And then I got a score of 0.01 on the leaderboard, which proves that Kaggle's dataset is converted from Stanford Dogs Dataset.
", "messages": [ { "id": "228870", "postDate": "10/08/2017 05:43:34", "content": "And then I got a score of 0.01 on the leaderboard, which proves that Kaggle's dataset is converted from Stanford Dogs Dataset.
", "rawMarkdown": "And then I got a score of 0.01 on the leaderboard, which proves that Kaggle's dataset is converted from Stanford Dogs Dataset.", "votes": null }, { "id": "228996", "postDate": "10/08/2017 14:11:45", "content": "在吗
", "rawMarkdown": "在吗", "votes": null }, { "id": "229042", "postDate": "10/08/2017 15:46:29", "content": "Of course this dataset is from Stanford Dogs Dataset. C&P from description: \"We extend our gratitude to the creators of the Stanford Dogs Dataset for making this competition possible: Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li.\"\nJust wondering, how did you manage to get 0.01? What model did give you best results? Did you use pretrained imagenet weights?
", "rawMarkdown": "Of course this dataset is from Stanford Dogs Dataset. C&P from description: \"We extend our gratitude to the creators of the Stanford Dogs Dataset for making this competition possible: Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li.\"\nJust wondering, how did you manage to get 0.01? What model did give you best results? Did you use pretrained imagenet weights?", "votes": null }, { "id": "229056", "postDate": "10/08/2017 16:31:27", "content": "I used five pre-trained models to predict features and connect them together, then I trained a simple classifier.
", "rawMarkdown": "I used five pre-trained models to predict features and connect them together, then I trained a simple classifier.", "votes": null }, { "id": "229080", "postDate": "10/08/2017 17:42:24", "content": "Since both the train and test set came from the Standord Dogs it is easy to train overfitted model on them. It is not really external data...
", "rawMarkdown": "Since both the train and test set came from the Standord Dogs it is easy to train overfitted model on them. It is not really external data...", "votes": null }, { "id": "229098", "postDate": "10/08/2017 18:42:49", "content": "I agree. Clever thinking by combining predictions from all dnn's! I think this is something that I wouldn't even think of, that's too far out of my scope :)
\nThere is one thing that confuses me: advancements are possible only if we find general solution and not exploits on datasets (although exploit exploration is good motivator for Kaggle team and organizers to prepare better datasets in future).
Is it common on Kaggle that datasets are \"poorly\" prepared? Is it common from Kagglers searching for dataset exploits or eval metrics exploits? I don't think there is something wrong about it, I just want to know what makes this community tick. I'm here last 3 weeks and I see that pattern, wondering if this actually makes sense.
\n\nI'm hoping for the future of better NN, advancements from collaboration of different minds is probably best strategy to get there.
", "rawMarkdown": "I agree. Clever thinking by combining predictions from all dnn's! I think this is something that I wouldn't even think of, that's too far out of my scope :) \nThere is one thing that confuses me: advancements are possible only if we find general solution and not exploits on datasets (although exploit exploration is good motivator for Kaggle team and organizers to prepare better datasets in future). \n \nIs it common on Kaggle that datasets are \"poorly\" prepared? Is it common from Kagglers searching for dataset exploits or eval metrics exploits? I don't think there is something wrong about it, I just want to know what makes this community tick. I'm here last 3 weeks and I see that pattern, wondering if this actually makes sense. \n \nI'm hoping for the future of better NN, advancements from collaboration of different minds is probably best strategy to get there.", "votes": null }, { "id": "229119", "postDate": "10/08/2017 21:13:48", "content": "I, too, have noticed people looking for exploits and shortcuts. Such explorations have their own value, but in this context they amount to cheating at solitaire.
", "rawMarkdown": "I, too, have noticed people looking for exploits and shortcuts. Such explorations have their own value, but in this context they amount to cheating at solitaire.", "votes": null }, { "id": "229145", "postDate": "10/09/2017 01:00:06", "content": "你好 我是一名在校的研究生 可以跟你组个队吗
", "rawMarkdown": "你好 我是一名在校的研究生 可以跟你组个队吗", "votes": null }, { "id": "229146", "postDate": "10/09/2017 01:04:39", "content": "Yes.
", "rawMarkdown": "Yes.", "votes": null }, { "id": "229266", "postDate": "10/09/2017 08:35:55", "content": "Exploiting is not the same as cheating. In this case cheating would be hand labeling and I will not take a second to comment on cheating. Yang used valid method for proving that test set are from same distribution of imagenet. I think this is just amazing way to prove how DNN arhitectures work when they are combined.
", "rawMarkdown": "Exploiting is not the same as cheating. In this case cheating would be hand labeling and I will not take a second to comment on cheating. Yang used valid method for proving that test set are from same distribution of imagenet. I think this is just amazing way to prove how DNN arhitectures work when they are combined.", "votes": null }, { "id": "229477", "postDate": "10/09/2017 17:49:48", "content": "I apologize for being unclear. I did not mean to accuse Yang Peiwen of cheating. I was comparing this educational competition to the game of Solitaire to suggest that using an exploit here is self-defeating.
\n\nSolitaire is an algorithmic sorting game for one player. It starts with a randomly-ordered deck of playing cards. The goal is to sort the deck by applying a limited set of discrete transformations. You win by completely sorting the deck. You lose if you reach a halting state (in which no permitted transformation will add order) prior to completely sorting the deck. The challenge is to minimize the number of moves required to win by choosing the best transformation at each intermediate state.
\n\nThis playground competition is an opportunity to practise specific skills under similarly artificial constraints. Like Solitaire, this is a single-player game. The leaderboard provides motivation and a sense of what kind of improvement is possible; but we're competing against ourselves, so the only way to \"win\" is to get better.
", "rawMarkdown": "I apologize for being unclear. I did not mean to accuse Yang Peiwen of cheating. I was comparing this educational competition to the game of Solitaire to suggest that using an exploit here is self-defeating.\n\nSolitaire is an algorithmic sorting game for one player. It starts with a randomly-ordered deck of playing cards. The goal is to sort the deck by applying a limited set of discrete transformations. You win by completely sorting the deck. You lose if you reach a halting state (in which no permitted transformation will add order) prior to completely sorting the deck. The challenge is to minimize the number of moves required to win by choosing the best transformation at each intermediate state. \n\nThis playground competition is an opportunity to practise specific skills under similarly artificial constraints. Like Solitaire, this is a single-player game. The leaderboard provides motivation and a sense of what kind of improvement is possible; but we're competing against ourselves, so the only way to \"win\" is to get better.", "votes": null }, { "id": "230180", "postDate": "10/11/2017 12:58:49", "content": "huo zhuo pei Shen ,hia hia
", "rawMarkdown": "huo zhuo pei Shen ,hia hia", "votes": null }, { "id": "249068", "postDate": "11/27/2017 16:44:25", "content": "Did you use FC layers to grab all the features of those models and output a classification score or another classifier?
", "rawMarkdown": "Did you use FC layers to grab all the features of those models and output a classification score or another classifier?", "votes": null }, { "id": "250904", "postDate": "11/30/2017 13:11:07", "content": "Are you running this on a gpu?
", "rawMarkdown": "Are you running this on a gpu?", "votes": null }, { "id": "251323", "postDate": "12/01/2017 01:57:43", "content": "Did you change all layers's weights in your fine-tune strategy when you train on this new task?? I tried to fine-tune many pre-trained models ,such as Resnet,Densenet.., but my validation(3000 photos) accuracy just got a little over 81%.
", "rawMarkdown": "Did you change all layers's weights in your fine-tune strategy when you train on this new task?? I tried to fine-tune many pre-trained models ,such as Resnet,Densenet.., but my validation(3000 photos) accuracy just got a little over 81%.", "votes": null }, { "id": "251400", "postDate": "12/01/2017 05:46:01", "content": "Come on guys. First one with 0 loss gonna shine forever.
", "rawMarkdown": "Come on guys. First one with 0 loss gonna shine forever.", "votes": null }, { "id": "253143", "postDate": "12/04/2017 14:09:38", "content": "Impossible, because there are many identical pictures in the training set and test set, but they are different classes, so model can not accurately determine the true labels of these pictures, and can only be submitted by multiple attempts to guess the result.
", "rawMarkdown": "Impossible, because there are many identical pictures in the training set and test set, but they are different classes, so model can not accurately determine the true labels of these pictures, and can only be submitted by multiple attempts to guess the result.", "votes": null }, { "id": "277152", "postDate": "02/02/2018 09:30:18", "content": "how about InceptionResNetV2?
", "rawMarkdown": "how about InceptionResNetV2?", "votes": null }, { "id": "287375", "postDate": "02/24/2018 15:23:27", "content": "it seems, it was possible. :(
", "rawMarkdown": "it seems, it was possible. :(", "votes": null }, { "id": "288276", "postDate": "02/27/2018 04:46:09", "content": "is it possible to get log loss less then 0.3 using kaggle dataset without stanford dataset
", "rawMarkdown": "is it possible to get log loss less then 0.3 using kaggle dataset without stanford dataset", "votes": null }, { "id": "288279", "postDate": "02/27/2018 04:51:56", "content": "Sure, inceptionv3 + resnet152_v1 can get 0.27143 on Public Leaderboard.
\n\nhttps://github.com/ypwhs/DogBreed_gluon
", "rawMarkdown": "Sure, inceptionv3 + resnet152_v1 can get 0.27143 on Public Leaderboard.\n\nhttps://github.com/ypwhs/DogBreed_gluon", "votes": null }, { "id": "288280", "postDate": "02/27/2018 04:52:27", "content": "Yes, my score ~0.14 is a result of using only kaggle train set.
", "rawMarkdown": "Yes, my score ~0.14 is a result of using only kaggle train set.", "votes": null }, { "id": "288304", "postDate": "02/27/2018 07:02:36", "content": "I wonder what the best score would really be for those only using the kaggle dataset :)
", "rawMarkdown": "I wonder what the best score would really be for those only using the kaggle dataset :)", "votes": null }, { "id": "415267", "postDate": "11/04/2018 19:10:51", "content": "On what metric do I get the output after submission? Getting 5.14 after an initial submission.
", "rawMarkdown": "On what metric do I get the output after submission? Getting 5.14 after an initial submission.", "votes": null }, { "id": "445361", "postDate": "12/26/2018 09:34:40", "content": "Did you use transfer learning or train the model from scratch ?
", "rawMarkdown": "Did you use transfer learning or train the model from scratch ?", "votes": null }, { "id": "445409", "postDate": "12/26/2018 11:55:24", "content": "transfer for me
", "rawMarkdown": "transfer for me", "votes": null }, { "id": "1187467", "postDate": "02/05/2021 13:16:06", "content": "I am using VGG16 model for transfer learning and then using some fully connected layer at bottom to train my model but my cross validation accuracy is stuck at 20% while my training accuracy is going near 95%. I have tried regularization but it's not working. Any idea what to do next ?
\nmodel on top of VGG 16
\n`def model():
\n model = Sequential()
model.add(Dense(128, activation='relu', input_dim = 8192))\nmodel.add(BatchNormalization())\n\nmodel.add(Dense(64, activation='relu'))\nmodel.add(BatchNormalization())\n\nmodel.add(Dense(120, activation='softmax'))\n\nreturn model\n`
\nloss and accuracy values over epochs
Epoch Train_Loss Train_Acc CV_Loss CV_Acc
\n0 1 4.716832 0.020848 4.454778 0.048900
\n1 2 4.086011 0.079432 3.849136 0.111247
\n2 3 3.436491 0.165269 3.596872 0.133252
\n3 4 2.999841 0.249709 3.495665 0.146699
\n4 5 2.663034 0.310738 3.480850 0.167482
\n… … … … … …
\n95 96 0.019482 0.994642 15.068735 0.169927
\n96 97 0.122124 0.962730 15.967104 0.160147
\n97 98 0.106633 0.966341 15.333890 0.169927
\n98 99 0.057931 0.980550 14.798896 0.165037
\n99 100 0.062000 0.980200 15.929802 0.167482
Thanks
\nIt was really helpful.
I am using VGG16 model for transfer learning and then using some fully connected layer at bottom to train my model but my cross validation accuracy is stuck at 20% while my training accuracy is going near 95%. I have tried regularization but it's not working. Any idea what to do next ?
\nmodel on top of VGG 16
\n`def model():
\n model = Sequential()
model.add(Dense(128, activation='relu', input_dim = 8192))\nmodel.add(BatchNormalization())\n\nmodel.add(Dense(64, activation='relu'))\nmodel.add(BatchNormalization())\n\nmodel.add(Dense(120, activation='softmax'))\n\nreturn model\n`
\nloss and accuracy values over epochs
Epoch Train_Loss Train_Acc CV_Loss CV_Acc
\n0 1 4.716832 0.020848 4.454778 0.048900
\n1 2 4.086011 0.079432 3.849136 0.111247
\n2 3 3.436491 0.165269 3.596872 0.133252
\n3 4 2.999841 0.249709 3.495665 0.146699
\n4 5 2.663034 0.310738 3.480850 0.167482
\n… … … … … …
\n95 96 0.019482 0.994642 15.068735 0.169927
\n96 97 0.122124 0.962730 15.967104 0.160147
\n97 98 0.106633 0.966341 15.333890 0.169927
\n98 99 0.057931 0.980550 14.798896 0.165037
\n99 100 0.062000 0.980200 15.929802 0.167482
Thanks
\nIt was really helpful.
在吗
", "votes": null, "replies": [] }, { "id": 229042, "author_name": "bostjanm", "author_url": "", "post_date": "10/08/2017 15:46:29", "content": "Of course this dataset is from Stanford Dogs Dataset. C&P from description: \"We extend our gratitude to the creators of the Stanford Dogs Dataset for making this competition possible: Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li.\"\nJust wondering, how did you manage to get 0.01? What model did give you best results? Did you use pretrained imagenet weights?
", "votes": null, "replies": [ { "id": 229056, "author_name": "yangpeiwen", "author_url": "", "post_date": "10/08/2017 16:31:27", "content": "I used five pre-trained models to predict features and connect them together, then I trained a simple classifier.
", "votes": null, "replies": [] }, { "id": 229080, "author_name": "gaborfodor", "author_url": "", "post_date": "10/08/2017 17:42:24", "content": "Since both the train and test set came from the Standord Dogs it is easy to train overfitted model on them. It is not really external data...
", "votes": null, "replies": [] }, { "id": 229098, "author_name": "bostjanm", "author_url": "", "post_date": "10/08/2017 18:42:49", "content": "I agree. Clever thinking by combining predictions from all dnn's! I think this is something that I wouldn't even think of, that's too far out of my scope :)
\nThere is one thing that confuses me: advancements are possible only if we find general solution and not exploits on datasets (although exploit exploration is good motivator for Kaggle team and organizers to prepare better datasets in future).
Is it common on Kaggle that datasets are \"poorly\" prepared? Is it common from Kagglers searching for dataset exploits or eval metrics exploits? I don't think there is something wrong about it, I just want to know what makes this community tick. I'm here last 3 weeks and I see that pattern, wondering if this actually makes sense.
\n\nI'm hoping for the future of better NN, advancements from collaboration of different minds is probably best strategy to get there.
", "votes": null, "replies": [] }, { "id": 229119, "author_name": "versalog", "author_url": "", "post_date": "10/08/2017 21:13:48", "content": "I, too, have noticed people looking for exploits and shortcuts. Such explorations have their own value, but in this context they amount to cheating at solitaire.
", "votes": null, "replies": [] }, { "id": 229146, "author_name": "yangpeiwen", "author_url": "", "post_date": "10/09/2017 01:04:39", "content": "Yes.
", "votes": null, "replies": [] }, { "id": 229266, "author_name": "bostjanm", "author_url": "", "post_date": "10/09/2017 08:35:55", "content": "Exploiting is not the same as cheating. In this case cheating would be hand labeling and I will not take a second to comment on cheating. Yang used valid method for proving that test set are from same distribution of imagenet. I think this is just amazing way to prove how DNN arhitectures work when they are combined.
", "votes": null, "replies": [] }, { "id": 229477, "author_name": "versalog", "author_url": "", "post_date": "10/09/2017 17:49:48", "content": "I apologize for being unclear. I did not mean to accuse Yang Peiwen of cheating. I was comparing this educational competition to the game of Solitaire to suggest that using an exploit here is self-defeating.
\n\nSolitaire is an algorithmic sorting game for one player. It starts with a randomly-ordered deck of playing cards. The goal is to sort the deck by applying a limited set of discrete transformations. You win by completely sorting the deck. You lose if you reach a halting state (in which no permitted transformation will add order) prior to completely sorting the deck. The challenge is to minimize the number of moves required to win by choosing the best transformation at each intermediate state.
\n\nThis playground competition is an opportunity to practise specific skills under similarly artificial constraints. Like Solitaire, this is a single-player game. The leaderboard provides motivation and a sense of what kind of improvement is possible; but we're competing against ourselves, so the only way to \"win\" is to get better.
", "votes": null, "replies": [] } ] }, { "id": 229145, "author_name": "zimygogo", "author_url": "", "post_date": "10/09/2017 01:00:06", "content": "你好 我是一名在校的研究生 可以跟你组个队吗
", "votes": null, "replies": [] }, { "id": 230180, "author_name": "wuzuping", "author_url": "", "post_date": "10/11/2017 12:58:49", "content": "huo zhuo pei Shen ,hia hia
", "votes": null, "replies": [] }, { "id": 249068, "author_name": "skinish", "author_url": "", "post_date": "11/27/2017 16:44:25", "content": "Did you use FC layers to grab all the features of those models and output a classification score or another classifier?
", "votes": null, "replies": [] }, { "id": 250904, "author_name": "jjabrams2", "author_url": "", "post_date": "11/30/2017 13:11:07", "content": "Are you running this on a gpu?
", "votes": null, "replies": [] }, { "id": 251323, "author_name": "ptlinyi", "author_url": "", "post_date": "12/01/2017 01:57:43", "content": "Did you change all layers's weights in your fine-tune strategy when you train on this new task?? I tried to fine-tune many pre-trained models ,such as Resnet,Densenet.., but my validation(3000 photos) accuracy just got a little over 81%.
", "votes": null, "replies": [] }, { "id": 251400, "author_name": "sermakarevich", "author_url": "", "post_date": "12/01/2017 05:46:01", "content": "Come on guys. First one with 0 loss gonna shine forever.
", "votes": null, "replies": [ { "id": 253143, "author_name": "yangpeiwen", "author_url": "", "post_date": "12/04/2017 14:09:38", "content": "Impossible, because there are many identical pictures in the training set and test set, but they are different classes, so model can not accurately determine the true labels of these pictures, and can only be submitted by multiple attempts to guess the result.
", "votes": null, "replies": [] }, { "id": 287375, "author_name": "morvayd", "author_url": "", "post_date": "02/24/2018 15:23:27", "content": "it seems, it was possible. :(
", "votes": null, "replies": [] } ] }, { "id": 277152, "author_name": "godsdusk", "author_url": "", "post_date": "02/02/2018 09:30:18", "content": "how about InceptionResNetV2?
", "votes": null, "replies": [] }, { "id": 288276, "author_name": "sainathreddy", "author_url": "", "post_date": "02/27/2018 04:46:09", "content": "is it possible to get log loss less then 0.3 using kaggle dataset without stanford dataset
", "votes": null, "replies": [ { "id": 288279, "author_name": "yangpeiwen", "author_url": "", "post_date": "02/27/2018 04:51:56", "content": "Sure, inceptionv3 + resnet152_v1 can get 0.27143 on Public Leaderboard.
\n\nhttps://github.com/ypwhs/DogBreed_gluon
", "votes": null, "replies": [] }, { "id": 288280, "author_name": "sermakarevich", "author_url": "", "post_date": "02/27/2018 04:52:27", "content": "Yes, my score ~0.14 is a result of using only kaggle train set.
", "votes": null, "replies": [] }, { "id": 288304, "author_name": "jamesrequa", "author_url": "", "post_date": "02/27/2018 07:02:36", "content": "I wonder what the best score would really be for those only using the kaggle dataset :)
", "votes": null, "replies": [] } ] }, { "id": 415267, "author_name": "hasibzunair", "author_url": "", "post_date": "11/04/2018 19:10:51", "content": "On what metric do I get the output after submission? Getting 5.14 after an initial submission.
", "votes": null, "replies": [] }, { "id": 445361, "author_name": "shivamkushwaha", "author_url": "", "post_date": "12/26/2018 09:34:40", "content": "Did you use transfer learning or train the model from scratch ?
", "votes": null, "replies": [ { "id": 445409, "author_name": "pukkinming", "author_url": "", "post_date": "12/26/2018 11:55:24", "content": "transfer for me
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "228870": "And then I got a score of 0.01 on the leaderboard, which proves that Kaggle's dataset is converted from Stanford Dogs Dataset.", "228996": "在吗", "229042": "Of course this dataset is from Stanford Dogs Dataset. C&P from description: \"We extend our gratitude to the creators of the Stanford Dogs Dataset for making this competition possible: Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li.\"\nJust wondering, how did you manage to get 0.01? What model did give you best results? Did you use pretrained imagenet weights?", "229056": "I used five pre-trained models to predict features and connect them together, then I trained a simple classifier.", "229080": "Since both the train and test set came from the Standord Dogs it is easy to train overfitted model on them. It is not really external data...", "229098": "I agree. Clever thinking by combining predictions from all dnn's! I think this is something that I wouldn't even think of, that's too far out of my scope :) \nThere is one thing that confuses me: advancements are possible only if we find general solution and not exploits on datasets (although exploit exploration is good motivator for Kaggle team and organizers to prepare better datasets in future). \n \nIs it common on Kaggle that datasets are \"poorly\" prepared? Is it common from Kagglers searching for dataset exploits or eval metrics exploits? I don't think there is something wrong about it, I just want to know what makes this community tick. I'm here last 3 weeks and I see that pattern, wondering if this actually makes sense. \n \nI'm hoping for the future of better NN, advancements from collaboration of different minds is probably best strategy to get there.", "229119": "I, too, have noticed people looking for exploits and shortcuts. Such explorations have their own value, but in this context they amount to cheating at solitaire.", "229145": "你好 我是一名在校的研究生 可以跟你组个队吗", "229146": "Yes.", "229266": "Exploiting is not the same as cheating. In this case cheating would be hand labeling and I will not take a second to comment on cheating. Yang used valid method for proving that test set are from same distribution of imagenet. I think this is just amazing way to prove how DNN arhitectures work when they are combined.", "229477": "I apologize for being unclear. I did not mean to accuse Yang Peiwen of cheating. I was comparing this educational competition to the game of Solitaire to suggest that using an exploit here is self-defeating.\n\nSolitaire is an algorithmic sorting game for one player. It starts with a randomly-ordered deck of playing cards. The goal is to sort the deck by applying a limited set of discrete transformations. You win by completely sorting the deck. You lose if you reach a halting state (in which no permitted transformation will add order) prior to completely sorting the deck. The challenge is to minimize the number of moves required to win by choosing the best transformation at each intermediate state. \n\nThis playground competition is an opportunity to practise specific skills under similarly artificial constraints. Like Solitaire, this is a single-player game. The leaderboard provides motivation and a sense of what kind of improvement is possible; but we're competing against ourselves, so the only way to \"win\" is to get better.", "230180": "huo zhuo pei Shen ,hia hia", "249068": "Did you use FC layers to grab all the features of those models and output a classification score or another classifier?", "250904": "Are you running this on a gpu?", "251323": "Did you change all layers's weights in your fine-tune strategy when you train on this new task?? I tried to fine-tune many pre-trained models ,such as Resnet,Densenet.., but my validation(3000 photos) accuracy just got a little over 81%.", "251400": "Come on guys. First one with 0 loss gonna shine forever.", "253143": "Impossible, because there are many identical pictures in the training set and test set, but they are different classes, so model can not accurately determine the true labels of these pictures, and can only be submitted by multiple attempts to guess the result.", "277152": "how about InceptionResNetV2?", "287375": "it seems, it was possible. :(", "288276": "is it possible to get log loss less then 0.3 using kaggle dataset without stanford dataset", "288279": "Sure, inceptionv3 + resnet152_v1 can get 0.27143 on Public Leaderboard.\n\nhttps://github.com/ypwhs/DogBreed_gluon", "288280": "Yes, my score ~0.14 is a result of using only kaggle train set.", "288304": "I wonder what the best score would really be for those only using the kaggle dataset :)", "415267": "On what metric do I get the output after submission? Getting 5.14 after an initial submission.", "445361": "Did you use transfer learning or train the model from scratch ?", "445409": "transfer for me", "1187467": "I am using VGG16 model for transfer learning and then using some fully connected layer at bottom to train my model but my cross validation accuracy is stuck at 20% while my training accuracy is going near 95%. I have tried regularization but it's not working. Any idea what to do next ?\n\nmodel on top of VGG 16\n\n`def model():\n model = Sequential()\n \n model.add(Dense(128, activation='relu', input_dim = 8192))\n model.add(BatchNormalization())\n \n model.add(Dense(64, activation='relu'))\n model.add(BatchNormalization())\n \n model.add(Dense(120, activation='softmax'))\n \n return model\n`\nloss and accuracy values over epochs\n\nEpoch\tTrain_Loss\tTrain_Acc CV_Loss\t CV_Acc\n0\t1\t4.716832\t0.020848\t4.454778\t0.048900\n1\t2\t4.086011\t 0.079432\t3.849136\t0.111247\n2\t3\t3.436491\t0.165269\t 3.596872\t0.133252\n3\t4\t2.999841\t0.249709\t3.495665\t0.146699\n4\t5\t2.663034\t0.310738\t3.480850\t0.167482\n...\t...\t...\t ...\t ...\t ...\n95\t96\t0.019482\t0.994642\t15.068735\t0.169927\n96\t97\t0.122124\t 0.962730\t15.967104\t0.160147\n97\t98\t0.106633\t0.966341\t15.333890\t0.169927\n98\t99\t0.057931\t0.980550\t14.798896\t0.165037\n99\t100\t0.062000\t0.980200\t15.929802\t0.167482", "3099790": "Thanks\nIt was really helpful." }, "source": "meta" }