An implementation of weight/bias initializers and batch normalization in Tensorflow.
MNIST database is used to show performance-comparison
In order to examine the effect of initializers and batch normalization, a simple network architecture called multilayer perceptrons (MLP) is employed.
MLP has following architecture.
- input layer : 784 nodes (MNIST images size)
- first hidden layer : 256 nodes
- second hidden layer : 256 nodes
- output layer : 10 nodes (number of class for MNIST)
The following initializers for weights/biases of network are considered.
- Weight Initializer
- normal, trucated normal
- xaiver : Understanding the difficulty of training deep feedforward neural networks
- he : Delving Deep into Rectifiers
- Bias Initializer
- normal, constant
- Weight Initializer : he > trauncated normal = xaiver > normal
- Bias Initilaizer : zero > normal
Sample results are following.
| Index | Weight Initializer | Bias Initializer | Accuracy |
|---|---|---|---|
| normal_w_normal_b_0.9451 | normal | normal | 0.9451 |
| normal_w_zero_b_0.9485 | normal | zero | 0.9485 |
| truncated_normal_w_normal_b_0.9788 | truncated_normal | normal | 0.9788 |
| truncated_normal_w_zero_b_0.9790 | truncated_normal | zero | 0.9790 |
| xavier_w_normal_b_0.9800 | xavier | normal | 0.9800 |
| xavier_w_zero_b_0.9806 | xavier | zero | 0.9806 |
| he_w_normal_b_0.9798 | he | normal | 0.9798 |
| he_w_zero_b_0.9811 | he | zero | 0.9811 |
Batch normalization improves performance of network in terms of final accuracy and convergence rate.
In this simulation, bias initalizer was zero-constant initializer and weight initializers were xavier / he.
Sample results are following.
| Index | Weight Initializer | Batch Normalization | Accuracy |
|---|---|---|---|
| xavier_woBN_0.9806 | xavier | Unused | 0.9806 |
| xavier_withBN_0.9812 | xavier | Used | 0.9812 |
| he_woBN_0.9811 | he | Unused | 0.9811 |
| he_withBN_0.9837 | he | Used | 0.9837 |
python run_main.py --weight-init <weight initializer> --bias-init <bias initializer> --batch-norm <True or False>
<weight initializer> must be selected in [normal, truncated_normal, xavier, he].
<bias initializer> must be selected in [normal, zero].
You may command like python run_main.py --weight-init xavier --bias-init zero --batch-norm True.
This implementation has been tested on Tensorflow r0.12.