Image classification Case 2
Different from the case 1, an image more wider, and different sizes .
Segmentation validation set, more images augmented operation, taken when defining model "fine tuning" of thinking.
BOTH
Generative Adversarial Networks (generated against neural network)
-
Generative learning:
Wherein no given data set, the data set to learn, and then generate some data (they do not belong to this data set, however, and is very similar to the data set)
-
definition:
A determination model to improve the model generating capacity, are unable to classify the data in the end is the data set or generated by the generator.
GAN model is to let the generator and classifier confrontation .
-
Loss function:
-
Classifier
classifier is a standard binary classification , which is a loss function cross entropy , approaching 1 is the real data set is generated from nearly 0 generator. -
Builder
from a Gaussian random sampling, the training function to maximize cross-entropy, because they want to generate more in line with real distribution.
-
-
skill:
The loss function generator into min (log), to avoid gradient disappears conditions.
DCGAN
Deep Convolutional Generative Adversarial Networks
application convolution deep network of GAN
-
data set
To use totensor conversion of data normalization, changes from [0,1] [1,1], so as to allow the format to match the output data and the input data.
-
Builder
There transpose convolutional layer, can grow in size, different convolution kernel, step, etc., may be set the output format to the size you want.
There are four basic blocks. -
Discriminator
It is a standard convolution neural network to do with the Leaky Rule activation function, compared with the Rule, not simply negative to 0, but turned him into αx.
-
Training process
GAN with substantially similar
