Exemplo n.º 1
0
    def test_plot_topn_connections_forward(self):
        """ Forward top 2 connections """

        weights = np.array([[[[-1, 0, 1], [+0, 0, 1]], [[1, 1, 2], [2, 3, 3]],
                             [[7, 1, 2], [2, 3, 4]]],
                            [[[+1, 1, 2], [-1, 1, 2]], [[1, 0, 6], [4, 1, 8]],
                             [[6, 1, 2], [2, 0, 2]]],
                            [[[-3, 0, 5], [+0, 1, 2]], [[0, 2, 4], [5, 1, 4]],
                             [[9, 1, 2], [4, 6, 6]]]])

        layer = Convo2D('test_2', '', 3, weights, weights)
        layer.set_coordinates(10, 0)
        prev_layer = Dense('test_prev', '', 2, np.ones((3, 2)), np.ones(
            (3, 2)))  # weights/grads of other do no matter
        prev_layer.set_coordinates(0, 0)

        strongest_idx, shapes = layer.plot_topn_connections_forward(
            prev_layer, 2, [1])

        assert strongest_idx.shape == (2, )
        assert (strongest_idx == np.array([1, 2])).all()
        assert len(
            shapes) == 2  # For each active Dense to each Convolution top 2
        assert isinstance(shapes[0], dict)
        assert shapes[0]['type'] == 'path'
Exemplo n.º 2
0
    def test_plot_topn_connections_backward(self):
        """ Backward top 2 connections """
        # def plot_topn_connections(self, backward_layer, topn, active_units, True)

        # 3 x 3 shape filters x 2 with 3 inputs => weights array of shape (3, 3, 2, 3)
        weights = np.array([[[[-1, 0, 1], [+0, 0, 1]], [[1, 1, 2], [2, 3, 3]],
                             [[7, 1, 2], [2, 3, 4]]],
                            [[[+1, 1, 2], [-1, 1, 2]], [[1, 0, 6], [4, 1, 8]],
                             [[6, 1, 2], [2, 0, 2]]],
                            [[[-3, 0, 5], [+0, 1, 2]], [[0, 2, 4], [5, 1, 4]],
                             [[9, 1, 2], [4, 6, 6]]]])
        # Max on convolution filters
        # [[9, 2, 6], [5, 6, 8]]
        layer = Convo2D('test_1', 3, weights, weights)
        layer.set_xoffset(10)

        prev_layer = Dense('test_prev', 2, np.ones((3, 2)), np.ones(
            (3, 2)))  # weights/grads of other do no matter
        prev_layer.set_xoffset(0)

        strongest_idx, shapes = layer.plot_topn_connections(
            prev_layer, 2, [1, 2], True)

        assert strongest_idx.shape == (2, )
        assert (strongest_idx == np.array([0, 1])).all()
        assert len(shapes) == 4  # Each Convolution to the top Dense
        assert isinstance(shapes[0], dict)
        assert shapes[0]['type'] == 'path'
Exemplo n.º 3
0
    def test_plot_topn_connections_backward(self):
        """ Backward top 2 connections """

        # 3 x 3 shape filters x 2 with 3 inputs => weights array of shape (3, 3, 2, 3)
        weights = self.weights_convo_3_3_4_3
        layer = Convo2D('test_1', '', 3, weights, weights)
        layer.set_coordinates(10, 0)

        prev_layer = Dense('test_prev', '', 4, np.ones((6, 4)), np.ones(
            (6, 4)))  # weights/grads of other do no matter
        prev_layer.set_coordinates(0, 0)

        strongest_idx, shapes = layer.plot_topn_connections_backward(
            prev_layer, 2, [1, 2])

        assert strongest_idx.shape == (2, )
        assert (strongest_idx == np.array([1, 3])).all()
        assert len(shapes) == 4  # Each Convolution to the top Dense
        assert isinstance(shapes[0], dict)
        assert shapes[0]['type'] == 'path'
Exemplo n.º 4
0
    def test_plot_topn_connections_backward_flatten_stride2d(self):
        """ Backward top 2 connections with stride on 2D and flatten """

        # 3 x 3 shape filters x 4 with 3 inputs => weights array of shape (3, 3, 4, 3)
        weights = self.weights_convo_3_3_4_3
        layer = Convo2D('test_1', '', 3, weights, weights,
                        flatten_output=True)  # <--
        layer.set_coordinates(10, 0)

        # Stride or pooling induce a fractional sampling factor (down-sampling)
        layer.append_sampling_factor(1 / np.array([2, 2, 1, 1]))  # <--

        prev_layer = Dense('test_prev', '', 4, np.ones((6, 4)), np.ones(
            (6, 4)))  # weights/grads of other do no matter
        prev_layer.set_coordinates(0, 0)

        strongest_idx, shapes = layer.plot_topn_connections_backward(
            prev_layer, 2, [1, 2])

        assert strongest_idx.shape == (2, )
        assert (strongest_idx == np.array([1, 3])).all()  # <--
        assert len(shapes) == 4  # Each Convolution to the top Dense
        assert isinstance(shapes[0], dict)
        assert shapes[0]['type'] == 'path'