feedforward_test.py 文件源码

def test_forward_gives_correct_output(self):
        params = Params({
                'input_dim': 2,
                'hidden_dims': 3,
                'activations': 'relu',
                'num_layers': 2
                })
        feedforward = FeedForward.from_params(params)

        constant_init = lambda tensor: torch.nn.init.constant(tensor, 1.)
        initializer = InitializerApplicator([(".*", constant_init)])
        initializer(feedforward)

        input_tensor = Variable(torch.FloatTensor([[-3, 1]]))
        output = feedforward(input_tensor).data.numpy()
        assert output.shape == (1, 3)
        # This output was checked by hand - ReLU makes output after first hidden layer [0, 0, 0],
        # which then gets a bias added in the second layer to be [1, 1, 1].
        assert_almost_equal(output, [[1, 1, 1]])