Training logs and accuracy reports are attached in the tail of the files.

• 01_cifar10_conv.py: Train and save a conv network with 4 conv layers, which will be used as a feature extractor. The final accuracy is about 88.70% for 200 epochs.

• 01_cifar10_conv_base.py: Define a new model with the same conv layers as 01_cifar10_conv.py, but with new FC layers. The conv layers are fixed and not updated when training, for 50 epochs, the result is about 87.44%.

• 02_cifar10_vgg11.py: The same as 01_cifar10_conv.py, but using VGG11's conv layers, which have 8 conv layers. When trained 300 epochs, the accuracy is about 90.79%。

• 02_cifar10_vgg11_base.py: The same as 01_cifar10_conv_base.py, but using the above 8-layer conv networks as a feature extractor. The accuracy is 90.73%.

• 03_cifar10_vgg16.py: The same as 02_cifar10_vgg11.py, but using VGG-16, which has 13 convolution layers. The accuracy is 92.41%.

• 03_cifar10_vgg16_base.py: The same as 02_cifar10_vgg11_base.py, but using VGG-16's conv layers as a feature extractor. The accuracy is also 92.41%%.

• Fixed the Conv Layers as feature extractor.
• Finetune the Conv Layers (VGG16 should be about 92%).
• Add self-contained connections.
• Test some update rules of self-contained layer.

• [1] Train CIFAR10 with PyTorch