spacetimeformer.spacetimeformer_model.nn package

class spacetimeformer.spacetimeformer_model.nn.attn.AttentionLayer(attention, d_model, n_heads, dropout_qkv=0.0, d_keys=None, d_values=None, mix=False)

Bases: torch.nn.modules.module.Module

forward(queries, keys, values, attn_mask, output_attn=False)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class spacetimeformer.spacetimeformer_model.nn.attn.BenchmarkAttention(*args, **kwargs)

Bases: torch.nn.modules.module.Module

forward(queries, keys, values, attn_mask, output_attn=False)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class spacetimeformer.spacetimeformer_model.nn.attn.FullAttention(mask_flag=False, scale=None, attention_dropout=0.1)

Bases: torch.nn.modules.module.Module

forward(queries, keys, values, attn_mask, output_attn=False)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class spacetimeformer.spacetimeformer_model.nn.attn.LocalAttentionLayer(attention, d_y, d_model, n_heads, dropout_qkv=0.0)

Bases: torch.nn.modules.module.Module

forward(queries, keys, values, attn_mask=None, output_attn=False)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class spacetimeformer.spacetimeformer_model.nn.attn.NystromSelfAttention(d_model, n_heads, num_landmarks=256, pinv_iterations=6, attention_dropout=0.0, residual=False, residual_conv_kernel=33, eps=1e-08)

Bases: torch.nn.modules.module.Module

forward(x, x_, x__, attn_mask=None, output_attn=False)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class spacetimeformer.spacetimeformer_model.nn.attn.PerformerAttention(mask_flag=False, dim_heads=None, ortho_scaling=0, feature_redraw_interval=1000, kernel='softmax')

Bases: performer_pytorch.performer_pytorch.FastAttention

forward(queries, keys, values, attn_mask, output_attn=False)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class spacetimeformer.spacetimeformer_model.nn.attn.ProbAttention(mask_flag=True, factor=5, scale=None, attention_dropout=0.1)

Bases: torch.nn.modules.module.Module

forward(queries, keys, values, attn_mask, output_attn=False)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class spacetimeformer.spacetimeformer_model.nn.data_dropout.DataDropout(dropout=None)

Bases: torch.nn.modules.module.Module

forward(embed)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class spacetimeformer.spacetimeformer_model.nn.decoder.Decoder(layers, norm_layer=None, emb_dropout=0.0, data_dropout=0.0)

Bases: torch.nn.modules.module.Module

forward(val_time_emb, space_emb, cross, x_mask=None, cross_mask=None)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class spacetimeformer.spacetimeformer_model.nn.decoder.DecoderLayer(global_self_attention, local_self_attention, global_cross_attention, local_cross_attention, d_model, d_ff=None, dropout_ff=0.1, activation='relu', post_norm=True, norm='layer')

Bases: torch.nn.modules.module.Module

forward(x, cross, x_mask=None, cross_mask=None)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class spacetimeformer.spacetimeformer_model.nn.embed.SpacetimeformerEmbedding(d_y, d_x, d_model=256, time_emb_dim=6, method='spatio-temporal', downsample_convs=1, start_token_len=0, null_value=None)

Bases: torch.nn.modules.module.Module

spatio_temporal_embed(y, x, is_encoder=True)

temporal_embed(y, x, is_encoder=True)

GIVEN = True

SPACE = True

TIME = True

VAL = True

training: bool

class spacetimeformer.spacetimeformer_model.nn.encoder.DownsampleConv(c_in)

Bases: torch.nn.modules.module.Module

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class spacetimeformer.spacetimeformer_model.nn.encoder.Encoder(attn_layers, conv_layers, norm_layer, emb_dropout=0.0, data_dropout=0.0)

Bases: torch.nn.modules.module.Module

forward(val_time_emb, space_emb, attn_mask=None, output_attn=False)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class spacetimeformer.spacetimeformer_model.nn.encoder.EncoderLayer(global_attention, local_attention, d_model, d_ff=None, dropout_ff=0.1, activation='relu', post_norm=True, norm='layer')

Bases: torch.nn.modules.module.Module

forward(x, attn_mask=None, output_attn=False)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class spacetimeformer.spacetimeformer_model.nn.encoder.Normalization(method, d_model=None)

Bases: torch.nn.modules.module.Module

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class spacetimeformer.spacetimeformer_model.nn.encoder.VariableDownsample(d_y, d_model)

Bases: torch.nn.modules.module.Module

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class spacetimeformer.spacetimeformer_model.nn.model.Spacetimeformer(d_y: int = 1, d_x: int = 4, start_token_len: int = 64, attn_factor: int = 5, d_model: int = 512, n_heads: int = 8, e_layers: int = 2, d_layers: int = 2, d_ff: int = 512, time_emb_dim: int = 6, dropout_emb: float = 0.05, dropout_token: float = 0.05, dropout_attn_out: float = 0.05, dropout_ff: float = 0.05, dropout_qkv: float = 0.05, global_self_attn: str = 'performer', local_self_attn: str = 'none', global_cross_attn: str = 'performer', local_cross_attn: str = 'none', performer_attn_kernel: str = 'relu', performer_redraw_interval: int = 250, embed_method: str = 'spatio-temporal', activation: str = 'gelu', post_norm: bool = True, norm: str = 'layer', initial_downsample_convs: int = 0, intermediate_downsample_convs: int = 0, device=device(type='cuda', index=0), null_value: Optional[float] = None, verbose: bool = True)

Bases: torch.nn.modules.module.Module

forward(x_enc, x_mark_enc, x_dec, x_mark_dec, enc_self_mask=None, dec_self_mask=None, dec_enc_mask=None, output_attention=False)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

PowerNorm code from https://github.com/sIncerass/powernorm/blob/master/fairseq/modules/norms/mask_powernorm.py

class spacetimeformer.spacetimeformer_model.nn.powernorm.MaskPowerNorm(num_features, eps=1e-05, alpha_fwd=0.9, alpha_bkw=0.9, affine=True, warmup_iters=10000, group_num=1)

Bases: torch.nn.modules.module.Module

An implementation of masked batch normalization, used for testing the numerical stability.

extra_repr()

Set the extra representation of the module

To print customized extra information, you should re-implement this method in your own modules. Both single-line and multi-line strings are acceptable.

forward(input, pad_mask=None, is_encoder=False)

input: T x B x C -> B x C x T

: B x C x T -> T x B x C

pad_mask: B x T (padding is True)

training: bool

class spacetimeformer.spacetimeformer_model.nn.scalenorm.ScaleNorm(dim, eps=1e-05)

Bases: torch.nn.modules.module.Module

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool