1, when used to make binary sigmoid activation function, number of channels, if one-dimensional output model, but with a loss function binary_crossentropy
2, when used to make binary classification softmax activation function, number of channels, if the number of classes of the model output, while the loss function with categorical_crossentropy
3, where the model function is activated when the default parameters, training and forecasting remember to be changed to correspond, otherwise just like me Baipao countless rounds, the results did not yet know what went wrong, and death
4, each running a model really want to make a record, a long time totally confused which is which, encourage each other