mbapy.dl_torch.m

This module provides classes for the MA TT (Multi-Attention Time-Transformer) model.

Classes

TransCfg

A class that represents the configuration for the transformer layers.

Parameters:
- hid_dim (int): The hidden dimension size.
- pf_dim (Optional[int]): The position-wise feedforward dimension size. Defaults to None.
- n_heads (int): The number of attention heads. Defaults to 8.
- n_layers (int): The number of transformer layers. Defaults to 3.
- dropout (float): The dropout rate. Defaults to 0.3.
- trans_layer (str): The type of transformer layer to use. Defaults to 'EncoderLayer'.
- out_layer (Optional[str]): The type of output layer to use. Defaults to None.
- q_len (int): The length of the query sequence. Defaults to -1.
- class_num (int): The number of classes. Defaults to -1.
- kwargs (Dict[str, Union[int, str, bool]]): Additional keyword arguments. Defaults to {}.

Methods:
- init(self, hid_dim:int, pf_dim:Optional[int] = None, n_heads:int = 8, n_layers:int = 3, dropout:float = 0.3, trans_layer:str = 'EncoderLayer', out_layer:Optional[str] = None, q_len:int = -1, class_num:int = -1, kwargs): Initializes the TransCfg object with the given arguments.
- str(self): Returns a string representation of the TransCfg object.
- toDict(self): Converts the TransCfg object to a dictionary.
- gen(self, layer:str = None, kwargs): Generates a transformer-like layer using the configuration.

Example:

cfg = TransCfg(hid_dim=512, n_heads=8, n_layers=6)
print(cfg)

LayerCfg

A class that represents the configuration for each layer in the MA TT model.

Parameters:
- inc (int): The input channel size.
- outc (int): The output channel size.
- kernel_size (int): The kernel size.
- stride (int): The stride size.
- layer (str): The type of layer to use.
- sa_layer (Optional[str]): The type of self-attention layer to use. Defaults to None.
- trans_layer (Optional[str]): The type of transformer layer to use. Defaults to None.
- avg_size (int): The size of the average pooling. Defaults to -1.
- trans_cfg (Optional[TransCfg]): The configuration for the transformer layers. Defaults to None.
- use_SA (bool): Whether to use self-attention. Defaults to False.
- use_trans (bool): Whether to use transformer layers. Defaults to False.

Methods:
- init(self, inc:int, outc:int, kernel_size:int, stride:int, layer:str, sa_layer:Optional[str] = None, trans_layer:Optional[str] = None, avg_size:int = -1, trans_cfg:Optional[TransCfg] = None, use_SA:bool = False, use_trans:bool = False): Initializes the LayerCfg object with the given arguments.
- str(self): Returns a string representation of the LayerCfg object.

Example:

cfg = LayerCfg(3, 64, 3, 1, 'Conv2d')
print(cfg)

COneDLayer

A class that represents a convolutional layer in the MA TT model.

Parameters:
- cfg (LayerCfg): The configuration for the layer.
- device (str): The device to use. Defaults to 'cuda'.

Methods:
- init(self, cfg:LayerCfg, device = 'cuda', **kwargs): Initializes the COneDLayer object with the given configuration and device.
- forward(self, x): Performs a forward pass of the COneDLayer.

Example:

layer = COneDLayer(cfg)
output = layer(x)

MAlayer

A class that represents a multi-attention layer in the MA TT model.

Parameters:
- cfg (LayerCfg): The configuration for the layer.

Methods:
- init(self, cfg:LayerCfg, **kwargs): Initializes the MAlayer object with the given configuration.
- forward(self, x): Performs a forward pass of the MAlayer.

Example:

layer = MAlayer(cfg)
output = layer(x)

MAvlayer

A class that represents a multi-attention layer with average pooling in the MA TT model.

Parameters:
- cfg (LayerCfg): The configuration for the layer.

Methods:
- init(self, cfg:LayerCfg, **kwargs): Initializes the MAvlayer object with the given configuration.
- forward(self, x): Performs a forward pass of the MAvlayer.

Example:

layer = MAvlayer(cfg)
output = layer(x)

MATTPE

A class that represents the MA TT model with positional encoding.

Parameters:
- args (GlobalSettings): The global settings for the model.
- cfg (list[LayerCfg]): The list of layer configurations.
- layer (MAvlayer): The type of layer to use.
- tail_trans_cfg (TransCfg): The configuration for the transformer layers. Defaults to None.

Methods:
- init(self, args: GlobalSettings, cfg:list[LayerCfg], layer:MAvlayer, tail_trans_cfg:TransCfg = None, **kwargs): Initializes the MATTPE object with the given arguments.
- forward(self, x): Performs a forward pass of the MATTPE model.

Example:

model = MATTPE(args, cfg, layer)
output = model(x)

SCANNTTP

A class that represents the MA TT model with SCANN layer.

Parameters:
- args (GlobalSettings): The global settings for the model.
- cfg (list[LayerCfg]): The list of layer configurations.
- layer (SCANlayer): The type of layer to use.
- tail_trans_cfg (TransCfg): The configuration for the transformer layers. Defaults to None.

Methods:
- init(self, args: GlobalSettings, cfg:list[LayerCfg], layer:SCANlayer, tail_trans_cfg:TransCfg = None, **kwargs): Initializes the SCANNTTP object with the given arguments.
- forward(self, x): Performs a forward pass of the SCANNTTP model.

Example:

model = SCANNTTP(args, cfg, layer)
output = model(x)

MATTP_ViT

A class that represents the MA TT model with ViT-like structure.

Parameters:
- args (GlobalSettings): The global settings for the model.
- cfg (list[LayerCfg]): The list of layer configurations.
- layer (MAvlayer): The type of layer to use.
- tail_trans_cfg (TransCfg): The configuration for the transformer layers. Defaults to None.

Methods:
- init(self, args: GlobalSettings, cfg:list[LayerCfg], layer:MAvlayer, tail_trans_cfg:TransCfg = None, **kwargs): Initializes the MATTP_ViT object with the given arguments.
- forward(self, x): Performs a forward pass of the MATTP_ViT model.

Example:

model = MATTP_ViT(args, cfg, layer)
output = model(x)

MATTP

A class that represents the MA TT model.

Parameters:
- args (GlobalSettings): The global settings for the model.
- cfg (list[LayerCfg]): The list of layer configurations.
- layer (MAlayer): The type of layer to use.
- tail_trans_cfg (TransCfg): The configuration for the transformer layers. Defaults to None.

Methods:
- init(self, args: GlobalSettings, cfg:list[LayerCfg], layer:MAlayer, tail_trans_cfg:TransCfg = None, **kwargs): Initializes the MATTP object with the given arguments.
- forward(self, x): Performs a forward pass of the MATTP model.

Example:

model = MATTP(args, cfg, layer)
output = model(x)

COneD

A class that represents the MA TT model with 1D convolutional layers.

Parameters:
- args (GlobalSettings): The global settings for the model.
- cfg (list[LayerCfg]): The list of layer configurations.
- layer (COneDLayer): The type of layer to use.
- tail_trans_cfg (TransCfg): The configuration for the transformer layers. Defaults to None.

Methods:
- init(self, args: GlobalSettings, cfg:list[LayerCfg], layer:COneDLayer, tail_trans_cfg:TransCfg = None, **kwargs): Initializes the COneD object with the given arguments.
- forward(self, x): Performs a forward pass of the COneD model.

Example:

model = COneD(args, cfg, layer)
output = model(x)

MAvTTP

A class that represents the MA TT model with average pooling and permute.

Parameters:
- args (GlobalSettings): The global settings for the model.
- cfg (list[LayerCfg]): The list of layer configurations.
- layer (MAvlayer): The type of layer to use.
- tail_trans_cfg (TransCfg): The configuration for the transformer layers. Defaults to None.

Methods:
- init(self, args: GlobalSettings, cfg:list[LayerCfg], layer:MAvlayer, tail_trans_cfg:TransCfg = None, **kwargs): Initializes the MAvTTP object with the given arguments.

Example:

model = MAvTTP(args, cfg, layer)
output = model(x)

MATTPE

A class that represents the MA TT model with positional encoding and permute.

Parameters:
- args (GlobalSettings): The global settings for the model.
- cfg (list[LayerCfg]): The list of layer configurations.
- layer (MAvlayer): The type of layer to use.
- tail_trans_cfg (TransCfg): The configuration for the transformer layers. Defaults to None.

Methods:
- init(self, args: GlobalSettings, cfg:list[LayerCfg], layer:MAvlayer, tail_trans_cfg:TransCfg = None, **kwargs): Initializes the MATTPE object with the given arguments.
- forward(self, x): Performs a forward pass of the MATTPE model.

Example:

model = MATTPE(args, cfg, layer)
output = model(x)

Constants

str2net

A dictionary that maps strings to network classes.

Example:

str2net = {
    'EncoderLayer':bb.EncoderLayer,
    'OutEncoderLayer':bb.OutEncoderLayer,
    'OutEncoderLayerAvg':bb.OutEncoderLayerAvg,
    'Trans':bb.Trans,
    'TransPE':bb.TransPE,
    'TransAvg':bb.TransAvg,

    'SeparableConv2d':bb.SeparableConv2d,
    'ResBlock':bb.ResBlock,
    'ResBlockR':bb.ResBlockR,
    'SABlock':bb.SABlock,
    'SABlockR':bb.SABlockR,
    'SABlock1D':bb.SABlock1D,
    'SABlock1DR':bb.SABlock1DR,

    'COneDLayer':COneDLayer,
    'MAlayer':MAlayer,
    'MAvlayer':MAvlayer,
    'SCANlayer':SCANlayer,
}