Weights were initialized using Xavier initialization, which ensures that our weights are neither too large nor too small; both cases impede the learning of the network. In Xavier initialization, the weights are assigned a value from a distribution with zero mean and a specific variance 2/(n_in+n_out), where n_in and n_out are respectively the numbers of inputs and output of the layer. To know more about Xavier initialization, you can refer to Glorot and Bengio's 2009 paper; details are given in the See also section.