fishy.models.deep.augformer_v2

AugFormerV2: Enhanced version of AugFormer with advancements from Parameter Golf.

Features implemented: - QK-Gain: Learnable per-head query scaling. - Parallel Residuals: Attention and MLP branches in parallel (GPT-J style). - Depth Recurrence: Selective looping of transformer blocks. - LeakyReLU(0.5)^2 Activation: Non-monotonic activation for efficiency. - TTT Support: forward_ttt method for entropy minimization at test-time.

Classes

class fishy.models.deep.augformer_v2.AugFormerV2(input_dim: int, output_dim: int, hidden_dim: int = 512, num_layers: int = 4, num_heads: int = 8, dropout: float = 0.3, num_views: int = 6, use_xsa: bool = False, use_qk_gain: bool = False, use_parallel_residuals: bool = False, recurrence_layers: List[int] | None = None, use_leaky_sq: bool = False, **kwargs)

Bases: AugFormer

Enhanced Augmentation-as-Sequence Transformer. Allows toggling Parameter Golf advancements for benchmarking.

__init__(input_dim: int, output_dim: int, hidden_dim: int = 512, num_layers: int = 4, num_heads: int = 8, dropout: float = 0.3, num_views: int = 6, use_xsa: bool = False, use_qk_gain: bool = False, use_parallel_residuals: bool = False, recurrence_layers: List[int] | None = None, use_leaky_sq: bool = False, **kwargs) → None

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor, return_attention: bool = False, *args, **kwargs) → Tensor | Tuple[Tensor, List]

Define 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.

class fishy.models.deep.augformer_v2.LeakyReluSq(neg_slope: float = 0.5)

Bases: Module

Non-monotonic activation used in high-performing Golf models.

__init__(neg_slope: float = 0.5)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor) → Tensor

Define 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.

class fishy.models.deep.augformer_v2.MultiHeadAttentionV2(embed_dim: int, num_heads: int, use_xsa: bool = False, use_qk_gain: bool = False)

Bases: MultiHeadAttention

__init__(embed_dim: int, num_heads: int, use_xsa: bool = False, use_qk_gain: bool = False)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor) → Tensor

Define 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.

class fishy.models.deep.augformer_v2.TransformerBlockV2(embed_dim: int, num_heads: int, mlp_ratio: int = 2, dropout: float = 0.1, use_xsa: bool = False, use_qk_gain: bool = False, use_parallel_residuals: bool = False, use_leaky_sq: bool = False)

Bases: TransformerBlock

__init__(embed_dim: int, num_heads: int, mlp_ratio: int = 2, dropout: float = 0.1, use_xsa: bool = False, use_qk_gain: bool = False, use_parallel_residuals: bool = False, use_leaky_sq: bool = False)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor) → Tensor

Define 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.

s