def forward(self, x):
    layer = nn.Linear(10, 64)          # new layer each call
    return self.head(torch.relu(layer(x)))