cnn

Bases: rpn

A Convolutional Neural Network (CNN) built on the RPN framework.

This CNN architecture consists of:

• Grid-based interdependence layers for convolution.
• Optional pooling layers for feature map compression.
• Fully connected perceptron layers for classification or regression.

The architecture supports various configurations, including patch-based structures, pooling strategies, and advanced processing features like batch normalization, dropout, and activation functions.

Attributes:

Name	Type	Description
h	int	Height of the input grid.
w	int	Width of the input grid.
channel_nums	list[int] | tuple[int]	Sequence of channel numbers for each convolutional layer.
fc_dims	list[int] | tuple[int]	Dimensions of the fully connected layers.
d	int	Depth of the input grid. Default is 1 (2D input).
fc_channel_num	int	Number of channels for the fully connected layers. Default is 1.
width	int	Number of parallel heads for each grid interdependence layer. Default is 1.
pooling_metric	str	Pooling metric used in pooling layers (e.g., 'batch_max'). Default is 'batch_max'.
pooling_layer_gaps	int	Number of layers before adding a pooling layer. Default is 2.
patch_size_half_after_pooling	bool	If True, reduces patch size by half after pooling. Default is False.
patch_shape	str	Shape of the patches ('cuboid', 'cylinder', or 'sphere'). Default is 'cuboid'.
with_batch_norm	bool	If True, applies batch normalization. Default is True.
with_relu	bool	If True, applies ReLU activation. Default is True.
with_softmax	bool	If True, applies softmax activation to the final output. Default is False.
with_residual	bool	If True, includes residual connections. Default is False.
with_dropout	bool	If True, applies dropout. Default is True.
device	str	Device for computation ('cpu' or 'cuda'). Default is 'cpu'.

Methods:

Name	Description
__init__	Initializes the CNN with grid-based convolutional layers, pooling layers, and fully connected layers.

Source code in tinybig/model/rpn_cnn.py

class cnn(rpn):
    """
    A Convolutional Neural Network (CNN) built on the RPN framework.

    This CNN architecture consists of:

    - Grid-based interdependence layers for convolution.
    - Optional pooling layers for feature map compression.
    - Fully connected perceptron layers for classification or regression.

    The architecture supports various configurations, including patch-based structures,
    pooling strategies, and advanced processing features like batch normalization,
    dropout, and activation functions.

    Attributes
    ----------
    h : int
        Height of the input grid.
    w : int
        Width of the input grid.
    channel_nums : list[int] | tuple[int]
        Sequence of channel numbers for each convolutional layer.
    fc_dims : list[int] | tuple[int]
        Dimensions of the fully connected layers.
    d : int
        Depth of the input grid. Default is 1 (2D input).
    fc_channel_num : int
        Number of channels for the fully connected layers. Default is 1.
    width : int
        Number of parallel heads for each grid interdependence layer. Default is 1.
    pooling_metric : str
        Pooling metric used in pooling layers (e.g., 'batch_max'). Default is 'batch_max'.
    pooling_layer_gaps : int
        Number of layers before adding a pooling layer. Default is 2.
    patch_size_half_after_pooling : bool
        If True, reduces patch size by half after pooling. Default is False.
    patch_shape : str
        Shape of the patches ('cuboid', 'cylinder', or 'sphere'). Default is 'cuboid'.
    with_batch_norm : bool
        If True, applies batch normalization. Default is True.
    with_relu : bool
        If True, applies ReLU activation. Default is True.
    with_softmax : bool
        If True, applies softmax activation to the final output. Default is False.
    with_residual : bool
        If True, includes residual connections. Default is False.
    with_dropout : bool
        If True, applies dropout. Default is True.
    device : str
        Device for computation ('cpu' or 'cuda'). Default is 'cpu'.

    Methods
    -------
    __init__(...)
        Initializes the CNN with grid-based convolutional layers, pooling layers, and fully connected layers.
    """
    def __init__(
        self,
        h: int, w: int,
        channel_nums: list[int] | tuple[int],
        fc_dims: list[int] | tuple[int],
        d: int = 1,
        fc_channel_num: int = 1,
        width: int = 1,
        pooling_metric: str = 'batch_max',
        pooling_layer_gaps: int = 2,
        patch_size_half_after_pooling: bool = False,
        name: str = 'rpn_cnn',
        # patch structure parameters for interdependence
        patch_shape: str = 'cuboid',
        p_h: int = None, p_h_prime: int = None,
        p_w: int = None, p_w_prime: int = None,
        p_d: int = 0, p_d_prime: int = None,
        p_r: int = None,
        cd_h: int = None, cd_w: int = None, cd_d: int = 1,
        packing_strategy: str = 'densest_packing',
        # patch structure parameters for compression
        pooling_patch_shape: str = None,
        pooling_p_h: int = None, pooling_p_h_prime: int = None,
        pooling_p_w: int = None, pooling_p_w_prime: int = None,
        pooling_p_d: int = None, pooling_p_d_prime: int = None,
        pooling_p_r: int = None,
        pooling_cd_h: int = None, pooling_cd_w: int = None, pooling_cd_d: int = None,
        pooling_packing_strategy: str = None,
        # output processing function parameters
        with_batch_norm: bool = True,
        with_relu: bool = True,
        with_softmax: bool = False,
        with_residual: bool = False,
        with_dropout: bool = True, p_pooling: float = 0.25, p_fc: float = 0.5,
        # parameter reconciliation function parameters
        with_dual_lphm: bool = False,
        with_lorr: bool = False, r: int = 3,
        enable_bias: bool = True,
        # parameter reconciliation function parameters
        with_perceptron_residual: bool = None,
        with_perceptron_dual_lphm: bool = None,
        with_perceptron_lorr: bool = None, perceptron_r: int = None,
        enable_perceptron_bias: bool = None,
        # other parameters
        device: str = 'cpu', *args, **kwargs
    ):
        """
        Initializes the CNN (Convolutional Neural Network) model.

        This method constructs the CNN architecture with configurable convolutional layers, optional pooling layers,
        and fully connected perceptron layers. Various processing features like batch normalization, dropout,
        and residual connections can also be enabled.

        Parameters
        ----------
        h : int
            Height of the input grid.
        w : int
            Width of the input grid.
        channel_nums : list[int] | tuple[int]
            Sequence of channel numbers for each convolutional layer.
        fc_dims : list[int] | tuple[int]
            Dimensions of the fully connected layers.
        d : int, optional
            Depth of the input grid. Default is 1 (2D input).
        fc_channel_num : int, optional
            Number of channels for the fully connected layers. Default is 1.
        width : int, optional
            Number of parallel heads for each grid interdependence layer. Default is 1.
        pooling_metric : str, optional
            Pooling metric used in pooling layers (e.g., 'batch_max'). Default is 'batch_max'.
        pooling_layer_gaps : int, optional
            Number of layers before adding a pooling layer. Default is 2.
        patch_size_half_after_pooling : bool, optional
            If True, reduces patch size by half after pooling. Default is False.
        name : str, optional
            Name of the CNN model. Default is 'rpn_cnn'.
        patch_shape : str, optional
            Shape of the patches ('cuboid', 'cylinder', or 'sphere'). Default is 'cuboid'.
        p_h, p_h_prime : int, optional
            Height and height prime of the patches. Default is None.
        p_w, p_w_prime : int, optional
            Width and width prime of the patches. Default is None.
        p_d, p_d_prime : int, optional
            Depth and depth prime of the patches. Default is 0 and None, respectively.
        p_r : int, optional
            Radius of spherical or cylindrical patches. Default is None.
        cd_h, cd_w, cd_d : int, optional
            Compression dimensions for height, width, and depth, respectively. Default is 1.
        packing_strategy : str, optional
            Strategy for patch packing. Default is 'densest_packing'.
        pooling_patch_shape : str, optional
            Shape of pooling patches. Default is None (same as `patch_shape`).
        pooling_p_h, pooling_p_h_prime : int, optional
            Height and height prime of pooling patches. Default is None.
        pooling_p_w, pooling_p_w_prime : int, optional
            Width and width prime of pooling patches. Default is None.
        pooling_p_d, pooling_p_d_prime : int, optional
            Depth and depth prime of pooling patches. Default is None.
        pooling_p_r : int, optional
            Radius of pooling patches. Default is None.
        pooling_cd_h, pooling_cd_w, pooling_cd_d : int, optional
            Compression dimensions for pooling patches. Default is None.
        pooling_packing_strategy : str, optional
            Packing strategy for pooling patches. Default is None (same as `packing_strategy`).
        with_batch_norm : bool, optional
            If True, applies batch normalization. Default is True.
        with_relu : bool, optional
            If True, applies ReLU activation. Default is True.
        with_softmax : bool, optional
            If True, applies softmax activation to the final output. Default is False.
        with_residual : bool, optional
            If True, includes residual connections. Default is False.
        with_dropout : bool, optional
            If True, applies dropout. Default is True.
        p_pooling : float, optional
            Dropout probability for pooling layers. Default is 0.25.
        p_fc : float, optional
            Dropout probability for fully connected layers. Default is 0.5.
        with_dual_lphm : bool, optional
            If True, enables dual low-parametric high-order interdependence. Default is False.
        with_lorr : bool, optional
            If True, enables low-rank parameterized interdependence. Default is False.
        r : int, optional
            Rank parameter for low-rank interdependence. Default is 3.
        enable_bias : bool, optional
            If True, includes bias in the layers. Default is True.
        with_perceptron_residual : bool, optional
            If True, includes residual connections in perceptron layers. Default is None (inherits from `with_residual`).
        with_perceptron_dual_lphm : bool, optional
            If True, enables dual low-parametric high-order interdependence in perceptron layers. Default is None.
        with_perceptron_lorr : bool, optional
            If True, enables low-rank parameterized interdependence in perceptron layers. Default is None.
        perceptron_r : int, optional
            Rank parameter for perceptron layers. Default is None.
        enable_perceptron_bias : bool, optional
            If True, includes bias in perceptron layers. Default is None.
        device : str, optional
            Device for computation ('cpu' or 'cuda'). Default is 'cpu'.
        *args : optional
            Additional positional arguments for layers.
        **kwargs : optional
            Additional keyword arguments for layers.

        Raises
        ------
        ValueError
            If input dimensions are invalid or improperly configured.
        """
        print('############# rpn-cnn model architecture ############')

        if pooling_patch_shape is None: pooling_patch_shape = patch_shape
        if pooling_p_h is None: pooling_p_h = p_h
        if pooling_p_h_prime is None: pooling_p_h_prime = p_h_prime
        if pooling_p_w is None: pooling_p_w = p_w
        if pooling_p_w_prime is None: pooling_p_w_prime = p_w_prime
        if pooling_p_d is None: pooling_p_d = p_d
        if pooling_p_d_prime is None: pooling_p_d_prime = p_d_prime
        if pooling_p_r is None: pooling_p_r = p_r
        if pooling_cd_h is None: pooling_cd_h = cd_h
        if pooling_cd_w is None: pooling_cd_w = cd_w
        if pooling_cd_d is None: pooling_cd_d = cd_d
        if pooling_packing_strategy is None: pooling_packing_strategy = packing_strategy

        if with_perceptron_residual is None: with_perceptron_residual = with_residual
        if with_perceptron_dual_lphm is None: with_perceptron_dual_lphm = with_dual_lphm
        if with_perceptron_lorr is None: with_perceptron_lorr = with_lorr
        if perceptron_r is None: perceptron_r = r
        if enable_perceptron_bias is None: enable_perceptron_bias = enable_bias

        layers = []
        for in_channel, out_channel in zip(channel_nums, channel_nums[1:]):
            print('conv in', h, w, d, in_channel)
            layer = grid_interdependence_layer(
                h=h, w=w, d=d,
                in_channel=in_channel, out_channel=out_channel,
                width=width,
                patch_shape=patch_shape,
                p_h=p_h, p_h_prime=p_h_prime,
                p_w=p_w, p_w_prime=p_w_prime,
                p_d=p_d, p_d_prime=p_d_prime,
                p_r=p_r,
                cd_h=cd_h, cd_w=cd_w, cd_d=cd_d,
                packing_strategy=packing_strategy,
                with_batch_norm=with_batch_norm,
                with_relu=with_relu,
                with_residual=with_residual,
                enable_bias=enable_bias,
                with_dual_lphm=with_dual_lphm,
                with_lorr=with_lorr, r=r,
                device=device, *args, **kwargs
            )
            h, w, d = layer.get_output_grid_shape()
            print('conv out', h, w, d, out_channel)
            layers.append(layer)

            # adding a pooling layer for a certain layer gaps
            if len(layers) % (pooling_layer_gaps+1) == pooling_layer_gaps:
                print('pooling in', h, w, d, out_channel)
                layer = grid_compression_layer(
                    h=h, w=w, d=d,
                    channel_num=out_channel,
                    pooling_metric=pooling_metric,
                    patch_shape=pooling_patch_shape,
                    p_h=pooling_p_h, p_h_prime=pooling_p_h_prime,
                    p_w=pooling_p_w, p_w_prime=pooling_p_w_prime,
                    p_d=pooling_p_d, p_d_prime=pooling_p_d_prime,
                    p_r=pooling_p_r,
                    cd_h=pooling_cd_h, cd_w=pooling_cd_w, cd_d=pooling_cd_d,
                    packing_strategy=pooling_packing_strategy,
                    with_dropout=with_dropout, p=p_pooling,
                    device=device, *args, **kwargs
                )
                h, w, d = layer.get_output_grid_shape()
                print('pooling out', h, w, d, out_channel)
                layers.append(layer)

                if patch_size_half_after_pooling:
                    print('halving patch size')
                    p_h, p_h_prime, p_w, p_w_prime, p_d, p_d_prime, p_r = parameter_scheduler(strategy='half', parameter_list=[p_h, p_h_prime, p_w, p_w_prime, p_d, p_d_prime, p_r])
                    pooling_p_h, pooling_p_h_prime, pooling_p_w, pooling_p_w_prime, pooling_p_d, pooling_p_d_prime, pooling_p_r = parameter_scheduler(strategy='half', parameter_list=[pooling_p_h, pooling_p_h_prime, pooling_p_w, pooling_p_w_prime, pooling_p_d, pooling_p_d_prime, pooling_p_r])

        # perceptron layers
        assert len(layers) >= 1
        m = layers[-1].get_n()
        dims = [m] + fc_dims
        for m, n in zip(dims, dims[1:]):
            print('fc in', m)
            layers.append(
                perceptron_layer(
                    m=m, n=n,
                    enable_bias=enable_perceptron_bias,
                    with_dual_lphm=with_perceptron_dual_lphm,
                    with_lorr=with_perceptron_lorr, r=perceptron_r,
                    with_residual=with_perceptron_residual,
                    channel_num=fc_channel_num,
                    width=width,
                    with_batch_norm=with_batch_norm and n != dims[-1],
                    with_relu=with_relu and n != dims[-1],
                    with_dropout=with_dropout and n != dims[-1], p=p_fc,
                    with_softmax=with_softmax and m == dims[-2] and n == dims[-1],
                    device=device, *args, **kwargs
                )
            )
            print('fc out', n)
        super().__init__(name=name, layers=layers, device=device, *args, **kwargs)

__init__(h, w, channel_nums, fc_dims, d=1, fc_channel_num=1, width=1, pooling_metric='batch_max', pooling_layer_gaps=2, patch_size_half_after_pooling=False, name='rpn_cnn', patch_shape='cuboid', p_h=None, p_h_prime=None, p_w=None, p_w_prime=None, p_d=0, p_d_prime=None, p_r=None, cd_h=None, cd_w=None, cd_d=1, packing_strategy='densest_packing', pooling_patch_shape=None, pooling_p_h=None, pooling_p_h_prime=None, pooling_p_w=None, pooling_p_w_prime=None, pooling_p_d=None, pooling_p_d_prime=None, pooling_p_r=None, pooling_cd_h=None, pooling_cd_w=None, pooling_cd_d=None, pooling_packing_strategy=None, with_batch_norm=True, with_relu=True, with_softmax=False, with_residual=False, with_dropout=True, p_pooling=0.25, p_fc=0.5, with_dual_lphm=False, with_lorr=False, r=3, enable_bias=True, with_perceptron_residual=None, with_perceptron_dual_lphm=None, with_perceptron_lorr=None, perceptron_r=None, enable_perceptron_bias=None, device='cpu', *args, **kwargs)

Initializes the CNN (Convolutional Neural Network) model.

This method constructs the CNN architecture with configurable convolutional layers, optional pooling layers, and fully connected perceptron layers. Various processing features like batch normalization, dropout, and residual connections can also be enabled.

Parameters:

Name	Type	Description	Default
h	int	Height of the input grid.	required
w	int	Width of the input grid.	required
channel_nums	list[int] | tuple[int]	Sequence of channel numbers for each convolutional layer.	required
fc_dims	list[int] | tuple[int]	Dimensions of the fully connected layers.	required
d	int	Depth of the input grid. Default is 1 (2D input).	1
fc_channel_num	int	Number of channels for the fully connected layers. Default is 1.	1
width	int	Number of parallel heads for each grid interdependence layer. Default is 1.	1
pooling_metric	str	Pooling metric used in pooling layers (e.g., 'batch_max'). Default is 'batch_max'.	'batch_max'
pooling_layer_gaps	int	Number of layers before adding a pooling layer. Default is 2.	2
patch_size_half_after_pooling	bool	If True, reduces patch size by half after pooling. Default is False.	False
name	str	Name of the CNN model. Default is 'rpn_cnn'.	'rpn_cnn'
patch_shape	str	Shape of the patches ('cuboid', 'cylinder', or 'sphere'). Default is 'cuboid'.	'cuboid'
p_h	int	Height and height prime of the patches. Default is None.	None
p_h_prime	int	Height and height prime of the patches. Default is None.	None
p_w	int	Width and width prime of the patches. Default is None.	None
p_w_prime	int	Width and width prime of the patches. Default is None.	None
p_d	int	Depth and depth prime of the patches. Default is 0 and None, respectively.	0
p_d_prime	int	Depth and depth prime of the patches. Default is 0 and None, respectively.	0
p_r	int	Radius of spherical or cylindrical patches. Default is None.	None
cd_h	int	Compression dimensions for height, width, and depth, respectively. Default is 1.	None
cd_w	int	Compression dimensions for height, width, and depth, respectively. Default is 1.	None
cd_d	int	Compression dimensions for height, width, and depth, respectively. Default is 1.	None
packing_strategy	str	Strategy for patch packing. Default is 'densest_packing'.	'densest_packing'
pooling_patch_shape	str	Shape of pooling patches. Default is None (same as patch_shape).	None
pooling_p_h	int	Height and height prime of pooling patches. Default is None.	None
pooling_p_h_prime	int	Height and height prime of pooling patches. Default is None.	None
pooling_p_w	int	Width and width prime of pooling patches. Default is None.	None
pooling_p_w_prime	int	Width and width prime of pooling patches. Default is None.	None
pooling_p_d	int	Depth and depth prime of pooling patches. Default is None.	None
pooling_p_d_prime	int	Depth and depth prime of pooling patches. Default is None.	None
pooling_p_r	int	Radius of pooling patches. Default is None.	None
pooling_cd_h	int	Compression dimensions for pooling patches. Default is None.	None
pooling_cd_w	int	Compression dimensions for pooling patches. Default is None.	None
pooling_cd_d	int	Compression dimensions for pooling patches. Default is None.	None
pooling_packing_strategy	str	Packing strategy for pooling patches. Default is None (same as packing_strategy).	None
with_batch_norm	bool	If True, applies batch normalization. Default is True.	True
with_relu	bool	If True, applies ReLU activation. Default is True.	True
with_softmax	bool	If True, applies softmax activation to the final output. Default is False.	False
with_residual	bool	If True, includes residual connections. Default is False.	False
with_dropout	bool	If True, applies dropout. Default is True.	True
p_pooling	float	Dropout probability for pooling layers. Default is 0.25.	0.25
p_fc	float	Dropout probability for fully connected layers. Default is 0.5.	0.5
with_dual_lphm	bool	If True, enables dual low-parametric high-order interdependence. Default is False.	False
with_lorr	bool	If True, enables low-rank parameterized interdependence. Default is False.	False
r	int	Rank parameter for low-rank interdependence. Default is 3.	3
enable_bias	bool	If True, includes bias in the layers. Default is True.	True
with_perceptron_residual	bool	If True, includes residual connections in perceptron layers. Default is None (inherits from with_residual).	None
with_perceptron_dual_lphm	bool	If True, enables dual low-parametric high-order interdependence in perceptron layers. Default is None.	None
with_perceptron_lorr	bool	If True, enables low-rank parameterized interdependence in perceptron layers. Default is None.	None
perceptron_r	int	Rank parameter for perceptron layers. Default is None.	None
enable_perceptron_bias	bool	If True, includes bias in perceptron layers. Default is None.	None
device	str	Device for computation ('cpu' or 'cuda'). Default is 'cpu'.	'cpu'
*args	optional	Additional positional arguments for layers.	()
**kwargs	optional	Additional keyword arguments for layers.	{}

Raises:

Type	Description
ValueError	If input dimensions are invalid or improperly configured.

Source code in tinybig/model/rpn_cnn.py

def __init__(
    self,
    h: int, w: int,
    channel_nums: list[int] | tuple[int],
    fc_dims: list[int] | tuple[int],
    d: int = 1,
    fc_channel_num: int = 1,
    width: int = 1,
    pooling_metric: str = 'batch_max',
    pooling_layer_gaps: int = 2,
    patch_size_half_after_pooling: bool = False,
    name: str = 'rpn_cnn',
    # patch structure parameters for interdependence
    patch_shape: str = 'cuboid',
    p_h: int = None, p_h_prime: int = None,
    p_w: int = None, p_w_prime: int = None,
    p_d: int = 0, p_d_prime: int = None,
    p_r: int = None,
    cd_h: int = None, cd_w: int = None, cd_d: int = 1,
    packing_strategy: str = 'densest_packing',
    # patch structure parameters for compression
    pooling_patch_shape: str = None,
    pooling_p_h: int = None, pooling_p_h_prime: int = None,
    pooling_p_w: int = None, pooling_p_w_prime: int = None,
    pooling_p_d: int = None, pooling_p_d_prime: int = None,
    pooling_p_r: int = None,
    pooling_cd_h: int = None, pooling_cd_w: int = None, pooling_cd_d: int = None,
    pooling_packing_strategy: str = None,
    # output processing function parameters
    with_batch_norm: bool = True,
    with_relu: bool = True,
    with_softmax: bool = False,
    with_residual: bool = False,
    with_dropout: bool = True, p_pooling: float = 0.25, p_fc: float = 0.5,
    # parameter reconciliation function parameters
    with_dual_lphm: bool = False,
    with_lorr: bool = False, r: int = 3,
    enable_bias: bool = True,
    # parameter reconciliation function parameters
    with_perceptron_residual: bool = None,
    with_perceptron_dual_lphm: bool = None,
    with_perceptron_lorr: bool = None, perceptron_r: int = None,
    enable_perceptron_bias: bool = None,
    # other parameters
    device: str = 'cpu', *args, **kwargs
):
    """
    Initializes the CNN (Convolutional Neural Network) model.

    This method constructs the CNN architecture with configurable convolutional layers, optional pooling layers,
    and fully connected perceptron layers. Various processing features like batch normalization, dropout,
    and residual connections can also be enabled.

    Parameters
    ----------
    h : int
        Height of the input grid.
    w : int
        Width of the input grid.
    channel_nums : list[int] | tuple[int]
        Sequence of channel numbers for each convolutional layer.
    fc_dims : list[int] | tuple[int]
        Dimensions of the fully connected layers.
    d : int, optional
        Depth of the input grid. Default is 1 (2D input).
    fc_channel_num : int, optional
        Number of channels for the fully connected layers. Default is 1.
    width : int, optional
        Number of parallel heads for each grid interdependence layer. Default is 1.
    pooling_metric : str, optional
        Pooling metric used in pooling layers (e.g., 'batch_max'). Default is 'batch_max'.
    pooling_layer_gaps : int, optional
        Number of layers before adding a pooling layer. Default is 2.
    patch_size_half_after_pooling : bool, optional
        If True, reduces patch size by half after pooling. Default is False.
    name : str, optional
        Name of the CNN model. Default is 'rpn_cnn'.
    patch_shape : str, optional
        Shape of the patches ('cuboid', 'cylinder', or 'sphere'). Default is 'cuboid'.
    p_h, p_h_prime : int, optional
        Height and height prime of the patches. Default is None.
    p_w, p_w_prime : int, optional
        Width and width prime of the patches. Default is None.
    p_d, p_d_prime : int, optional
        Depth and depth prime of the patches. Default is 0 and None, respectively.
    p_r : int, optional
        Radius of spherical or cylindrical patches. Default is None.
    cd_h, cd_w, cd_d : int, optional
        Compression dimensions for height, width, and depth, respectively. Default is 1.
    packing_strategy : str, optional
        Strategy for patch packing. Default is 'densest_packing'.
    pooling_patch_shape : str, optional
        Shape of pooling patches. Default is None (same as `patch_shape`).
    pooling_p_h, pooling_p_h_prime : int, optional
        Height and height prime of pooling patches. Default is None.
    pooling_p_w, pooling_p_w_prime : int, optional
        Width and width prime of pooling patches. Default is None.
    pooling_p_d, pooling_p_d_prime : int, optional
        Depth and depth prime of pooling patches. Default is None.
    pooling_p_r : int, optional
        Radius of pooling patches. Default is None.
    pooling_cd_h, pooling_cd_w, pooling_cd_d : int, optional
        Compression dimensions for pooling patches. Default is None.
    pooling_packing_strategy : str, optional
        Packing strategy for pooling patches. Default is None (same as `packing_strategy`).
    with_batch_norm : bool, optional
        If True, applies batch normalization. Default is True.
    with_relu : bool, optional
        If True, applies ReLU activation. Default is True.
    with_softmax : bool, optional
        If True, applies softmax activation to the final output. Default is False.
    with_residual : bool, optional
        If True, includes residual connections. Default is False.
    with_dropout : bool, optional
        If True, applies dropout. Default is True.
    p_pooling : float, optional
        Dropout probability for pooling layers. Default is 0.25.
    p_fc : float, optional
        Dropout probability for fully connected layers. Default is 0.5.
    with_dual_lphm : bool, optional
        If True, enables dual low-parametric high-order interdependence. Default is False.
    with_lorr : bool, optional
        If True, enables low-rank parameterized interdependence. Default is False.
    r : int, optional
        Rank parameter for low-rank interdependence. Default is 3.
    enable_bias : bool, optional
        If True, includes bias in the layers. Default is True.
    with_perceptron_residual : bool, optional
        If True, includes residual connections in perceptron layers. Default is None (inherits from `with_residual`).
    with_perceptron_dual_lphm : bool, optional
        If True, enables dual low-parametric high-order interdependence in perceptron layers. Default is None.
    with_perceptron_lorr : bool, optional
        If True, enables low-rank parameterized interdependence in perceptron layers. Default is None.
    perceptron_r : int, optional
        Rank parameter for perceptron layers. Default is None.
    enable_perceptron_bias : bool, optional
        If True, includes bias in perceptron layers. Default is None.
    device : str, optional
        Device for computation ('cpu' or 'cuda'). Default is 'cpu'.
    *args : optional
        Additional positional arguments for layers.
    **kwargs : optional
        Additional keyword arguments for layers.

    Raises
    ------
    ValueError
        If input dimensions are invalid or improperly configured.
    """
    print('############# rpn-cnn model architecture ############')

    if pooling_patch_shape is None: pooling_patch_shape = patch_shape
    if pooling_p_h is None: pooling_p_h = p_h
    if pooling_p_h_prime is None: pooling_p_h_prime = p_h_prime
    if pooling_p_w is None: pooling_p_w = p_w
    if pooling_p_w_prime is None: pooling_p_w_prime = p_w_prime
    if pooling_p_d is None: pooling_p_d = p_d
    if pooling_p_d_prime is None: pooling_p_d_prime = p_d_prime
    if pooling_p_r is None: pooling_p_r = p_r
    if pooling_cd_h is None: pooling_cd_h = cd_h
    if pooling_cd_w is None: pooling_cd_w = cd_w
    if pooling_cd_d is None: pooling_cd_d = cd_d
    if pooling_packing_strategy is None: pooling_packing_strategy = packing_strategy

    if with_perceptron_residual is None: with_perceptron_residual = with_residual
    if with_perceptron_dual_lphm is None: with_perceptron_dual_lphm = with_dual_lphm
    if with_perceptron_lorr is None: with_perceptron_lorr = with_lorr
    if perceptron_r is None: perceptron_r = r
    if enable_perceptron_bias is None: enable_perceptron_bias = enable_bias

    layers = []
    for in_channel, out_channel in zip(channel_nums, channel_nums[1:]):
        print('conv in', h, w, d, in_channel)
        layer = grid_interdependence_layer(
            h=h, w=w, d=d,
            in_channel=in_channel, out_channel=out_channel,
            width=width,
            patch_shape=patch_shape,
            p_h=p_h, p_h_prime=p_h_prime,
            p_w=p_w, p_w_prime=p_w_prime,
            p_d=p_d, p_d_prime=p_d_prime,
            p_r=p_r,
            cd_h=cd_h, cd_w=cd_w, cd_d=cd_d,
            packing_strategy=packing_strategy,
            with_batch_norm=with_batch_norm,
            with_relu=with_relu,
            with_residual=with_residual,
            enable_bias=enable_bias,
            with_dual_lphm=with_dual_lphm,
            with_lorr=with_lorr, r=r,
            device=device, *args, **kwargs
        )
        h, w, d = layer.get_output_grid_shape()
        print('conv out', h, w, d, out_channel)
        layers.append(layer)

        # adding a pooling layer for a certain layer gaps
        if len(layers) % (pooling_layer_gaps+1) == pooling_layer_gaps:
            print('pooling in', h, w, d, out_channel)
            layer = grid_compression_layer(
                h=h, w=w, d=d,
                channel_num=out_channel,
                pooling_metric=pooling_metric,
                patch_shape=pooling_patch_shape,
                p_h=pooling_p_h, p_h_prime=pooling_p_h_prime,
                p_w=pooling_p_w, p_w_prime=pooling_p_w_prime,
                p_d=pooling_p_d, p_d_prime=pooling_p_d_prime,
                p_r=pooling_p_r,
                cd_h=pooling_cd_h, cd_w=pooling_cd_w, cd_d=pooling_cd_d,
                packing_strategy=pooling_packing_strategy,
                with_dropout=with_dropout, p=p_pooling,
                device=device, *args, **kwargs
            )
            h, w, d = layer.get_output_grid_shape()
            print('pooling out', h, w, d, out_channel)
            layers.append(layer)

            if patch_size_half_after_pooling:
                print('halving patch size')
                p_h, p_h_prime, p_w, p_w_prime, p_d, p_d_prime, p_r = parameter_scheduler(strategy='half', parameter_list=[p_h, p_h_prime, p_w, p_w_prime, p_d, p_d_prime, p_r])
                pooling_p_h, pooling_p_h_prime, pooling_p_w, pooling_p_w_prime, pooling_p_d, pooling_p_d_prime, pooling_p_r = parameter_scheduler(strategy='half', parameter_list=[pooling_p_h, pooling_p_h_prime, pooling_p_w, pooling_p_w_prime, pooling_p_d, pooling_p_d_prime, pooling_p_r])

    # perceptron layers
    assert len(layers) >= 1
    m = layers[-1].get_n()
    dims = [m] + fc_dims
    for m, n in zip(dims, dims[1:]):
        print('fc in', m)
        layers.append(
            perceptron_layer(
                m=m, n=n,
                enable_bias=enable_perceptron_bias,
                with_dual_lphm=with_perceptron_dual_lphm,
                with_lorr=with_perceptron_lorr, r=perceptron_r,
                with_residual=with_perceptron_residual,
                channel_num=fc_channel_num,
                width=width,
                with_batch_norm=with_batch_norm and n != dims[-1],
                with_relu=with_relu and n != dims[-1],
                with_dropout=with_dropout and n != dims[-1], p=p_fc,
                with_softmax=with_softmax and m == dims[-2] and n == dims[-1],
                device=device, *args, **kwargs
            )
        )
        print('fc out', n)
    super().__init__(name=name, layers=layers, device=device, *args, **kwargs)