- Pretrained model is used
- freezing all layer except x number of layer (experimental)
- ImageDataGenerator is used for preprocessing images
- In output Softmax activation is used as it is multiclass
- Total num epochs : 50
- Adam optimizer is used with 0.0001 learning rate
- Dropout & early_stop function used to avoid overfitting.
- Changes
- Rotation Range changed from 20 to 10
- Width , Height shift & zoom range changed from 0.2 to 0.1
- Freezing all layer
- Model saved based on max validation accuracy
- Roboflow dataset is used (Previous ModelV01 was on Kaggle Dataset)
- Total seven model is trained and measured
- Removed
- Shear range Removed
- Dropout, New layer (experimented previously)
- Early stop function removed
- Added
- Preprocessing function added
- Shuffle added
- Advantage
- Accuracy increased
- Complexity reduced
- Less number of params
- Accurate ROC,AUC curve
- Accurate performance matrices
- Changes
- Kaggle and Roboflow both dataset is Measured separately
- Roboflow dataset decreases after filtering
- Removed
- Few number of images removed after filtering process
- Added
- Image filtering added
- labels added for image preview and other task
- Total training time measured
- Epoch number measured by highest training and validation accuracy
- Advantage
- Black images detected
- Two dataset models are compared
- Accurate result analysis
- Changes
- Kaggle and Roboflow dataset are combined
- Removed
- Nothing removed
- Added
- More comparison added
- Advantage
- Accuracy improved