Unitary CLIP

Contrastive pre-training of an unitary encoder

Layers

source

RotaryMultiheadAttention


def RotaryMultiheadAttention(
    in_dim:int, embed_dim:int, num_heads:int, bias:bool=True, p_rope:float=1.0, max_seq_len:int=4096,
    base_rope:float=10000, enable_qk_norm:bool=False
)->None:

MultiheadAttention described in the paper: Attention Is All You Need (https://arxiv.org/abs/1706.03762). We add a rotary position encoding (RoPE).

The attention core is F.scaled_dot_attention from pytorch. Could be switched to https://github.com/Dao-AILab/flash-attention or xFormers.

source

FeedForwardBlock


def FeedForwardBlock(
    in_dim:int, hidden_dim:int, dropout:float=0.0
)->None:

A small dense feed-forward network as used in transformers. Assumes channel last. Inspired by https://arxiv.org/pdf/2401.11605 and added from https://arxiv.org/pdf/2002.05202 a modification to SiGLU structure.

Unitary-text encoder

source

UnitaryEncoderAttnBlock


def UnitaryEncoderAttnBlock(
    ch:int, y_emb_size:int, num_heads:int, dropout:float=0.0, p_rope:float=1.0, base_rope:float=10000
)->None:

A self-attention block with 2d-RoPE.

source

UnitaryTextEncoderConfig


def UnitaryTextEncoderConfig(
    text_embed_ch:int, text_encoding_ch:int, text_attn_num_heads:int, text_attn_depth:int, unitary_encoding_ch:int,
    unitary_downscale_factor:int, main_num_heads:int, main_depth:int, use_rope:bool, p_rope:float, base_rope:float,
    dropout:float
)->None:

source

UnitaryTextEncoder


def UnitaryTextEncoder(
    text_embed_ch:int, text_encoding_ch:int, text_attn_num_heads:int, text_attn_depth:int, unitary_encoding_ch:int,
    unitary_downscale_factor:int, main_num_heads:int, main_depth:int, use_rope:bool, p_rope:float, base_rope:float,
    dropout:float
)->None:

A basic nn.Module with IO functionality.

Circuit encoder

source

SelfAttnBlock


def SelfAttnBlock(
    ch:int, num_heads:int, dropout:float=0.0, p_rope:float=1.0, base_rope:float=10000
)->None:

A self-attention block with RoPE.

source

PackingTransformer


def PackingTransformer(
    ch:int, depth:int, num_heads:int, dropout:float=0.0, p_rope:float=1.0, base_rope:float=10000
)->None:

The first stage packing/unpacking transformers of the CirDiT model. Applies a RoPE for time dimension only, not on spatial dimension.

source

CoreTransformer


def CoreTransformer(
    ch:int, depth:int, num_heads:int, dropout:float=0.0, p_rope:float=1.0, base_rope:float=10000
)->None:

The main transformer of the CirDiT model. Applies a RoPE for time dimension.

source

CircuitEncoderConfig


def CircuitEncoderConfig(
    embedder_config:dict, ch_packing:int, ch_core:int, depth_packing:int, depth_core:int, num_heads_packing:int,
    num_heads_core:int, dropout:float, p_rope:float, base_rope:float
)->None:

source

CircuitEncoder


def CircuitEncoder(
    embedder_config:Optional, ch_packing:int, ch_core:int, depth_packing:int, depth_core:int, num_heads_packing:int,
    num_heads_core:int, dropout:float=0.0, p_rope:float=1.0, base_rope:float=10000, embedder:Optional=None
)->None:

A basic nn.Module with IO functionality.

Unitary CLIP model

source

UnitaryCLIPConfig


def UnitaryCLIPConfig(
    text_encoder_config:dict, clip_embed_size:int
)->None:

source

UnitaryCLIP


def UnitaryCLIP(
    text_encoder_config:Optional, unitary_text_encoder:UnitaryTextEncoder, circuit_encoder:CircuitEncoder,
    clip_embed_size:int, text_encoder:Optional=None
)->None:

A basic nn.Module with IO functionality.

Back to top