Using https://www.vision.caltech.edu/datasets/cub_200_2011/ dataset
- Dataset info
- Number of categories: 200
- Number of images: 11,788
- Annotations per image: 15 Part Locations, 312 Binary Attributes, 1 Bounding Box
- Modeling
- implement a fully connected neural network (multi-layer perceptron):
- Resize the images to be no larger than 32x32.
- Use the sequential model API in keras to build your network using dense layers
- ResNet-101
- Use the pretrained ResNet-101 network available in Keras.
- Visualization
- t-SNE
- filters learned by each model
- implement a fully connected neural network (multi-layer perceptron):
-
Using 64 * 64 as image size as a step for handle overfitting from the last step
-
step_1
- Using ResNet101 pre-trained model when trainable = False (freeze the layers of the model )
- using flattened layer then adding the output layer with the number of classes and SoftMax as an activation function
-
step_2
- Using ResNet101 pre-trained model when trainable = False but choose 10 last layers to be unfrozeed
- Using a flattened layer then adding the output layer with the number of classes and SoftMax as an activation function with sparse_categorical_crossentropy for the optimizer
