Dogs vs. Cats Redux: Kernels Edition is a Kaggle's competition where using the dataset you have to create classifier that can classify images as dog or cat. The evaluation metirc for this competition is cross entropy.
- Getting started
- Model info
- Model predictions
- Visualizing feature maps
- Issues
- Latest model info
- License
The notebook is available on Kaggle to work in the same environment where this notebook was created i.e. use the same version packages used, etc...
Count plot for the labels
There are total 16 kernels (versions) for this project on Kaggle. Here only info for the best, latest and longest trained model info is given, if you want to look at other version then go to the main kernel and change the versions.
The best score for this competition is 3.47 (loss, kernel)
Here VGG16 model is used as the base model and the entire model is trained for 1 epoch. Below is the snapshot of the last epoch.
Model performance against testing data
The latest score for this competition is 4.68 (loss, kernel)
Here InceptionV3 model is used as the base model and the entire model is trained for 1 epochs. Below is the snapshot of the last epoch.
Model performance against testing data
The latest score for this competition is 13.72 (loss, kernel)
Here VGG16 model is used as the base model and the entire model is trained for 30 epochs. Below is the snapshot of the last epoch.
Accuracy and loss curves for this model
Model performance against testing data
Lastest model's predictions on testing data (part of training data)
If you want to see the visualizations for the filters for VGG16 and InceptionV3 then they are available in kernel_vgg16 and kernel_inceptionv3 respectively.
Animal image
1st conv layer
11th conv layer
17th conv layer
Animal image
1st conv layer
3rd conv layer
Mid conv layer
Last conv layer
Only one issue that I faced and that's during the testing(on the test data which is a part of the training data) process and use the other way to get right results.
The issue is explained in the latest kernel's Evaluation section but the summary is: while using the ImageDataGenerator for the testing data with the flow_from_dataframe method where y_col and the target_size was given, the predictions done using the ImageDataGenerator was wrong, this is because the confusion matrix for these predictions were giving reuslts (bad results) as compared to the results giving the evaluate method on tensorflow.keras.Model. So to avoid this issue I just took the images path for the testing data and use the ImageDataGenerator with the flow method on it without the labels info. After doing that the results were in sync with the results of the evaluate method on tensorflow.keras.Model.
Testing data evaluation results
Using the previous method
Using the new (right) method
The confusion matrix for the longest and the best models had this issue, can be viewed in their respective kernels. This issue is solved for the latest kernel.