I have trained a couple of deepfakes that took on average 40-50 hours to train on an 11gb GPU such as GTX 1080 Ti and then swap the faces on each frame in the video. The video was about a minute in length.
\n\nBut I wonder if detecting a deepfake could be equally taxing on the GPUs? Or is that a pretty fast process in terms of computing requirements?
\n\nI think if it's like an inference approach then it shouldn't be that GPU intensive. But would love to know your thoughts.
", "messages": [ { "id": "931687", "postDate": "07/16/2020 11:30:10", "content": "I have trained a couple of deepfakes that took on average 40-50 hours to train on an 11gb GPU such as GTX 1080 Ti and then swap the faces on each frame in the video. The video was about a minute in length.
\n\nBut I wonder if detecting a deepfake could be equally taxing on the GPUs? Or is that a pretty fast process in terms of computing requirements?
\n\nI think if it's like an inference approach then it shouldn't be that GPU intensive. But would love to know your thoughts.
", "rawMarkdown": "I have trained a couple of deepfakes that took on average 40-50 hours to train on an 11gb GPU such as GTX 1080 Ti and then swap the faces on each frame in the video. The video was about a minute in length.\n\nBut I wonder if detecting a deepfake could be equally taxing on the GPUs? Or is that a pretty fast process in terms of computing requirements?\n\nI think if it's like an inference approach then it shouldn't be that GPU intensive. But would love to know your thoughts.", "votes": null }, { "id": "931729", "postDate": "07/16/2020 12:10:36", "content": "Since the competition was limited to 9 hours GPU time, this gives about 8 seconds per video.
\n\nNote: 8 seconds total includes all activities such as reading the video, extracting frames, and testing 20-40 frames (images) on all your ensemble models, and coming up with the final prediction for that video.
", "rawMarkdown": "Since the competition was limited to 9 hours GPU time, this gives about 8 seconds per video.\n\nNote: 8 seconds total includes all activities such as reading the video, extracting frames, and testing 20-40 frames (images) on all your ensemble models, and coming up with the final prediction for that video.", "votes": null }, { "id": "931738", "postDate": "07/16/2020 12:17:37", "content": "Wow. That's quite a lot of computing in a very short amount of time. I'm trying to understand another thing here. Since you mentioned that the GPU time was limited. Did that mean the same GPU for everyone?
", "rawMarkdown": "Wow. That's quite a lot of computing in a very short amount of time. I'm trying to understand another thing here. Since you mentioned that the GPU time was limited. Did that mean the same GPU for everyone?", "votes": null }, { "id": "931792", "postDate": "07/16/2020 12:57:57", "content": "\"9 hours run-time on Kaggle's P100 GPUs\" is mentioned in the project Requirements
", "rawMarkdown": "\"9 hours run-time on Kaggle's P100 GPUs\" is mentioned in the project [Requirements](https://www.kaggle.com/c/deepfake-detection-challenge/overview/code-requirements)", "votes": null }, { "id": "931802", "postDate": "07/16/2020 13:11:04", "content": "Thanks. Forgot to look at that.
", "rawMarkdown": "Thanks. Forgot to look at that.", "votes": null } ], "comments": [ { "id": 931729, "author_name": "sirishks", "author_url": "", "post_date": "07/16/2020 12:10:36", "content": "Since the competition was limited to 9 hours GPU time, this gives about 8 seconds per video.
\n\nNote: 8 seconds total includes all activities such as reading the video, extracting frames, and testing 20-40 frames (images) on all your ensemble models, and coming up with the final prediction for that video.
", "votes": null, "replies": [ { "id": 931738, "author_name": "genesis96839", "author_url": "", "post_date": "07/16/2020 12:17:37", "content": "Wow. That's quite a lot of computing in a very short amount of time. I'm trying to understand another thing here. Since you mentioned that the GPU time was limited. Did that mean the same GPU for everyone?
", "votes": null, "replies": [] }, { "id": 931792, "author_name": "sirishks", "author_url": "", "post_date": "07/16/2020 12:57:57", "content": "\"9 hours run-time on Kaggle's P100 GPUs\" is mentioned in the project Requirements
", "votes": null, "replies": [] }, { "id": 931802, "author_name": "genesis96839", "author_url": "", "post_date": "07/16/2020 13:11:04", "content": "Thanks. Forgot to look at that.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "931687": "I have trained a couple of deepfakes that took on average 40-50 hours to train on an 11gb GPU such as GTX 1080 Ti and then swap the faces on each frame in the video. The video was about a minute in length.\n\nBut I wonder if detecting a deepfake could be equally taxing on the GPUs? Or is that a pretty fast process in terms of computing requirements?\n\nI think if it's like an inference approach then it shouldn't be that GPU intensive. But would love to know your thoughts.", "931729": "Since the competition was limited to 9 hours GPU time, this gives about 8 seconds per video.\n\nNote: 8 seconds total includes all activities such as reading the video, extracting frames, and testing 20-40 frames (images) on all your ensemble models, and coming up with the final prediction for that video.", "931738": "Wow. That's quite a lot of computing in a very short amount of time. I'm trying to understand another thing here. Since you mentioned that the GPU time was limited. Did that mean the same GPU for everyone?", "931792": "\"9 hours run-time on Kaggle's P100 GPUs\" is mentioned in the project [Requirements](https://www.kaggle.com/c/deepfake-detection-challenge/overview/code-requirements)", "931802": "Thanks. Forgot to look at that." }, "source": "meta" }