Your model is just predicting 0's. In the test set exactly 96.05% of the targets are 0's, this explains why you're getting 96.05% accuracy. The actual metric used in the kaggle leaderboard is Area under ROC not '1-error_rate'. Your Area under ROC score would be 0.5. Try oversampling to balance the test data before feeding it to the CNN
Hey! Great videos first and foremost. I modified the model slightly (has 95.6% train and test acc) but when I make predictions on the actual test data on Kaggle and submit for scoring, the score received is 0.46723. Why is there such a large discrepancy? What is the reason?
@@KGPTalkie after noticing some issues in my predictions, I revised the code to achieve the correct accuracy of 0.65773. The reason for the discrepancy is because the evaluation of the model on Kaggle is the "area under the ROC curve". Since our training data is around 96% satisfied customers. The model is naive and simply predicts all customers are satisfied, thus obtaining ~96% accuracy. I'm working now to try and fix this issue with a combination of 1) under sampling the satisfied customer pool 2) modifying the loss function to more harshly penalize classifying a dissatisfied customer as satisfied. Any other tips?
@LAXMI KANT Hi, Sir, Thank you for making this video. I have been having an issue, as I wanted to test out the model that I have built. I am using a new CSV file, with data (8 inputs) and am trying to pass it through the model.predict() function... BUT i keep getting this error "ValueError: Error when checking input: expected conv1d_input to have 3 dimensions, but got array with shape (8, 1)" ---- I have tried to reshape the data but have been unsuccessful. Any help would be greatly appreciated.
Its really interesting video, but you must use train and test data sets in the library, but you use only the train set and split train set to train and test... anyway, your videos are perfect!
Awesome video! It would be great if you can also download the test from Kaggle and show how to evaluate the model by that dataset and how to prepare it for submission. Great work, thanks again.
When I fit the model, I got this error "AttributeError: 'function' object has no attribute 'shape'" >>>history=model.fit(X_train,y_train,epochs=epochs,validation_data=(X_test,y_test),verbose=1)
Your model is just predicting 0's. In the test set exactly 96.05% of the targets are 0's, this explains why you're getting 96.05% accuracy. The actual metric used in the kaggle leaderboard is Area under ROC not '1-error_rate'. Your Area under ROC score would be 0.5. Try oversampling to balance the test data before feeding it to the CNN
Iam a beginner,
Cnn is used for image classification...
Here we doing on numerical data.
When to use conv1D and conv2D
1d for datasets and 2d for images
please make vides image dateset preprocessing that are not publically available. Thanks
Yeah sure. I will try to do it.
This model is not working for kaggle. There is need to do advance feature engineering. Getting only 50% accuracy
Hey! Great videos first and foremost. I modified the model slightly (has 95.6% train and test acc) but when I make predictions on the actual test data on Kaggle and submit for scoring, the score received is 0.46723. Why is there such a large discrepancy? What is the reason?
Try to use balance and unbalance both. And see how it is performing.
@@KGPTalkie after noticing some issues in my predictions, I revised the code to achieve the correct accuracy of 0.65773.
The reason for the discrepancy is because the evaluation of the model on Kaggle is the "area under the ROC curve". Since our training data is around 96% satisfied customers. The model is naive and simply predicts all customers are satisfied, thus obtaining ~96% accuracy.
I'm working now to try and fix this issue with a combination of 1) under sampling the satisfied customer pool 2) modifying the loss function to more harshly penalize classifying a dissatisfied customer as satisfied.
Any other tips?
@@redhotchilipeppers38 Please let us know if are getting better results. Also if possible, please share us the code.
@LAXMI KANT Hi, Sir, Thank you for making this video. I have been having an issue, as I wanted to test out the model that I have built. I am using a new CSV file, with data (8 inputs) and am trying to pass it through the model.predict() function... BUT i keep getting this error "ValueError: Error when checking input: expected conv1d_input to have 3 dimensions, but got array with shape (8, 1)" ---- I have tried to reshape the data but have been unsuccessful. Any help would be greatly appreciated.
You need to reshape the data.. Follow exact procedure for testing data as you did for training data.
Its really interesting video, but you must use train and test data sets in the library, but you use only the train set and split train set to train and test...
anyway, your videos are perfect!
Thank you for the watching ❤️. We do not have label for test set so can't test the accuracy therefore I had splitted train set in train and test set.
Awesome video! It would be great if you can also download the test from Kaggle and show how to evaluate the model by that dataset and how to prepare it for submission.
Great work, thanks again.
Thank you so much for watching ❤️ 😍
please sir make the videos on medical image analysis
Thank you for your feedback. I was also thinking the same. Please stay tuned for more videos. Thanks for watching.
When I fit the model, I got this error "AttributeError: 'function' object has no attribute 'shape'" >>>history=model.fit(X_train,y_train,epochs=epochs,validation_data=(X_test,y_test),verbose=1)
Please compare with original file. Link is given in the video description.
@@KGPTalkie Thank you for your reply. I have already solve this issue. Your lecture is really amazing.
Thank you!