def _init_weights()

in pytorch-transformers/pytorch_transformers/modeling_transfo_xl.py [0:0]

• 38 lines of code
• 28 McCabe index (conditional complexity)

    def _init_weights(self, m):
        """ Initialize the weights.
        """
        classname = m.__class__.__name__
        if classname.find('Linear') != -1:
            if hasattr(m, 'weight') and m.weight is not None:
                self._init_weight(m.weight)
            if hasattr(m, 'bias') and m.bias is not None:
                self._init_bias(m.bias)
        elif classname.find('AdaptiveEmbedding') != -1:
            if hasattr(m, 'emb_projs'):
                for i in range(len(m.emb_projs)):
                    if m.emb_projs[i] is not None:
                        nn.init.normal_(m.emb_projs[i], 0.0, self.config.proj_init_std)
        elif classname.find('Embedding') != -1:
            if hasattr(m, 'weight'):
                self._init_weight(m.weight)
        elif classname.find('ProjectedAdaptiveLogSoftmax') != -1:
            if hasattr(m, 'cluster_weight') and m.cluster_weight is not None:
                self._init_weight(m.cluster_weight)
            if hasattr(m, 'cluster_bias') and m.cluster_bias is not None:
                self._init_bias(m.cluster_bias)
            if hasattr(m, 'out_projs'):
                for i in range(len(m.out_projs)):
                    if m.out_projs[i] is not None:
                        nn.init.normal_(m.out_projs[i], 0.0, self.config.proj_init_std)
        elif classname.find('LayerNorm') != -1:
            if hasattr(m, 'weight'):
                nn.init.normal_(m.weight, 1.0, self.config.init_std)
            if hasattr(m, 'bias') and m.bias is not None:
                self._init_bias(m.bias)
        else:
            if hasattr(m, 'r_emb'):
                self._init_weight(m.r_emb)
            if hasattr(m, 'r_w_bias'):
                self._init_weight(m.r_w_bias)
            if hasattr(m, 'r_r_bias'):
                self._init_weight(m.r_r_bias)
            if hasattr(m, 'r_bias'):
                self._init_bias(m.r_bias)