Computer VisionMachine Learning
A compact Keras CNN, augmentation, and a learning-rate ablation.
Notebook with recorded results.
TensorFlowKerasscikit-learn
Overview
Three convolution blocks (32, 64, 128 filters) with max pooling, a dense layer, and dropout, plus experiments on augmentation and learning rate.
Recorded results
- Test accuracy
- 0.741
- Baseline, 10 epochs
- Source: CNN_for_CIFAR10.ipynb, cell 4
- With augmentation
- 0.742
- Rotation, shift, flip, zoom
- Source: CNN_for_CIFAR10.ipynb, cell 10
- Learning rate 0.01 / 0.1
- 0.425 / 0.101
- Same model; higher rates diverge
- Source: CNN_for_CIFAR10.ipynb, cell 15