ml_autoregressive_model

Introducton

An convolutional neural network (CNN) based autoregressive model for image generation on a calligraphy dataset.

Model overview

1. Causal convolution will be performed on training images.
2. The convoluted images enter network blocks which contain Resnet networks and TenserDense layers.
3. Images pass though a tf.nn.elu activation function.
4. Images pass though a CNN layer and a Sigmoid function.

Evaluation

Binary cross entropy loss against test set:
Before loading weight: 0.6972017288208008
After loading weight: 0.202795147895813
After fitting: 0.19266654551029205

Accuracy: 0.9200

Demonstration with Chinese calligrahy inputs