CV2 and MTCNN pipeline will take me close to the maximum allowed time for predicting the test set. I am extracting all 300 frames from every video and would like to continue doing this preferably.
\n\nShould I be optimising my pipeline or do I need to extract fewer frames?
", "messages": [ { "id": 763863, "postDate": "2020-03-04T23:03:49.133Z", "content": "\npd.pivot_table(ALL_VIDEOS_SUMMARY_DF, index=\"frame_count\", columns=\"label\"\n , values=\"file_name\",aggfunc=len, margins=True, fill_value=0\n ).astype(\"int\")\n
| FrameCount / Label | FAKE |REAL |All|\n| --- | --- | --- | --- |\n| 0 | 8 | 0 | 8 |\n| 83 | 45 | 5 | 50 |\n| 99 | 98 | 7 | 105 | \n| 100 | 32 | 5 | 37 | \n| 109 | 3 | 3 | 6 | \n| 125 | 15 | 5 | 20 | \n| 134 | 10 | 5 | 15 | \n| 136 | 126 | 9 | 135 | \n| 143 | 34 | 5 | 39 | \n| 146 | 35 | 5 | 40 | \n| 147 | 108 | 9 | 117 | \n| 167 | 90 | 5 | 95 | \n| 193 | 13 | 5 | 18 | \n| 236 | 36 | 6 | 42 | \n| 237 | 35 | 5 | 40 | \n| 239 | 139 | 31 | 170 | \n| 240 | 535 | 45 | 580 | \n| 241 | 1832 | 152 | 1984 | \n| 242 | 17 | 5 | 22 | \n| 249 | 119 | 15 | 134 | \n| 250 | 18 | 5 | 23 | \n| 255 | 74 | 4 | 78 | \n| 269 | 144 | 16 | 160 | \n| 270 | 23 | 5 | 28 | \n| 288 | 215 | 11 | 226 | \n| 290 | 15 | 7 | 22 | \n| 292 | 95 | 5 | 100 | \n| 294 | 52 | 5 | 57 | \n| 295 | 25 | 6 | 31 | \n| 296 | 380 | 22 | 402 | \n| 297 | 903 | 69 | 972 | \n| 298 | 3598 | 299 | 3897 | \n| 299 | 2847 | 239 | 3086 | \n| 300 | 80514 | 16998 | 97512 | \n| 301 | 6920 | 981 | 7901 | \n| 302 | 777 | 132 | 909 | \n| 303 | 20 | 13 | 33 | \n| 600 | 35 | 5 | 40\n| 601 | 15 | 5 | 20 | \n| All | 100000 | 19154 | 119154 |
", "rawMarkdown": "```\npd.pivot_table(ALL_VIDEOS_SUMMARY_DF, index=\"frame_count\", columns=\"label\"\n , values=\"file_name\",aggfunc=len, margins=True, fill_value=0\n ).astype(\"int\")\n```\n\n\n| FrameCount / Label \t| FAKE\t|REAL\t|All|\n| --- | --- | --- | --- |\n| 0\t| 8\t| 0\t| 8 |\n| 83 | 45 |\t5 |\t50 |\n| 99 | 98 | \t7 | \t105 | \n| 100 | \t32 | \t5 | \t37 | \n| 109 | \t3 | \t3 | \t6 | \n| 125 | \t15 | \t5 | \t20 | \n| 134 | \t10 | \t5 | \t15 | \n| 136 | \t126 | \t9 | \t135 | \n| 143 | \t34 | \t5 | \t39 | \n| 146 | \t35 | \t5 | \t40 | \n| 147 | \t108 | \t9 | \t117 | \n| 167 | \t90 | \t5 | \t95 | \n| 193 | \t13 | \t5 | \t18 | \n| 236 | \t36 | \t6 | \t42 | \n| 237 | \t35 | \t5 | \t40 | \n| 239 | \t139 | \t31 | \t170 | \n| 240 | \t535 | \t45 | \t580 | \n| 241 | \t1832 | \t152\t | 1984 | \n| 242 | \t17 | \t5 | \t22 | \n| 249 | \t119 | \t15 | \t134 | \n| 250 | \t18 | \t5 | \t23 | \n| 255 | \t74 | \t4 | \t78 | \n| 269 | \t144 | \t16 | \t160 | \n| 270 | \t23 | \t5 | \t28 | \n| 288 | \t215 | \t11 | \t226 | \n| 290 | \t15 | \t7 | \t22 | \n| 292 | \t95 | \t5 | \t100 | \n| 294 | \t52 | \t5 | \t57 | \n| 295 | \t25 | \t6 | \t31 | \n| 296 | \t380 | \t22 | \t402 | \n| 297 | \t903 | \t69 | \t972 | \n| 298 | \t3598 | \t299 | \t3897 | \n| 299 | \t2847 | \t239 | \t3086 | \n| 300 | \t80514 | \t16998 | \t97512 | \n| 301 | \t6920 | \t981 | \t7901 | \n| 302 | \t777 | \t132 | \t909 | \n| 303 | \t20 | \t13 | \t33 | \n| 600 | 35 | \t5 | \t40\n| 601 | 15 | \t5 | \t20 | \n| All\t| 100000 | \t19154 | \t119154 | ", "votes": 3, "replies": [ { "id": 764781, "postDate": "2020-03-05T22:22:54.117Z", "content": "Thanks for this table, very interesting.
", "rawMarkdown": "Thanks for this table, very interesting." } ] }, { "id": 764258, "postDate": "2020-03-05T09:22:26.550Z", "content": "There's also no point doing your face detection on full res images. You can reduce the resolution and then upsample the bounding boxes.
", "rawMarkdown": "There's also no point doing your face detection on full res images. You can reduce the resolution and then upsample the bounding boxes.", "votes": 4, "replies": [ { "id": 764783, "postDate": "2020-03-05T22:24:30.827Z", "content": "great point
", "rawMarkdown": "great point" } ] }, { "id": 763870, "postDate": "2020-03-04T23:17:03.077Z", "content": "Even in the public training set, only 97 K out of 119 K videos have exactly 300 frames (see the pivot table below). Some have as low as 83 frames, while some others have 600+ frames. There is no way to know how many frames your algorithm will have to deal with while processing private test set. So, instead of hard-coding the number of frames I would recommend to treat this # as a parameter of your model and to see how you can fine-tune this parameter to reach a balance between performance and prediction quality.
", "rawMarkdown": "Even in the public training set, only 97 K out of 119 K videos have exactly 300 frames (see the pivot table below). Some have as low as 83 frames, while some others have 600+ frames. There is no way to know how many frames your algorithm will have to deal with while processing private test set. So, instead of hard-coding the number of frames I would recommend to treat this # as a parameter of your model and to see how you can fine-tune this parameter to reach a balance between performance and prediction quality.", "votes": 4 }, { "id": 763744, "postDate": "2020-03-04T20:04:55.973Z", "content": "As suggested by @akashnandi, try getting less frames when predicting. Also, have you tried running your code in parallel? Finally, if your function run time can be optimized by say 10% or more, then you might get under the limit.\nBest of luck!
", "rawMarkdown": "As suggested by @akashnandi, try getting less frames when predicting. Also, have you tried running your code in parallel? Finally, if your function run time can be optimized by say 10% or more, then you might get under the limit.\nBest of luck! ", "votes": 1, "replies": [ { "id": 763751, "postDate": "2020-03-04T20:15:39.410Z", "content": "cheers
", "rawMarkdown": "cheers" } ] }, { "id": 763711, "postDate": "2020-03-04T19:10:44.817Z", "content": "Well, don't extract 300 frames. Try 1 FPS. Should be decent.
", "rawMarkdown": "Well, don't extract 300 frames. Try 1 FPS. Should be decent.", "votes": 2, "replies": [ { "id": 763724, "postDate": "2020-03-04T19:32:53.863Z", "content": "might as well, at least to begin with :p
", "rawMarkdown": "might as well, at least to begin with :p" } ] }, { "id": 764167, "postDate": "2020-03-05T07:41:23.240Z", "content": "I bet even top10 guys don't use 300 frames for inference. 17 Frames can get you to 0.46lb.
", "rawMarkdown": "I bet even top10 guys don't use 300 frames for inference. 17 Frames can get you to 0.46lb. " }, { "id": 763690, "postDate": "2020-03-04T18:39:08.097Z", "content": "CV2 and MTCNN pipeline will take me close to the maximum allowed time for predicting the test set. I am extracting all 300 frames from every video and would like to continue doing this preferably.
\n\nShould I be optimising my pipeline or do I need to extract fewer frames?
", "rawMarkdown": "CV2 and MTCNN pipeline will take me close to the maximum allowed time for predicting the test set. I am extracting all 300 frames from every video and would like to continue doing this preferably.\n\nShould I be optimising my pipeline or do I need to extract fewer frames?" }, { "id": 763834, "postDate": "2020-03-04T22:25:02.217Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 763863, "author_name": "Vlad Pavlov", "author_url": "", "post_date": "2020-03-04T23:03:49.133000", "content": "\npd.pivot_table(ALL_VIDEOS_SUMMARY_DF, index=\"frame_count\", columns=\"label\"\n , values=\"file_name\",aggfunc=len, margins=True, fill_value=0\n ).astype(\"int\")\n
| FrameCount / Label | FAKE |REAL |All|\n| --- | --- | --- | --- |\n| 0 | 8 | 0 | 8 |\n| 83 | 45 | 5 | 50 |\n| 99 | 98 | 7 | 105 | \n| 100 | 32 | 5 | 37 | \n| 109 | 3 | 3 | 6 | \n| 125 | 15 | 5 | 20 | \n| 134 | 10 | 5 | 15 | \n| 136 | 126 | 9 | 135 | \n| 143 | 34 | 5 | 39 | \n| 146 | 35 | 5 | 40 | \n| 147 | 108 | 9 | 117 | \n| 167 | 90 | 5 | 95 | \n| 193 | 13 | 5 | 18 | \n| 236 | 36 | 6 | 42 | \n| 237 | 35 | 5 | 40 | \n| 239 | 139 | 31 | 170 | \n| 240 | 535 | 45 | 580 | \n| 241 | 1832 | 152 | 1984 | \n| 242 | 17 | 5 | 22 | \n| 249 | 119 | 15 | 134 | \n| 250 | 18 | 5 | 23 | \n| 255 | 74 | 4 | 78 | \n| 269 | 144 | 16 | 160 | \n| 270 | 23 | 5 | 28 | \n| 288 | 215 | 11 | 226 | \n| 290 | 15 | 7 | 22 | \n| 292 | 95 | 5 | 100 | \n| 294 | 52 | 5 | 57 | \n| 295 | 25 | 6 | 31 | \n| 296 | 380 | 22 | 402 | \n| 297 | 903 | 69 | 972 | \n| 298 | 3598 | 299 | 3897 | \n| 299 | 2847 | 239 | 3086 | \n| 300 | 80514 | 16998 | 97512 | \n| 301 | 6920 | 981 | 7901 | \n| 302 | 777 | 132 | 909 | \n| 303 | 20 | 13 | 33 | \n| 600 | 35 | 5 | 40\n| 601 | 15 | 5 | 20 | \n| All | 100000 | 19154 | 119154 |
", "votes": 3, "replies": [ { "id": 764781, "author_name": "OptimalFit", "author_url": "", "post_date": "2020-03-05T22:22:54.117000", "content": "Thanks for this table, very interesting.
", "votes": 0, "replies": [] } ] }, { "id": 764258, "author_name": "James Howard", "author_url": "", "post_date": "2020-03-05T09:22:26.550000", "content": "There's also no point doing your face detection on full res images. You can reduce the resolution and then upsample the bounding boxes.
", "votes": 4, "replies": [ { "id": 764783, "author_name": "OptimalFit", "author_url": "", "post_date": "2020-03-05T22:24:30.827000", "content": "great point
", "votes": 0, "replies": [] } ] }, { "id": 763870, "author_name": "Vlad Pavlov", "author_url": "", "post_date": "2020-03-04T23:17:03.077000", "content": "Even in the public training set, only 97 K out of 119 K videos have exactly 300 frames (see the pivot table below). Some have as low as 83 frames, while some others have 600+ frames. There is no way to know how many frames your algorithm will have to deal with while processing private test set. So, instead of hard-coding the number of frames I would recommend to treat this # as a parameter of your model and to see how you can fine-tune this parameter to reach a balance between performance and prediction quality.
", "votes": 4, "replies": [] }, { "id": 763744, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2020-03-04T20:04:55.973000", "content": "As suggested by @akashnandi, try getting less frames when predicting. Also, have you tried running your code in parallel? Finally, if your function run time can be optimized by say 10% or more, then you might get under the limit.\nBest of luck!
", "votes": 1, "replies": [ { "id": 763751, "author_name": "OptimalFit", "author_url": "", "post_date": "2020-03-04T20:15:39.410000", "content": "cheers
", "votes": 0, "replies": [] } ] }, { "id": 763711, "author_name": "Akash", "author_url": "", "post_date": "2020-03-04T19:10:44.817000", "content": "Well, don't extract 300 frames. Try 1 FPS. Should be decent.
", "votes": 2, "replies": [ { "id": 763724, "author_name": "OptimalFit", "author_url": "", "post_date": "2020-03-04T19:32:53.863000", "content": "might as well, at least to begin with :p
", "votes": 0, "replies": [] } ] }, { "id": 764167, "author_name": "petya", "author_url": "", "post_date": "2020-03-05T07:41:23.240000", "content": "I bet even top10 guys don't use 300 frames for inference. 17 Frames can get you to 0.46lb.
", "votes": 0, "replies": [] }, { "id": 763834, "author_name": "", "author_url": "", "post_date": "2020-03-04T22:25:02.217000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "763863": "```\npd.pivot_table(ALL_VIDEOS_SUMMARY_DF, index=\"frame_count\", columns=\"label\"\n , values=\"file_name\",aggfunc=len, margins=True, fill_value=0\n ).astype(\"int\")\n```\n\n\n| FrameCount / Label \t| FAKE\t|REAL\t|All|\n| --- | --- | --- | --- |\n| 0\t| 8\t| 0\t| 8 |\n| 83 | 45 |\t5 |\t50 |\n| 99 | 98 | \t7 | \t105 | \n| 100 | \t32 | \t5 | \t37 | \n| 109 | \t3 | \t3 | \t6 | \n| 125 | \t15 | \t5 | \t20 | \n| 134 | \t10 | \t5 | \t15 | \n| 136 | \t126 | \t9 | \t135 | \n| 143 | \t34 | \t5 | \t39 | \n| 146 | \t35 | \t5 | \t40 | \n| 147 | \t108 | \t9 | \t117 | \n| 167 | \t90 | \t5 | \t95 | \n| 193 | \t13 | \t5 | \t18 | \n| 236 | \t36 | \t6 | \t42 | \n| 237 | \t35 | \t5 | \t40 | \n| 239 | \t139 | \t31 | \t170 | \n| 240 | \t535 | \t45 | \t580 | \n| 241 | \t1832 | \t152\t | 1984 | \n| 242 | \t17 | \t5 | \t22 | \n| 249 | \t119 | \t15 | \t134 | \n| 250 | \t18 | \t5 | \t23 | \n| 255 | \t74 | \t4 | \t78 | \n| 269 | \t144 | \t16 | \t160 | \n| 270 | \t23 | \t5 | \t28 | \n| 288 | \t215 | \t11 | \t226 | \n| 290 | \t15 | \t7 | \t22 | \n| 292 | \t95 | \t5 | \t100 | \n| 294 | \t52 | \t5 | \t57 | \n| 295 | \t25 | \t6 | \t31 | \n| 296 | \t380 | \t22 | \t402 | \n| 297 | \t903 | \t69 | \t972 | \n| 298 | \t3598 | \t299 | \t3897 | \n| 299 | \t2847 | \t239 | \t3086 | \n| 300 | \t80514 | \t16998 | \t97512 | \n| 301 | \t6920 | \t981 | \t7901 | \n| 302 | \t777 | \t132 | \t909 | \n| 303 | \t20 | \t13 | \t33 | \n| 600 | 35 | \t5 | \t40\n| 601 | 15 | \t5 | \t20 | \n| All\t| 100000 | \t19154 | \t119154 | ", "764258": "There's also no point doing your face detection on full res images. You can reduce the resolution and then upsample the bounding boxes.", "763870": "Even in the public training set, only 97 K out of 119 K videos have exactly 300 frames (see the pivot table below). Some have as low as 83 frames, while some others have 600+ frames. There is no way to know how many frames your algorithm will have to deal with while processing private test set. So, instead of hard-coding the number of frames I would recommend to treat this # as a parameter of your model and to see how you can fine-tune this parameter to reach a balance between performance and prediction quality.", "763744": "As suggested by @akashnandi, try getting less frames when predicting. Also, have you tried running your code in parallel? Finally, if your function run time can be optimized by say 10% or more, then you might get under the limit.\nBest of luck! ", "763711": "Well, don't extract 300 frames. Try 1 FPS. Should be decent.", "764167": "I bet even top10 guys don't use 300 frames for inference. 17 Frames can get you to 0.46lb. ", "763690": "CV2 and MTCNN pipeline will take me close to the maximum allowed time for predicting the test set. I am extracting all 300 frames from every video and would like to continue doing this preferably.\n\nShould I be optimising my pipeline or do I need to extract fewer frames?", "763834": "" } }