Whats your best score without using the leak? Since it'll be reset in a few days i think a few LB scores (as well as some public kernels) will be corrected; I'm wondering if anyone (other than @ryches who has been doing an incredible job) was able to get sub 0.6 with a face detection model.
", "messages": [ { "id": 707369, "postDate": "2019-12-31T17:41:09.413Z", "content": "My current score is without any leaks and speech modeling. It's a simple dnn binary classification model trained with a single frame (frame #150) from each video and on a small subset of the total training data.
", "rawMarkdown": "My current score is without any leaks and speech modeling. It's a simple dnn binary classification model trained with a single frame (frame #150) from each video and on a small subset of the total training data.", "votes": 26, "replies": [ { "id": 707392, "postDate": "2019-12-31T18:25:25.280Z", "content": "What resolution are you operating on?
", "rawMarkdown": "What resolution are you operating on? ", "votes": 1 }, { "id": 707403, "postDate": "2019-12-31T18:39:31.903Z", "content": "faces extracted at 300
", "rawMarkdown": "faces extracted at 300", "votes": 1 }, { "id": 707429, "postDate": "2019-12-31T20:02:18.397Z", "content": "Thanks, for sharing your experience Sandeep!
", "rawMarkdown": "Thanks, for sharing your experience Sandeep! ", "votes": 1 }, { "id": 707508, "postDate": "2020-01-01T01:07:49.750Z", "content": "Have you tried different resolutions? I'm curious about the effect that this has on the models but haven't had the opportunity to try yet, I'm keeping mine quiet low so far
", "rawMarkdown": "Have you tried different resolutions? I'm curious about the effect that this has on the models but haven't had the opportunity to try yet, I'm keeping mine quiet low so far" }, { "id": 707509, "postDate": "2020-01-01T01:11:22.937Z", "content": "I think lower resolution will decrease the accuracy of the model. There will less and less detail until the real faces the same blurry as the fake ones.
", "rawMarkdown": "I think lower resolution will decrease the accuracy of the model. There will less and less detail until the real faces the same blurry as the fake ones. " }, { "id": 707510, "postDate": "2020-01-01T01:14:59.073Z", "content": "I haven't tried any other resolutions. Lower resolutions are lossy but higher resolutions are expensive to compute. Given how the videos are, 300x300 is a reasonable box for faces.
", "rawMarkdown": "I haven't tried any other resolutions. Lower resolutions are lossy but higher resolutions are expensive to compute. Given how the videos are, 300x300 is a reasonable box for faces.", "votes": 1 }, { "id": 707515, "postDate": "2020-01-01T01:22:51.080Z", "content": "Unless you use aws. You can use about 500 hours of Tesla V100!(based on what I got: 500 credits, probably will get more in the future)
", "rawMarkdown": "Unless you use aws. You can use about 500 hours of Tesla V100!(based on what I got: 500 credits, probably will get more in the future)" }, { "id": 707525, "postDate": "2020-01-01T01:39:26.450Z", "content": "Thats true but I meant the storage and processing cost and not really the dollar cost.
", "rawMarkdown": "Thats true but I meant the storage and processing cost and not really the dollar cost." }, { "id": 707567, "postDate": "2020-01-01T04:29:51.970Z", "content": "Just to confirm, is your current score with or without speech modeling?
", "rawMarkdown": "Just to confirm, is your current score with or without speech modeling?" }, { "id": 707810, "postDate": "2020-01-01T14:17:13.710Z", "content": "I haven't done any speech modeling yet.
", "rawMarkdown": "I haven't done any speech modeling yet.", "votes": 1 }, { "id": 708094, "postDate": "2020-01-02T01:03:56.157Z", "content": "Sorry to ask a dumb question, but what is a dnn? I am familiar with cnn but not dnn. Id it an acronym for a \"deep neural net\" of unspecified architecture?
", "rawMarkdown": "Sorry to ask a dumb question, but what is a dnn? I am familiar with cnn but not dnn. Id it an acronym for a \"deep neural net\" of unspecified architecture?" }, { "id": 708100, "postDate": "2020-01-02T01:32:49.407Z", "content": "You are right and I should have made it clearer. A deep neural net with just some cnn and fully connected layers.
", "rawMarkdown": "You are right and I should have made it clearer. A deep neural net with just some cnn and fully connected layers.", "votes": 1 }, { "id": 708437, "postDate": "2020-01-02T09:56:10.293Z", "content": "faces extracted at 300? Did you mean the faces were extracted from frames with resolution 300x300, or you extracted faces from original frames and resized them into 300x300?
", "rawMarkdown": "faces extracted at 300? Did you mean the faces were extracted from frames with resolution 300x300, or you extracted faces from original frames and resized them into 300x300?", "votes": 1 }, { "id": 708677, "postDate": "2020-01-02T14:40:15.083Z", "content": "extracted from original frames and resized to 300
", "rawMarkdown": "extracted from original frames and resized to 300" }, { "id": 708935, "postDate": "2020-01-02T20:59:07.560Z", "content": "I'm interested in how you might have dealt with videos that have more than one face?
", "rawMarkdown": "I'm interested in how you might have dealt with videos that have more than one face? ", "votes": 1 }, { "id": 708962, "postDate": "2020-01-02T21:56:47.080Z", "content": "My current pipeline is bare minimum. I only retain the face with highest probability and only if the probability is greater than 0.95.
", "rawMarkdown": "My current pipeline is bare minimum. I only retain the face with highest probability and only if the probability is greater than 0.95.", "votes": 2 }, { "id": 708984, "postDate": "2020-01-02T22:25:03.563Z", "content": "Hi Sandeep,\n Thanks for your valuable information. May I ask how you \"retain the face with highest probability and only if the probability is greater than 0.95\"? Also, how large is you training set right now?
", "rawMarkdown": "Hi Sandeep,\n Thanks for your valuable information. May I ask how you \"retain the face with highest probability and only if the probability is greater than 0.95\"? Also, how large is you training set right now?\n" }, { "id": 708985, "postDate": "2020-01-02T22:27:32.753Z", "content": "Hi Sandeep, do you also sample 150 frames per test video? I use cv2.cascade classifier to detect & extract faces in 10 frames/video, and my submission kernel runs very slow....
", "rawMarkdown": "Hi Sandeep, do you also sample 150 frames per test video? I use cv2.cascade classifier to detect & extract faces in 10 frames/video, and my submission kernel runs very slow...." }, { "id": 708992, "postDate": "2020-01-02T22:36:01.250Z", "content": "training set - roughly 15k videos.\nface probabilities - I'm using mtcnn for face detection and it takes an optional argument to return probabilities along with the detected faces.\nframes per video - I'm sampling exactly 1 frame per video and it's the frame from the middle of the video stream.
", "rawMarkdown": "training set - roughly 15k videos.\nface probabilities - I'm using mtcnn for face detection and it takes an optional argument to return probabilities along with the detected faces.\nframes per video - I'm sampling exactly 1 frame per video and it's the frame from the middle of the video stream.", "votes": 1 }, { "id": 709014, "postDate": "2020-01-02T23:46:33.803Z", "content": "Hi Sandeep, thank you for your info!\nWas wondering if you could share how you do the inference at the moment? Do you also take the middle frame of the video of the testing data?
", "rawMarkdown": "Hi Sandeep, thank you for your info!\nWas wondering if you could share how you do the inference at the moment? Do you also take the middle frame of the video of the testing data?" }, { "id": 709034, "postDate": "2020-01-03T00:44:47.740Z", "content": "@sattree Yeah! You're right.
", "rawMarkdown": "@sattree Yeah! You're right." }, { "id": 709043, "postDate": "2020-01-03T01:02:47.363Z", "content": "@rafiko1 yes it's the same process.\n@unkownhihi I understand your concern but there's no secret to it. It's all pretty basic as of now.
", "rawMarkdown": "@rafiko1 yes it's the same process.\n@unkownhihi I understand your concern but there's no secret to it. It's all pretty basic as of now.", "votes": 1 }, { "id": 709138, "postDate": "2020-01-03T04:34:06.047Z", "content": "Hi Sandeep, \n Great thanks for your reply! Let me try to do it and see if I can get a better score.
\n\nI also tried the package \"face_recognition\" mentioned in the kernel by someone else, but I failed to install it in the submission kernel and never finish running it. Did you successfully \"pip install mtcnn\" in your submission kernel?
\n\nThanks!
", "rawMarkdown": "Hi Sandeep, \n Great thanks for your reply! Let me try to do it and see if I can get a better score.\n\n I also tried the package \"face_recognition\" mentioned in the kernel by someone else, but I failed to install it in the submission kernel and never finish running it. Did you successfully \"pip install mtcnn\" in your submission kernel?\n\n Thanks!\n" }, { "id": 709419, "postDate": "2020-01-03T13:43:11.343Z", "content": "Here's a great kernel https://www.kaggle.com/timesler/facial-recognition-model-in-pytorch to get started. I'm using timesler's implementation of mtcnn in facenet_pytorch.
", "rawMarkdown": "Here's a great kernel https://www.kaggle.com/timesler/facial-recognition-model-in-pytorch to get started. I'm using timesler's implementation of mtcnn in facenet_pytorch.", "votes": 2 }, { "id": 710190, "postDate": "2020-01-04T12:33:30.353Z", "content": "@sattree did you just use the model's raw predictions or did you tweak the threshold?
", "rawMarkdown": "@sattree did you just use the model's raw predictions or did you tweak the threshold?" }, { "id": 710290, "postDate": "2020-01-04T14:51:19.513Z", "content": "no tweaking on the threshold.
", "rawMarkdown": "no tweaking on the threshold." }, { "id": 711382, "postDate": "2020-01-06T02:12:44.130Z", "content": "@sattree Have you set the margin parameter of MTCNN ? Or directly use the detected face area to train?
", "rawMarkdown": "@sattree Have you set the margin parameter of MTCNN ? Or directly use the detected face area to train?" }, { "id": 711430, "postDate": "2020-01-06T04:53:02.757Z", "content": "margin=16
", "rawMarkdown": "margin=16", "votes": 1 }, { "id": 711923, "postDate": "2020-01-06T17:01:27.453Z", "rawMarkdown": "" } ] }, { "id": 704630, "postDate": "2019-12-27T18:42:51.533Z", "content": "Whats your best score without using the leak? Since it'll be reset in a few days i think a few LB scores (as well as some public kernels) will be corrected; I'm wondering if anyone (other than @ryches who has been doing an incredible job) was able to get sub 0.6 with a face detection model.
", "rawMarkdown": "Whats your best score without using the leak? Since it'll be reset in a few days i think a few LB scores (as well as some public kernels) will be corrected; I'm wondering if anyone (other than @ryches who has been doing an incredible job) was able to get sub 0.6 with a face detection model.", "votes": 24 }, { "id": 706910, "postDate": "2019-12-31T01:48:23.283Z", "content": "Pretty interesting no one else has reacted to this thread. We'll see what happens after the leak is removed.
", "rawMarkdown": "Pretty interesting no one else has reacted to this thread. We'll see what happens after the leak is removed. ", "votes": 3 }, { "id": 707201, "postDate": "2019-12-31T12:03:10.277Z", "content": "On my validation set of size 4000 (2000 Fake + 2000 Real), I got a loss (BCE) of 0.34 and an accuracy of 89%. I am trying to submit my kernel, but in the last 24 hrs, I have made two unnecessary submissions and I am not able to submit my solution right now. I will let you know once the submission is done. Let's see what the score will be on the public test set.\nOn those 400 visible test videos, the model is doing very good (accuracy wise). Although, the model is doing kinda bad on Real videos of one particular person in this set.
", "rawMarkdown": "On my validation set of size 4000 (2000 Fake + 2000 Real), I got a loss (BCE) of 0.34 and an accuracy of 89%. I am trying to submit my kernel, but in the last 24 hrs, I have made two unnecessary submissions and I am not able to submit my solution right now. I will let you know once the submission is done. Let's see what the score will be on the public test set.\nOn those 400 visible test videos, the model is doing very good (accuracy wise). Although, the model is doing kinda bad on Real videos of one particular person in this set.", "votes": 2 }, { "id": 707147, "postDate": "2019-12-31T10:24:56.787Z", "content": "I'm in the same boat. Until now, all my NN face classification models converge to a 0.5 probability. The only ones which do not are severely overfitting.
", "rawMarkdown": "I'm in the same boat. Until now, all my NN face classification models converge to a 0.5 probability. The only ones which do not are severely overfitting.", "votes": 1 }, { "id": 706736, "postDate": "2019-12-30T18:54:02.017Z", "content": "My best without the leak is .69294. I've since trained my model to a better confidence ( 4 more NN layers on top of a Conv network ), but I'm having trouble getting a score with the new model. The original method used 6 frames from the video, but the new model seems to time out with that number of frames.
", "rawMarkdown": "My best without the leak is .69294. I've since trained my model to a better confidence ( 4 more NN layers on top of a Conv network ), but I'm having trouble getting a score with the new model. The original method used 6 frames from the video, but the new model seems to time out with that number of frames.", "votes": 1, "replies": [ { "id": 706851, "postDate": "2019-12-30T23:16:54.037Z", "content": "At least it is better than all 0.5 loss. My model can't even be better than .69314(all 0.5 loss) and is, unfortunately, learning towards to predict all 0.5 as I add more data and train more epochs. Need improvement. 😦
", "rawMarkdown": "At least it is better than all 0.5 loss. My model can't even be better than .69314(all 0.5 loss) and is, unfortunately, learning towards to predict all 0.5 as I add more data and train more epochs. Need improvement. 😦 " }, { "id": 708982, "postDate": "2020-01-02T22:17:42.400Z", "content": "Hi Jon, when you say Conv model, can you specify which network? I tried one DenseNet backbone with a new head, it seems not working very well...
\n\nThanks
", "rawMarkdown": "Hi Jon, when you say Conv model, can you specify which network? I tried one DenseNet backbone with a new head, it seems not working very well...\n\nThanks" } ] }, { "id": 712205, "postDate": "2020-01-07T00:04:18.070Z", "content": "I got a better score. 0.69203. Just a little bit better than 0.5 but still a big improvement. I made the notebook public. link \nThe only version that got that score is version 15.
", "rawMarkdown": "I got a better score. 0.69203. Just a little bit better than 0.5 but still a big improvement. I made the notebook public. [link ](https://www.kaggle.com/unkownhihi/starter-kernel-with-cnn-ll-lb-0-69306-no-leak)\nThe only version that got that score is version 15.", "votes": 2, "replies": [ { "id": 716000, "postDate": "2020-01-11T06:05:49.037Z", "content": "Hi, thank you very much for sharing, I am using your kernel as a starting point.
", "rawMarkdown": "Hi, thank you very much for sharing, I am using your kernel as a starting point. " } ] }, { "id": 707897, "postDate": "2020-01-01T16:42:27.710Z", "content": "My score is clean, FYI
", "rawMarkdown": "My score is clean, FYI", "votes": 2 }, { "id": 711380, "postDate": "2020-01-06T02:12:24.760Z", "rawMarkdown": "", "isDeleted": true }, { "id": 705270, "postDate": "2019-12-28T17:42:19.567Z", "rawMarkdown": "", "votes": 1, "isDeleted": true }, { "id": 708188, "postDate": "2020-01-02T04:00:28.930Z", "content": "Thanks for the excellent post
", "rawMarkdown": "Thanks for the excellent post \n\n" } ], "comments": [ { "id": 707369, "author_name": "Sandeep Attree", "author_url": "", "post_date": "2019-12-31T17:41:09.413000", "content": "My current score is without any leaks and speech modeling. It's a simple dnn binary classification model trained with a single frame (frame #150) from each video and on a small subset of the total training data.
", "votes": 26, "replies": [ { "id": 707392, "author_name": "ryches", "author_url": "", "post_date": "2019-12-31T18:25:25.280000", "content": "What resolution are you operating on?
", "votes": 1, "replies": [] }, { "id": 707403, "author_name": "Sandeep Attree", "author_url": "", "post_date": "2019-12-31T18:39:31.903000", "content": "faces extracted at 300
", "votes": 1, "replies": [] }, { "id": 707429, "author_name": "Debanga Raj Neog", "author_url": "", "post_date": "2019-12-31T20:02:18.397000", "content": "Thanks, for sharing your experience Sandeep!
", "votes": 1, "replies": [] }, { "id": 707508, "author_name": "Pedro Bernardo", "author_url": "", "post_date": "2020-01-01T01:07:49.750000", "content": "Have you tried different resolutions? I'm curious about the effect that this has on the models but haven't had the opportunity to try yet, I'm keeping mine quiet low so far
", "votes": 0, "replies": [] }, { "id": 707509, "author_name": "Shangqiu Li", "author_url": "", "post_date": "2020-01-01T01:11:22.937000", "content": "I think lower resolution will decrease the accuracy of the model. There will less and less detail until the real faces the same blurry as the fake ones.
", "votes": 0, "replies": [] }, { "id": 707510, "author_name": "Sandeep Attree", "author_url": "", "post_date": "2020-01-01T01:14:59.073000", "content": "I haven't tried any other resolutions. Lower resolutions are lossy but higher resolutions are expensive to compute. Given how the videos are, 300x300 is a reasonable box for faces.
", "votes": 1, "replies": [] }, { "id": 707515, "author_name": "Shangqiu Li", "author_url": "", "post_date": "2020-01-01T01:22:51.080000", "content": "Unless you use aws. You can use about 500 hours of Tesla V100!(based on what I got: 500 credits, probably will get more in the future)
", "votes": 0, "replies": [] }, { "id": 707525, "author_name": "Sandeep Attree", "author_url": "", "post_date": "2020-01-01T01:39:26.450000", "content": "Thats true but I meant the storage and processing cost and not really the dollar cost.
", "votes": 0, "replies": [] }, { "id": 707567, "author_name": "Manideep", "author_url": "", "post_date": "2020-01-01T04:29:51.970000", "content": "Just to confirm, is your current score with or without speech modeling?
", "votes": 0, "replies": [] }, { "id": 707810, "author_name": "Sandeep Attree", "author_url": "", "post_date": "2020-01-01T14:17:13.710000", "content": "I haven't done any speech modeling yet.
", "votes": 1, "replies": [] }, { "id": 708094, "author_name": "pete", "author_url": "", "post_date": "2020-01-02T01:03:56.157000", "content": "Sorry to ask a dumb question, but what is a dnn? I am familiar with cnn but not dnn. Id it an acronym for a \"deep neural net\" of unspecified architecture?
", "votes": 0, "replies": [] }, { "id": 708100, "author_name": "Sandeep Attree", "author_url": "", "post_date": "2020-01-02T01:32:49.407000", "content": "You are right and I should have made it clearer. A deep neural net with just some cnn and fully connected layers.
", "votes": 1, "replies": [] }, { "id": 708437, "author_name": "xxn-xx", "author_url": "", "post_date": "2020-01-02T09:56:10.293000", "content": "faces extracted at 300? Did you mean the faces were extracted from frames with resolution 300x300, or you extracted faces from original frames and resized them into 300x300?
", "votes": 1, "replies": [] }, { "id": 708677, "author_name": "Sandeep Attree", "author_url": "", "post_date": "2020-01-02T14:40:15.083000", "content": "extracted from original frames and resized to 300
", "votes": 0, "replies": [] }, { "id": 708935, "author_name": "ryches", "author_url": "", "post_date": "2020-01-02T20:59:07.560000", "content": "I'm interested in how you might have dealt with videos that have more than one face?
", "votes": 1, "replies": [] }, { "id": 708962, "author_name": "Sandeep Attree", "author_url": "", "post_date": "2020-01-02T21:56:47.080000", "content": "My current pipeline is bare minimum. I only retain the face with highest probability and only if the probability is greater than 0.95.
", "votes": 2, "replies": [] }, { "id": 708984, "author_name": "", "author_url": "", "post_date": "2020-01-02T22:25:03.563000", "content": "Hi Sandeep,\n Thanks for your valuable information. May I ask how you \"retain the face with highest probability and only if the probability is greater than 0.95\"? Also, how large is you training set right now?
", "votes": 0, "replies": [] }, { "id": 708985, "author_name": "", "author_url": "", "post_date": "2020-01-02T22:27:32.753000", "content": "Hi Sandeep, do you also sample 150 frames per test video? I use cv2.cascade classifier to detect & extract faces in 10 frames/video, and my submission kernel runs very slow....
", "votes": 0, "replies": [] }, { "id": 708992, "author_name": "Sandeep Attree", "author_url": "", "post_date": "2020-01-02T22:36:01.250000", "content": "training set - roughly 15k videos.\nface probabilities - I'm using mtcnn for face detection and it takes an optional argument to return probabilities along with the detected faces.\nframes per video - I'm sampling exactly 1 frame per video and it's the frame from the middle of the video stream.
", "votes": 1, "replies": [] }, { "id": 709014, "author_name": "Rafi Hai", "author_url": "", "post_date": "2020-01-02T23:46:33.803000", "content": "Hi Sandeep, thank you for your info!\nWas wondering if you could share how you do the inference at the moment? Do you also take the middle frame of the video of the testing data?
", "votes": 0, "replies": [] }, { "id": 709034, "author_name": "Shangqiu Li", "author_url": "", "post_date": "2020-01-03T00:44:47.740000", "content": "@sattree Yeah! You're right.
", "votes": 0, "replies": [] }, { "id": 709043, "author_name": "Sandeep Attree", "author_url": "", "post_date": "2020-01-03T01:02:47.363000", "content": "@rafiko1 yes it's the same process.\n@unkownhihi I understand your concern but there's no secret to it. It's all pretty basic as of now.
", "votes": 1, "replies": [] }, { "id": 709138, "author_name": "", "author_url": "", "post_date": "2020-01-03T04:34:06.047000", "content": "Hi Sandeep, \n Great thanks for your reply! Let me try to do it and see if I can get a better score.
\n\nI also tried the package \"face_recognition\" mentioned in the kernel by someone else, but I failed to install it in the submission kernel and never finish running it. Did you successfully \"pip install mtcnn\" in your submission kernel?
\n\nThanks!
", "votes": 0, "replies": [] }, { "id": 709419, "author_name": "Sandeep Attree", "author_url": "", "post_date": "2020-01-03T13:43:11.343000", "content": "Here's a great kernel https://www.kaggle.com/timesler/facial-recognition-model-in-pytorch to get started. I'm using timesler's implementation of mtcnn in facenet_pytorch.
", "votes": 2, "replies": [] }, { "id": 710190, "author_name": "Polychronis Charitidis", "author_url": "", "post_date": "2020-01-04T12:33:30.353000", "content": "@sattree did you just use the model's raw predictions or did you tweak the threshold?
", "votes": 0, "replies": [] }, { "id": 710290, "author_name": "Sandeep Attree", "author_url": "", "post_date": "2020-01-04T14:51:19.513000", "content": "no tweaking on the threshold.
", "votes": 0, "replies": [] }, { "id": 711382, "author_name": "Chason", "author_url": "", "post_date": "2020-01-06T02:12:44.130000", "content": "@sattree Have you set the margin parameter of MTCNN ? Or directly use the detected face area to train?
", "votes": 0, "replies": [] }, { "id": 711430, "author_name": "Sandeep Attree", "author_url": "", "post_date": "2020-01-06T04:53:02.757000", "content": "margin=16
", "votes": 1, "replies": [] }, { "id": 711923, "author_name": "Sumukh Aithal ", "author_url": "", "post_date": "2020-01-06T17:01:27.453000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 706910, "author_name": "ryches", "author_url": "", "post_date": "2019-12-31T01:48:23.283000", "content": "Pretty interesting no one else has reacted to this thread. We'll see what happens after the leak is removed.
", "votes": 3, "replies": [] }, { "id": 707201, "author_name": "Manideep", "author_url": "", "post_date": "2019-12-31T12:03:10.277000", "content": "On my validation set of size 4000 (2000 Fake + 2000 Real), I got a loss (BCE) of 0.34 and an accuracy of 89%. I am trying to submit my kernel, but in the last 24 hrs, I have made two unnecessary submissions and I am not able to submit my solution right now. I will let you know once the submission is done. Let's see what the score will be on the public test set.\nOn those 400 visible test videos, the model is doing very good (accuracy wise). Although, the model is doing kinda bad on Real videos of one particular person in this set.
", "votes": 2, "replies": [] }, { "id": 707147, "author_name": "dagnelies", "author_url": "", "post_date": "2019-12-31T10:24:56.787000", "content": "I'm in the same boat. Until now, all my NN face classification models converge to a 0.5 probability. The only ones which do not are severely overfitting.
", "votes": 1, "replies": [] }, { "id": 706736, "author_name": "Jon Macpherson", "author_url": "", "post_date": "2019-12-30T18:54:02.017000", "content": "My best without the leak is .69294. I've since trained my model to a better confidence ( 4 more NN layers on top of a Conv network ), but I'm having trouble getting a score with the new model. The original method used 6 frames from the video, but the new model seems to time out with that number of frames.
", "votes": 1, "replies": [ { "id": 706851, "author_name": "Shangqiu Li", "author_url": "", "post_date": "2019-12-30T23:16:54.037000", "content": "At least it is better than all 0.5 loss. My model can't even be better than .69314(all 0.5 loss) and is, unfortunately, learning towards to predict all 0.5 as I add more data and train more epochs. Need improvement. 😦
", "votes": 0, "replies": [] }, { "id": 708982, "author_name": "", "author_url": "", "post_date": "2020-01-02T22:17:42.400000", "content": "Hi Jon, when you say Conv model, can you specify which network? I tried one DenseNet backbone with a new head, it seems not working very well...
\n\nThanks
", "votes": 0, "replies": [] } ] }, { "id": 712205, "author_name": "Shangqiu Li", "author_url": "", "post_date": "2020-01-07T00:04:18.070000", "content": "I got a better score. 0.69203. Just a little bit better than 0.5 but still a big improvement. I made the notebook public. link \nThe only version that got that score is version 15.
", "votes": 2, "replies": [ { "id": 716000, "author_name": "Owen Xu", "author_url": "", "post_date": "2020-01-11T06:05:49.037000", "content": "Hi, thank you very much for sharing, I am using your kernel as a starting point.
", "votes": 0, "replies": [] } ] }, { "id": 707897, "author_name": "notbad", "author_url": "", "post_date": "2020-01-01T16:42:27.710000", "content": "My score is clean, FYI
", "votes": 2, "replies": [] }, { "id": 711380, "author_name": "", "author_url": "", "post_date": "2020-01-06T02:12:24.760000", "content": "", "votes": 0, "replies": [] }, { "id": 705270, "author_name": "", "author_url": "", "post_date": "2019-12-28T17:42:19.567000", "content": "", "votes": 1, "replies": [] }, { "id": 708188, "author_name": "Ahmed Mohamed", "author_url": "", "post_date": "2020-01-02T04:00:28.930000", "content": "Thanks for the excellent post
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "707369": "My current score is without any leaks and speech modeling. It's a simple dnn binary classification model trained with a single frame (frame #150) from each video and on a small subset of the total training data.", "704630": "Whats your best score without using the leak? Since it'll be reset in a few days i think a few LB scores (as well as some public kernels) will be corrected; I'm wondering if anyone (other than @ryches who has been doing an incredible job) was able to get sub 0.6 with a face detection model.", "706910": "Pretty interesting no one else has reacted to this thread. We'll see what happens after the leak is removed. ", "707201": "On my validation set of size 4000 (2000 Fake + 2000 Real), I got a loss (BCE) of 0.34 and an accuracy of 89%. I am trying to submit my kernel, but in the last 24 hrs, I have made two unnecessary submissions and I am not able to submit my solution right now. I will let you know once the submission is done. Let's see what the score will be on the public test set.\nOn those 400 visible test videos, the model is doing very good (accuracy wise). Although, the model is doing kinda bad on Real videos of one particular person in this set.", "707147": "I'm in the same boat. Until now, all my NN face classification models converge to a 0.5 probability. The only ones which do not are severely overfitting.", "706736": "My best without the leak is .69294. I've since trained my model to a better confidence ( 4 more NN layers on top of a Conv network ), but I'm having trouble getting a score with the new model. The original method used 6 frames from the video, but the new model seems to time out with that number of frames.", "712205": "I got a better score. 0.69203. Just a little bit better than 0.5 but still a big improvement. I made the notebook public. [link ](https://www.kaggle.com/unkownhihi/starter-kernel-with-cnn-ll-lb-0-69306-no-leak)\nThe only version that got that score is version 15.", "707897": "My score is clean, FYI", "711380": "", "705270": "", "708188": "Thanks for the excellent post \n\n" } }