I don't know if its because my model isn't complex enough, or I am underfitting (I train on 150 epochs!), but my accuracy seems too low for the model (0.14). I use ImageDataGenerator for augmentation, too. I would be grateful if someone could help me out!
\n\nMy Kernel: https://www.kaggle.com/arjunrao2000/baseline-keras-model
", "messages": [ { "id": "566215", "postDate": "07/01/2019 22:58:44", "content": "I don't know if its because my model isn't complex enough, or I am underfitting (I train on 150 epochs!), but my accuracy seems too low for the model (0.14). I use ImageDataGenerator for augmentation, too. I would be grateful if someone could help me out!
\n\nMy Kernel: https://www.kaggle.com/arjunrao2000/baseline-keras-model
", "rawMarkdown": "I don't know if its because my model isn't complex enough, or I am underfitting (I train on 150 epochs!), but my accuracy seems too low for the model (0.14). I use ImageDataGenerator for augmentation, too. I would be grateful if someone could help me out!\n\nMy Kernel: [https://www.kaggle.com/arjunrao2000/baseline-keras-model](https://www.kaggle.com/arjunrao2000/baseline-keras-model)", "votes": null }, { "id": "566331", "postDate": "07/02/2019 04:15:21", "content": "Your model is not learning anything useful. You have two choices --do more to prevent over-fitting (handling drop-out properly, for example). You could also try to use transfer learning. That is, take a pre-trained model and use it as the first layers of a neural net classifier. Also, make sure that your data augmentations make sense.
", "rawMarkdown": "Your model is not learning anything useful. You have two choices --do more to prevent over-fitting (handling drop-out properly, for example). You could also try to use transfer learning. That is, take a pre-trained model and use it as the first layers of a neural net classifier. Also, make sure that your data augmentations make sense.", "votes": null }, { "id": "566449", "postDate": "07/02/2019 07:15:48", "content": "After glancing your code , I found something unusual : \npython\ntrain_generator = train_datagen.flow_from_dataframe(\n dataframe = train_df,\n directory = train_dir,\n validation_split = 0.2,\n x_col = 'id_code',\n y_col = 'diagnosis',\n target_size = (32,32),\n class_mode = 'categorical',\n batch_size = 32,\n subset = 'training'\n)\n\nSo here you are using categorical target variable implies target is one hot encoded. But when you are creating the model \npython\nmodel.add(Flatten())\nmodel.add(Dense(512,activation = 'relu'))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(5,activation = 'sigmoid'))\n\nI think you should use softmax not sigmoid. Other suggestions are ,
(32,32) which is very less (because we are given images which are shapes in like 2000's ), 0.14 public score and 0.14 is not accuracy , its cohens kappa. Try to incorporate kappa score on validation in your code. I am not that good with keras , so may be some opinions are wrong.
", "rawMarkdown": "After glancing your code , I found something unusual : \n```python\ntrain_generator = train_datagen.flow_from_dataframe(\n dataframe = train_df,\n directory = train_dir,\n validation_split = 0.2,\n x_col = 'id_code',\n y_col = 'diagnosis',\n target_size = (32,32),\n class_mode = 'categorical',\n batch_size = 32,\n subset = 'training'\n)\n```\nSo here you are using categorical target variable implies target is one hot encoded. But when you are creating the model \n```python\nmodel.add(Flatten())\nmodel.add(Dense(512,activation = 'relu'))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(5,activation = 'sigmoid'))\n```\nI think you should use `softmax` not `sigmoid`. Other suggestions are , \n\n- you are resizing the images to `(32,32)` which is very less (because we are given images which are shapes in like 2000's ), \n- as bryan suggested use pretrained models which may boost your accuracy. \n - Your kernel got `0.14` public score and `0.14` is not accuracy , its cohens kappa. Try to incorporate kappa score on validation in your code. \n\n\nI am not that good with keras , so may be some opinions are wrong.", "votes": null }, { "id": "568789", "postDate": "07/05/2019 12:48:40", "content": "Thank you so much! Really helpful advice!
", "rawMarkdown": "Thank you so much! Really helpful advice!", "votes": null }, { "id": "568790", "postDate": "07/05/2019 12:49:10", "content": "This is great! I'll definitely try a pre-trained model. Any suggestions for this competition?
", "rawMarkdown": "This is great! I'll definitely try a pre-trained model. Any suggestions for this competition?", "votes": null } ], "comments": [ { "id": 566331, "author_name": "puremath86", "author_url": "", "post_date": "07/02/2019 04:15:21", "content": "Your model is not learning anything useful. You have two choices --do more to prevent over-fitting (handling drop-out properly, for example). You could also try to use transfer learning. That is, take a pre-trained model and use it as the first layers of a neural net classifier. Also, make sure that your data augmentations make sense.
", "votes": null, "replies": [ { "id": 568790, "author_name": "arjunrao2000", "author_url": "", "post_date": "07/05/2019 12:49:10", "content": "This is great! I'll definitely try a pre-trained model. Any suggestions for this competition?
", "votes": null, "replies": [] } ] }, { "id": 566449, "author_name": "suchith0312", "author_url": "", "post_date": "07/02/2019 07:15:48", "content": "After glancing your code , I found something unusual : \npython\ntrain_generator = train_datagen.flow_from_dataframe(\n dataframe = train_df,\n directory = train_dir,\n validation_split = 0.2,\n x_col = 'id_code',\n y_col = 'diagnosis',\n target_size = (32,32),\n class_mode = 'categorical',\n batch_size = 32,\n subset = 'training'\n)\n\nSo here you are using categorical target variable implies target is one hot encoded. But when you are creating the model \npython\nmodel.add(Flatten())\nmodel.add(Dense(512,activation = 'relu'))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(5,activation = 'sigmoid'))\n\nI think you should use softmax not sigmoid. Other suggestions are ,
(32,32) which is very less (because we are given images which are shapes in like 2000's ), 0.14 public score and 0.14 is not accuracy , its cohens kappa. Try to incorporate kappa score on validation in your code. I am not that good with keras , so may be some opinions are wrong.
", "votes": null, "replies": [ { "id": 568789, "author_name": "arjunrao2000", "author_url": "", "post_date": "07/05/2019 12:48:40", "content": "Thank you so much! Really helpful advice!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "566215": "I don't know if its because my model isn't complex enough, or I am underfitting (I train on 150 epochs!), but my accuracy seems too low for the model (0.14). I use ImageDataGenerator for augmentation, too. I would be grateful if someone could help me out!\n\nMy Kernel: [https://www.kaggle.com/arjunrao2000/baseline-keras-model](https://www.kaggle.com/arjunrao2000/baseline-keras-model)", "566331": "Your model is not learning anything useful. You have two choices --do more to prevent over-fitting (handling drop-out properly, for example). You could also try to use transfer learning. That is, take a pre-trained model and use it as the first layers of a neural net classifier. Also, make sure that your data augmentations make sense.", "566449": "After glancing your code , I found something unusual : \n```python\ntrain_generator = train_datagen.flow_from_dataframe(\n dataframe = train_df,\n directory = train_dir,\n validation_split = 0.2,\n x_col = 'id_code',\n y_col = 'diagnosis',\n target_size = (32,32),\n class_mode = 'categorical',\n batch_size = 32,\n subset = 'training'\n)\n```\nSo here you are using categorical target variable implies target is one hot encoded. But when you are creating the model \n```python\nmodel.add(Flatten())\nmodel.add(Dense(512,activation = 'relu'))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(5,activation = 'sigmoid'))\n```\nI think you should use `softmax` not `sigmoid`. Other suggestions are , \n\n- you are resizing the images to `(32,32)` which is very less (because we are given images which are shapes in like 2000's ), \n- as bryan suggested use pretrained models which may boost your accuracy. \n - Your kernel got `0.14` public score and `0.14` is not accuracy , its cohens kappa. Try to incorporate kappa score on validation in your code. \n\n\nI am not that good with keras , so may be some opinions are wrong.", "568789": "Thank you so much! Really helpful advice!", "568790": "This is great! I'll definitely try a pre-trained model. Any suggestions for this competition?" }, "source": "meta" }