def __init__()

in ss_baselines/savi/models/visual_cnn.py [0:0]

• 70 lines of code
• 12 McCabe index (conditional complexity)

    def __init__(self, observation_space, output_size, extra_rgb=False):
        super().__init__()
        self._output_size = output_size
        if "rgb" in observation_space.spaces and not extra_rgb:
            self._n_input_rgb = observation_space.spaces["rgb"].shape[2]
        else:
            self._n_input_rgb = 0

        if "depth" in observation_space.spaces:
            self._n_input_depth = observation_space.spaces["depth"].shape[2]
        else:
            self._n_input_depth = 0

        if "semantic" in observation_space.spaces:
            self._n_input_semantic = 6
        else:
            self._n_input_semantic = 0

        # kernel size for different CNN layers
        self._cnn_layers_kernel_size = [(8, 8), (4, 4), (3, 3)]

        # strides for different CNN layers
        self._cnn_layers_stride = [(4, 4), (2, 2), (2, 2)]

        if self._n_input_rgb > 0:
            cnn_dims = np.array(
                observation_space.spaces["rgb"].shape[:2], dtype=np.float32
            )
        elif self._n_input_depth > 0:
            cnn_dims = np.array(
                observation_space.spaces["depth"].shape[:2], dtype=np.float32
            )
        elif self._n_input_semantic > 0:
            cnn_dims = np.array(
                observation_space.spaces["semantic"].shape[:2], dtype=np.float32
            )

        if self.is_blind:
            self.cnn = nn.Sequential()
        else:
            self._input_shape = (self._n_input_rgb + self._n_input_depth + self._n_input_semantic,
                                 int(cnn_dims[0]), int(cnn_dims[1]))
            for kernel_size, stride in zip(
                self._cnn_layers_kernel_size, self._cnn_layers_stride
            ):
                cnn_dims = self._conv_output_dim(
                    dimension=cnn_dims,
                    padding=np.array([0, 0], dtype=np.float32),
                    dilation=np.array([1, 1], dtype=np.float32),
                    kernel_size=np.array(kernel_size, dtype=np.float32),
                    stride=np.array(stride, dtype=np.float32),
                )

            self.cnn = nn.Sequential(
                nn.Conv2d(
                    in_channels=self._n_input_rgb + self._n_input_depth + self._n_input_semantic,
                    out_channels=32,
                    kernel_size=self._cnn_layers_kernel_size[0],
                    stride=self._cnn_layers_stride[0],
                ),
                nn.ReLU(True),
                nn.Conv2d(
                    in_channels=32,
                    out_channels=64,
                    kernel_size=self._cnn_layers_kernel_size[1],
                    stride=self._cnn_layers_stride[1],
                ),
                nn.ReLU(True),
                nn.Conv2d(
                    in_channels=64,
                    out_channels=64,
                    kernel_size=self._cnn_layers_kernel_size[2],
                    stride=self._cnn_layers_stride[2],
                ),
                #  nn.ReLU(True),
                Flatten(),
                nn.Linear(64 * cnn_dims[0] * cnn_dims[1], output_size),
                nn.ReLU(True),
            )

        self.layer_init()