replace inplace copy with nn.Parameter

Author: tgisaturday

taming/modules/vqvae/quantize.py

@@ -342,8 +342,8 @@ def __init__(self, n_embed, embedding_dim, beta, decay=0.99, eps=1e-5,
         self.embedding = nn.Embedding(self.n_embed, self.embedding_dim)
         self.embedding.weight.requires_grad = False
         self.cluster_size = nn.Parameter(torch.zeros(n_embed),requires_grad=False)
-        self.embed_avg = nn.Parameter(torch.Tensor(self.n_embed, self.embedding_dim),requires_grad=False)
-        self.embed_avg.data.copy_(self.embedding.weight.data)
+        self.embed_avg = nn.Parameter(torch.randn(self.n_embed, self.embedding_dim),requires_grad=False)
+
         self.remap = remap
         if self.remap is not None:
             self.register_buffer("used", torch.tensor(np.load(self.remap)))