CNN to classify the cifar-10 database by using a vgg16 trained on Imagenet as base. The approach is to transfer learn using the first three blocks (top layers) of vgg16 network and adding FC layers on top of them and train it on CIFAR-10.
Trained using two approaches for 250 epochs:
- Keeping the base model's layer fixed, and
- By training end-to-end
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv_bn_relu (ConvBNRelu) (None, 32, 32, 64) 2048
_________________________________________________________________
conv_bn_relu_1 (ConvBNRelu) (None, 32, 32, 64) 37184
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 16, 16, 64) 0
_________________________________________________________________
conv_bn_relu_2 (ConvBNRelu) (None, 16, 16, 128) 74368
_________________________________________________________________
conv_bn_relu_3 (ConvBNRelu) (None, 16, 16, 128) 148096
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 8, 8, 128) 0
_________________________________________________________________
conv_bn_relu_4 (ConvBNRelu) (None, 8, 8, 256) 296192
_________________________________________________________________
conv_bn_relu_5 (ConvBNRelu) (None, 8, 8, 256) 591104
_________________________________________________________________
conv_bn_relu_6 (ConvBNRelu) (None, 8, 8, 256) 591104
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 4, 4, 256) 0
_________________________________________________________________
conv_bn_relu_7 (ConvBNRelu) (None, 4, 4, 512) 1182208
_________________________________________________________________
conv_bn_relu_8 (ConvBNRelu) (None, 4, 4, 512) 2361856
_________________________________________________________________
conv_bn_relu_9 (ConvBNRelu) (None, 4, 4, 512) 2361856
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 2, 2, 512) 0
_________________________________________________________________
conv_bn_relu_10 (ConvBNRelu) (None, 2, 2, 512) 2361856
_________________________________________________________________
conv_bn_relu_11 (ConvBNRelu) (None, 2, 2, 512) 2361856
_________________________________________________________________
conv_bn_relu_12 (ConvBNRelu) (None, 2, 2, 512) 2361856
_________________________________________________________________
max_pooling2d_4 (MaxPooling2 (None, 1, 1, 512) 0
_________________________________________________________________
flatten (Flatten) (None, 512) 0
_________________________________________________________________
dropout_13 (Dropout) (None, 1, 1, 512) 0
_________________________________________________________________
dense (Dense) (None, 512) 262656
_________________________________________________________________
batch_normalization_13 (Batc (None, 512) 2048
_________________________________________________________________
dense_1 (Dense) (None, 10) 5130
_________________________________________________________________
activation (Activation) (None, 10) 0
=================================================================
Total params: 15,001,418
Trainable params: 14,991,946
Non-trainable params: 9,472
_________________________________________________________________
training_epochs = 250
batch_size = 128
learning_rate = 0.1
momentum = 0.9
lr_decay = 1e-6
lr_drop = 20
Source Files:
-
vgg16.py
- load_images() : load cifar-10 images (train, test)
- normalization() : normalization cifar-10 images
- ConvBNRelu : create conv layer with relu, batchnorm
- VGG16Model : create deep learning model based vgg16
- train() : train VGG16Model with cifar-10 images
- main() : main function that Initial images and model then, call train function
-
cifar10vgg_custom.h5 : trained model's weights
| Model | Validation Accuracy |
|---|---|
| VGG-16 | 93.15% |
| ResNet-20 | 91.52% |
| ResNet-32 | 92.53% |
| ResNet-44 | 93.16% |
| ResNet-56 | 93.21% |
| ResNet-110 | 93.90% |
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent