
\n700 * 12 = 8,400 iterations
\n 90 minutes to run inference (5400 seconds)
\n5400 / 8400 = 0.643 sec/iter (I had this backwards)
Edit (Hopefully for clarity): I was trying to figure out the maximum time each 5 second chunk can take including audio loading, preprocessing, and model prediction.
", "messages": [ { "id": "3179002", "postDate": "04/14/2025 21:50:40", "content": "I tried to math this out and this is what I'm getting. The problem is that I've definitely submitted ensembles that take longer than that, but don't time out. I was hoping someone could spot my math error or tell me what the slowest per iteration my model(s) can be.
\n~700 Clips 1 minute long
\n60s / 5s = 12 segments
\n700 * 12 = 8,400 iterations
\n 90 minutes to run inference (5400 seconds)
\n5400 / 8400 = 0.643 sec/iter (I had this backwards)
Edit (Hopefully for clarity): I was trying to figure out the maximum time each 5 second chunk can take including audio loading, preprocessing, and model prediction.
", "rawMarkdown": "I tried to math this out and this is what I'm getting. The problem is that I've definitely submitted ensembles that take longer than that, but don't time out. I was hoping someone could spot my math error or tell me what the slowest per iteration my model(s) can be.\n\n\n\n~700 Clips 1 minute long\n60s / 5s = 12 segments\n700 * 12 = 8,400 iterations\n 90 minutes to run inference (5400 seconds)\n5400 / 8400 = 0.643 sec/iter (I had this backwards)\n\nEdit (Hopefully for clarity): I was trying to figure out the maximum time each 5 second chunk can take including audio loading, preprocessing, and model prediction.", "votes": null }, { "id": "3179004", "postDate": "04/14/2025 21:55:41", "content": "Loading audio part is just once ? and multiple models are predicting on the same mel spec - right ?
\nIf you were to load audio 5 times and predict 5 times - that might give time out.
", "rawMarkdown": "Loading audio part is just once ? and multiple models are predicting on the same mel spec - right ? \n\nIf you were to load audio 5 times and predict 5 times - that might give time out.", "votes": null }, { "id": "3179018", "postDate": "04/14/2025 22:34:13", "content": "I'm loading each 1 minute clip once and then computing melspec on each 5 second segment. I was trying to come up with the slowest each iteration of my model could be. My observed iterations per seconds (seconds per iteration) is based on the tqdm output while inferring on the train soundscapes.
", "rawMarkdown": "I'm loading each 1 minute clip once and then computing melspec on each 5 second segment. I was trying to come up with the slowest each iteration of my model could be. My observed iterations per seconds (seconds per iteration) is based on the tqdm output while inferring on the train soundscapes.", "votes": null }, { "id": "3179019", "postDate": "04/14/2025 22:36:52", "content": "Oh I see what you mean though. I'm looking for total iteration of the pipeline not necessarily just the model inference part.
", "rawMarkdown": "Oh I see what you mean though. I'm looking for total iteration of the pipeline not necessarily just the model inference part.", "votes": null } ], "comments": [ { "id": 3179004, "author_name": "rashmibanthia", "author_url": "", "post_date": "04/14/2025 21:55:41", "content": "Loading audio part is just once ? and multiple models are predicting on the same mel spec - right ?
\nIf you were to load audio 5 times and predict 5 times - that might give time out.
", "votes": null, "replies": [ { "id": 3179018, "author_name": "willrice", "author_url": "", "post_date": "04/14/2025 22:34:13", "content": "I'm loading each 1 minute clip once and then computing melspec on each 5 second segment. I was trying to come up with the slowest each iteration of my model could be. My observed iterations per seconds (seconds per iteration) is based on the tqdm output while inferring on the train soundscapes.
", "votes": null, "replies": [ { "id": 3179019, "author_name": "willrice", "author_url": "", "post_date": "04/14/2025 22:36:52", "content": "Oh I see what you mean though. I'm looking for total iteration of the pipeline not necessarily just the model inference part.
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3179002": "I tried to math this out and this is what I'm getting. The problem is that I've definitely submitted ensembles that take longer than that, but don't time out. I was hoping someone could spot my math error or tell me what the slowest per iteration my model(s) can be.\n\n\n\n~700 Clips 1 minute long\n60s / 5s = 12 segments\n700 * 12 = 8,400 iterations\n 90 minutes to run inference (5400 seconds)\n5400 / 8400 = 0.643 sec/iter (I had this backwards)\n\nEdit (Hopefully for clarity): I was trying to figure out the maximum time each 5 second chunk can take including audio loading, preprocessing, and model prediction.", "3179004": "Loading audio part is just once ? and multiple models are predicting on the same mel spec - right ? \n\nIf you were to load audio 5 times and predict 5 times - that might give time out.", "3179018": "I'm loading each 1 minute clip once and then computing melspec on each 5 second segment. I was trying to come up with the slowest each iteration of my model could be. My observed iterations per seconds (seconds per iteration) is based on the tqdm output while inferring on the train soundscapes.", "3179019": "Oh I see what you mean though. I'm looking for total iteration of the pipeline not necessarily just the model inference part." }, "source": "meta" }