Progress Report
Internship @ Yıldız Teknik Üniversitesi
by Srikote Naewchampa (Bamboo)
18/07/17
TensorFlow
Basic Components
- Tensor
- Computational Graph
- Session
- Placeholder
- Variables
Neural Network
Linear Classifier = Single-Layer NN
Trying to Understand (Intuitively)
- Each input in each neuron
- Each input in all neurons
- A batch of inputs in all neurons
Training
- Cost/loss function
- Step size or learning rate
- Gradient Descent
- Gradient Check
- Back propagation
- Hyperparameters
Activation Functions
- Non-linearity
- Sigmoid
- ReLU (Recommended)
- Others (Maxout, Tanh, ...)
Regularization
- Prevent overfitting
- L1 and L2
- Dropout
"Use as many layers as possible, and use regularization to control overfitting"
Convolutional
Neural Network
Convolutional Layer
- Extract features
- Common hyperparameter values
- Receptive field = 3
- Stride = 1
- Zero padding = 1
Pooling Layer
- Pool outstanding features
- Common hyperparameter values
- Receptive field = 2
- Stride = 2
- Overlapping Pooling , F = 3, S = 2
Fully-Connected Layer
- Looks like linear classifier
- Fully connected
- Convolutional Layer with ...
- Receptive field = Origianl size
Layer Patterns
INPUT->[(CONV->RELU)*N ->POOL]*M->[FC->RELU]*K->FC
Common, N <= 3, K < 3
Suggestions
- Try 3-channel input images
- Focus on accuracy
- And best practice
- After that, performance
Thanks.
deck
By whcwhc78
deck
- 408