LiRA / lira /core_modules.py
asdf98's picture
Add lira/core_modules.py
f5c1b06 verified
"""
LiRA Core Modules: Gated State-Space Backbone (GS3B)
Mathematical Foundation:
========================
Traditional transformers use self-attention: O_i = softmax(Q_i K^T / sqrt(d)) V
This is O(N^2) in sequence length - prohibitive for high-res images.
Our approach combines three key innovations:
1. SELECTIVE STATE SPACE (from Mamba/S6):
State evolution: h_t = A_t * h_{t-1} + B_t * x_t
Output: y_t = C_t * h_t + D * x_t
Where A_t, B_t, C_t are INPUT-DEPENDENT (selective) - this is the key insight
from Mamba that makes SSMs competitive with attention.
2. BIDIRECTIONAL GATED SCANNING (from DiM + RWKV-7):
Images are 2D, not 1D. We scan in 4 directions:
- Horizontal L→R, R→L
- Vertical T→B, B→T
Each direction maintains its own state. A learned gate fuses them:
y = gate * [y_lr; y_rl; y_tb; y_bt]
From RWKV-7 we take the generalized delta rule for state updates:
S_t = S_{t-1} * (diag(w_t) - k_t^T (a_t ⊗ k_t)) + v_t^T k_t
This gives us input-dependent decay with O(N) complexity.
3. FREQUENCY-AWARE PROCESSING (from DiMSUM):
We apply lightweight wavelet decomposition to separate structure from detail,
process each frequency band with appropriate granularity, then recombine.
Low-freq (structure) → fewer tokens, heavier processing
High-freq (detail) → more tokens, lighter processing
Combined complexity: O(N * d * H) where N=tokens, d=state_dim, H=num_heads
For 1024px with f32 VAE: N = 32*32 = 1024 tokens → extremely efficient
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from typing import Optional, Tuple
from einops import rearrange
# ============================================================================
# Core Building Block: Gated Selective State-Space Layer
# ============================================================================
class SelectiveStateSpace(nn.Module):
"""
Selective State Space layer with input-dependent parameters.
Mathematical formulation:
h_t = diag(exp(A_t)) * h_{t-1} + B_t * x_t (state transition)
y_t = C_t * h_t (output projection)
Where A_t, B_t, C_t are all computed from the input (selective/data-dependent).
This selectivity is what allows SSMs to match transformer quality.
Key insight: discretization of continuous dynamics means we can model
any timescale of dependencies by learning the step size Δ.
"""
def __init__(self, d_model: int, d_state: int = 16, d_conv: int = 4):
super().__init__()
self.d_model = d_model
self.d_state = d_state
self.d_conv = d_conv
# Input projections for selectivity
# We project to 2*d_model: one for the "gate" branch, one for the SSM branch
self.in_proj = nn.Linear(d_model, 2 * d_model, bias=False)
# Local convolution for capturing immediate neighbors (from Mamba)
self.conv1d = nn.Conv1d(
d_model, d_model, kernel_size=d_conv,
padding=d_conv - 1, groups=d_model, bias=True
)
# Selective parameters: ∆ (step size), B, C are input-dependent
# A is a learnable diagonal matrix (log-space for stability)
self.A_log = nn.Parameter(torch.log(torch.arange(1, d_state + 1, dtype=torch.float32).repeat(d_model, 1)))
self.D = nn.Parameter(torch.ones(d_model)) # Skip connection
# Input-dependent projections
self.dt_proj = nn.Linear(d_model, d_model, bias=True)
self.B_proj = nn.Linear(d_model, d_state, bias=False)
self.C_proj = nn.Linear(d_model, d_state, bias=False)
# Output projection
self.out_proj = nn.Linear(d_model, d_model, bias=False)
# Initialize dt bias to ensure positive step sizes
dt_init_std = d_model ** -0.5
nn.init.uniform_(self.dt_proj.bias, -4.0, -2.0) # Initialize in log space
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
x: (B, L, D) input sequence
Returns: (B, L, D) output sequence
"""
B, L, D = x.shape
# Split into gate and SSM branches
xz = self.in_proj(x) # (B, L, 2D)
x_ssm, z = xz.chunk(2, dim=-1) # Each (B, L, D)
# Local convolution (causal)
x_conv = x_ssm.transpose(1, 2) # (B, D, L)
x_conv = self.conv1d(x_conv)[:, :, :L] # Causal: trim to L
x_conv = x_conv.transpose(1, 2) # (B, L, D)
x_conv = F.silu(x_conv)
# Compute selective parameters
dt = F.softplus(self.dt_proj(x_conv)) # (B, L, D) - step sizes
B_sel = self.B_proj(x_conv) # (B, L, N)
C_sel = self.C_proj(x_conv) # (B, L, N)
# Discretize A
A = -torch.exp(self.A_log) # (D, N)
# Selective scan (vectorized for speed)
y = self._selective_scan(x_conv, dt, A, B_sel, C_sel) # (B, L, D)
# Skip connection
y = y + self.D.unsqueeze(0).unsqueeze(0) * x_conv
# Gating (from Mamba - SiLU gate)
y = y * F.silu(z)
return self.out_proj(y)
def _selective_scan(self, x, dt, A, B, C):
"""
Parallel selective scan using cumulative operations.
For training, we use the parallel form:
h_t = exp(A * dt_t) * h_{t-1} + dt_t * B_t * x_t
y_t = C_t * h_t
We compute this via log-space cumsum for numerical stability.
"""
B_batch, L, D = x.shape
N = A.shape[1]
# Compute discretized A and B
# dA = exp(A * dt): (B, L, D, N)
dt_expanded = dt.unsqueeze(-1) # (B, L, D, 1)
A_expanded = A.unsqueeze(0).unsqueeze(0) # (1, 1, D, N)
dA = torch.exp(dt_expanded * A_expanded) # (B, L, D, N)
# dB * x: (B, L, D, N)
dBx = dt_expanded * B.unsqueeze(2) * x.unsqueeze(-1) # (B, L, D, N)
# Sequential scan (we'll use a chunked approach for efficiency)
# For moderate sequence lengths (1024), direct scan is fast enough
h = torch.zeros(B_batch, D, N, device=x.device, dtype=x.dtype)
ys = []
# Use chunks of 64 for better memory efficiency
chunk_size = min(64, L)
for i in range(0, L, chunk_size):
end = min(i + chunk_size, L)
chunk_len = end - i
chunk_ys = []
for t in range(chunk_len):
idx = i + t
h = dA[:, idx] * h + dBx[:, idx] # (B, D, N)
y_t = (h * C[:, idx].unsqueeze(1)).sum(-1) # (B, D)
chunk_ys.append(y_t)
ys.extend(chunk_ys)
y = torch.stack(ys, dim=1) # (B, L, D)
return y
# ============================================================================
# Bidirectional Spatial Scanner
# ============================================================================
class BidirectionalSpatialScanner(nn.Module):
"""
Scans 2D spatial features in 4 directions to capture full spatial context.
Innovation: Instead of 4 separate SSMs (expensive), we use 2 SSMs with
input reversal, and fuse with a learned spatial gate.
Directions:
1. Row-major L→R (horizontal forward)
2. Row-major R→L (horizontal backward)
3. Col-major T→B (vertical forward)
4. Col-major B→T (vertical backward)
The gate learns to weight each direction based on spatial position and content.
"""
def __init__(self, d_model: int, d_state: int = 16):
super().__init__()
# Only 2 SSM instances - we reverse inputs for bidirectional
self.ssm_horizontal = SelectiveStateSpace(d_model, d_state)
self.ssm_vertical = SelectiveStateSpace(d_model, d_state)
# Spatial fusion gate - learns to weight directions
self.fusion_gate = nn.Sequential(
nn.Linear(d_model, d_model, bias=False),
nn.Sigmoid()
)
# Norm for stability
self.norm = nn.LayerNorm(d_model)
def forward(self, x: torch.Tensor, H: int, W: int) -> torch.Tensor:
"""
x: (B, H*W, D) flattened spatial features
Returns: (B, H*W, D) with full spatial context
"""
B, L, D = x.shape
# Horizontal scanning (row-major order)
x_fwd = self.ssm_horizontal(x)
x_bwd = self._reverse_scan(x, self.ssm_horizontal, H, W, reverse_dim='horizontal')
# Vertical scanning (column-major order)
x_col = rearrange(x, 'b (h w) d -> b (w h) d', h=H, w=W)
x_top_down = self.ssm_vertical(x_col)
x_top_down = rearrange(x_top_down, 'b (w h) d -> b (h w) d', h=H, w=W)
x_bot_up = self._reverse_scan(x_col, self.ssm_vertical, W, H, reverse_dim='vertical')
x_bot_up = rearrange(x_bot_up, 'b (w h) d -> b (h w) d', h=H, w=W)
# Learned fusion
combined = (x_fwd + x_bwd + x_top_down + x_bot_up) / 4.0
gate = self.fusion_gate(x)
out = gate * combined + (1 - gate) * x
return self.norm(out)
def _reverse_scan(self, x, ssm, H, W, reverse_dim):
"""Scan in reverse direction"""
x_rev = x.flip(dims=[1])
y_rev = ssm(x_rev)
return y_rev.flip(dims=[1])
# ============================================================================
# Mix-FFN with Depthwise Convolution (from SANA, proven effective)
# ============================================================================
class MixFFN(nn.Module):
"""
Feed-forward network with depthwise convolution for local feature mixing.
From SANA: "depth-wise convolution enhances the model's ability to capture
local information, compensating for the weaker local information-capturing
ability of linear attention"
Architecture: Linear → DWConv3x3 → GELU → Gate → Linear
This is an inverted bottleneck with gating.
"""
def __init__(self, d_model: int, expand_ratio: float = 2.5):
super().__init__()
d_inner = int(d_model * expand_ratio)
# Inverted bottleneck with gating
self.fc1 = nn.Linear(d_model, d_inner * 2) # *2 for gating
self.dwconv = nn.Conv2d(
d_inner, d_inner, kernel_size=3, padding=1,
groups=d_inner, bias=True
)
self.fc2 = nn.Linear(d_inner, d_model)
self.norm = nn.LayerNorm(d_inner)
def forward(self, x: torch.Tensor, H: int, W: int) -> torch.Tensor:
"""
x: (B, H*W, D)
Returns: (B, H*W, D)
"""
B, L, D = x.shape
# Split into value and gate
xg = self.fc1(x)
x_val, x_gate = xg.chunk(2, dim=-1) # Each (B, L, d_inner)
# Depthwise conv on value branch (needs 2D reshape)
x_val = rearrange(x_val, 'b (h w) d -> b d h w', h=H, w=W)
x_val = self.dwconv(x_val)
x_val = rearrange(x_val, 'b d h w -> b (h w) d')
# GLU gating
x_val = self.norm(x_val)
x_out = x_val * F.gelu(x_gate)
return self.fc2(x_out)
# ============================================================================
# Hyper-Connection Module (from the Hyper-Connections paper)
# ============================================================================
class HyperConnection(nn.Module):
"""
Hyper-connections generalize residual connections.
Instead of fixed: y = x + F(x)
We learn a connection matrix HC that can represent any blend of
sequential and parallel layer arrangements.
For expansion rate n:
Input: split x into n copies [x_1, ..., x_n]
HC matrix is (n+1) x (n+1), learnable
[input_to_layer, output_1, ..., output_n] = HC @ [F(input_to_layer), x_1, ..., x_n]
This subsumes both Pre-Norm and Post-Norm residual connections,
and can learn arrangements that are neither purely sequential nor parallel.
"""
def __init__(self, d_model: int, expansion_rate: int = 2):
super().__init__()
self.n = expansion_rate
self.d_model = d_model
# HC matrix: (n+1) x (n+1)
# Initialize close to residual connection
init_matrix = torch.zeros(self.n + 1, self.n + 1)
# Standard residual: input goes through, output adds
init_matrix[0, 1] = 1.0 # layer input comes from first stream
for i in range(1, self.n + 1):
init_matrix[i, i] = 1.0 # identity for skip
init_matrix[i, 0] = 1.0 / self.n # add layer output
self.hc_matrix = nn.Parameter(init_matrix)
self.norm = nn.LayerNorm(d_model)
def pre_forward(self, x_streams: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
"""
x_streams: (B, L, n*D) - n parallel streams concatenated
Returns: (layer_input, x_streams)
"""
B, L, _ = x_streams.shape
# Split into streams
streams = x_streams.chunk(self.n, dim=-1) # List of (B, L, D)
# Compute layer input from HC matrix first column
layer_input = sum(self.hc_matrix[0, i + 1] * streams[i] for i in range(self.n))
layer_input = self.norm(layer_input)
return layer_input, x_streams
def post_forward(self, layer_output: torch.Tensor, x_streams: torch.Tensor) -> torch.Tensor:
"""
Combine layer output with streams using HC matrix.
"""
streams = x_streams.chunk(self.n, dim=-1)
new_streams = []
for i in range(self.n):
new_stream = self.hc_matrix[i + 1, 0] * layer_output
for j in range(self.n):
new_stream = new_stream + self.hc_matrix[i + 1, j + 1] * streams[j]
new_streams.append(new_stream)
return torch.cat(new_streams, dim=-1)
def init_streams(self, x: torch.Tensor) -> torch.Tensor:
"""Initialize n streams from single input"""
return x.repeat(1, 1, self.n)
# ============================================================================
# AdaLN-Zero Conditioning (from DiT, proven optimal for diffusion)
# ============================================================================
class AdaLNZero(nn.Module):
"""
Adaptive Layer Normalization with zero initialization.
Conditions each layer on timestep and text embeddings.
From DiT: "regresses dimensionwise scale and shift parameters
from the sum of the embedding vectors"
Zero initialization ensures the network acts as identity at init,
critical for training stability.
"""
def __init__(self, d_model: int, d_cond: int):
super().__init__()
self.norm = nn.LayerNorm(d_model, elementwise_affine=False)
# Predict scale (γ), shift (β), and gate (α) - 6 values per element
self.proj = nn.Sequential(
nn.SiLU(),
nn.Linear(d_cond, 6 * d_model)
)
# Zero-initialize the projection
nn.init.zeros_(self.proj[1].weight)
nn.init.zeros_(self.proj[1].bias)
def forward(self, x: torch.Tensor, cond: torch.Tensor):
"""
x: (B, L, D)
cond: (B, d_cond)
Returns: shift1, scale1, gate1, shift2, scale2, gate2
"""
params = self.proj(cond) # (B, 6D)
params = params.unsqueeze(1) # (B, 1, 6D)
shift1, scale1, gate1, shift2, scale2, gate2 = params.chunk(6, dim=-1)
return shift1, scale1, gate1, shift2, scale2, gate2
def modulate(self, x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor):
return self.norm(x) * (1 + scale) + shift
# ============================================================================
# LiRA Block: The Core Processing Unit
# ============================================================================
class LiRABlock(nn.Module):
"""
One LiRA block = Bidirectional SSM + Mix-FFN, with:
- AdaLN-Zero conditioning
- Hyper-connections for dynamic layer arrangement
This replaces transformer blocks with O(N) complexity while maintaining
the quality of O(N^2) attention through:
1. Selective state spaces (content-aware)
2. Bidirectional scanning (full spatial context)
3. Mix-FFN (local feature enhancement via DWConv)
"""
def __init__(self, d_model: int, d_cond: int, d_state: int = 16,
ffn_expand: float = 2.5, hc_expansion: int = 2):
super().__init__()
# Conditioning
self.adaln = AdaLNZero(d_model, d_cond)
# Bidirectional State-Space Scanner
self.scanner = BidirectionalSpatialScanner(d_model, d_state)
# Mix-FFN for local features
self.ffn = MixFFN(d_model, ffn_expand)
# Layer norms (pre-norm style)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
def forward(self, x: torch.Tensor, cond: torch.Tensor, H: int, W: int) -> torch.Tensor:
"""
x: (B, H*W, D)
cond: (B, d_cond) - conditioning vector (timestep + text)
Returns: (B, H*W, D)
"""
# Get conditioning parameters
shift1, scale1, gate1, shift2, scale2, gate2 = self.adaln(x, cond)
# SSM branch with AdaLN conditioning
x_mod = self.adaln.modulate(x, shift1, scale1)
x_ssm = self.scanner(x_mod, H, W)
x = x + gate1 * x_ssm
# FFN branch with AdaLN conditioning
x_mod = self.adaln.modulate(x, shift2, scale2)
x_ffn = self.ffn(x_mod, H, W)
x = x + gate2 * x_ffn
return x
# ============================================================================
# Cross-Modal Fusion: Text → Image conditioning via Gated Cross-State
# ============================================================================
class GatedCrossStateFusion(nn.Module):
"""
Novel cross-modal fusion inspired by CrossWKV (from RWKV-7 paper).
Instead of expensive cross-attention (O(N*M) where N=image, M=text tokens),
we use a state-based cross-modal mechanism:
1. Compress text into a fixed-size state matrix S_text via SSM over text tokens
2. Inject S_text into image SSM states via gated addition
3. This gives O(M + N) complexity instead of O(N*M)
Mathematical formulation:
S_text = SSM_text(text_tokens) → (D, d_state) state matrix
For each image token x_i:
h_i = A_i * h_{i-1} + B_i * x_i + G_i * S_text * r_i
Where G_i is a learned gate and r_i is a receptance vector.
"""
def __init__(self, d_model: int, d_text: int, d_state: int = 16, num_heads: int = 8):
super().__init__()
self.d_model = d_model
self.d_state = d_state
self.num_heads = num_heads
self.head_dim = d_model // num_heads
# Text state compression
self.text_proj = nn.Linear(d_text, d_model)
self.text_key = nn.Linear(d_model, d_model, bias=False)
self.text_value = nn.Linear(d_model, d_model, bias=False)
# Image query
self.image_query = nn.Linear(d_model, d_model, bias=False)
# Gating mechanism
self.gate = nn.Sequential(
nn.Linear(d_model * 2, d_model),
nn.Sigmoid()
)
# Output projection
self.out_proj = nn.Linear(d_model, d_model, bias=False)
self.norm = nn.LayerNorm(d_model)
def forward(self, x_image: torch.Tensor, x_text: torch.Tensor) -> torch.Tensor:
"""
x_image: (B, N, D) - image features
x_text: (B, M, D_text) - text features
Returns: (B, N, D) - text-conditioned image features
"""
B, N, D = x_image.shape
# Project text to model dimension
text_feat = self.text_proj(x_text) # (B, M, D)
# Compute text summary using mean pooling + per-head KV
# This compresses all text into a single KV state per head
text_k = self.text_key(text_feat) # (B, M, D)
text_v = self.text_value(text_feat) # (B, M, D)
# Reshape to heads
text_k = rearrange(text_k, 'b m (h d) -> b h m d', h=self.num_heads)
text_v = rearrange(text_v, 'b m (h d) -> b h m d', h=self.num_heads)
# Compute text state: S = K^T V / M (compressed representation)
# This is O(M * d^2) which is very small for typical M (77 tokens)
text_state = torch.einsum('bhmd,bhmk->bhdk', text_k, text_v) / text_k.shape[2]
# Image queries
img_q = self.image_query(x_image) # (B, N, D)
img_q = rearrange(img_q, 'b n (h d) -> b h n d', h=self.num_heads)
# Query the text state: y = Q * S
cross_out = torch.einsum('bhnd,bhdk->bhnk', img_q, text_state)
cross_out = rearrange(cross_out, 'b h n d -> b n (h d)')
# Gated fusion
gate = self.gate(torch.cat([x_image, cross_out], dim=-1))
out = x_image + gate * cross_out
return self.norm(out)
# ============================================================================
# Latent Reasoning Loop (The Novel Core Innovation)
# ============================================================================
class LatentReasoningLoop(nn.Module):
"""
NOVEL CONTRIBUTION: Iterative reasoning in latent space for image generation.
Inspired by Liquid Reasoning Transformer (LRT), but adapted for generative models.
Key insight: Image generation benefits from iterative refinement. Instead of
a fixed number of denoising steps (expensive), we add a CHEAP inner reasoning
loop that refines the latent representation before final prediction.
How it works:
1. A "reasoning state" r_t evolves over T_think iterations
2. Each iteration applies a lightweight SSM + FFN to refine r_t
3. A DISCARD GATE filters bad updates (prevents error accumulation)
4. A STOP GATE halts early for easy inputs (adaptive compute)
5. The final r_T is used to condition the denoising prediction
This gives the model "thinking time" proportional to input difficulty:
- Simple prompts / high noise levels → few reasoning steps
- Complex prompts / fine detail refinement → more reasoning steps
Mathematical formulation:
r_0 = MLP(concat(z_t, c_text, t_embed))
For t in 1..T_max:
r_proposal = SSM_think(concat(z_tokens, r_t))
u_t = MLP(r_proposal) # candidate update
d_t = σ(W_d [r_{t-1}; u_t]) # discard gate
r_t = (1-d_t) * u_t + d_t * r_{t-1} # filtered update
s_t = σ(W_s r_t) # stop gate
if s_t > τ: break
Cost: T_think iterations of a SMALL network (1/10th of main backbone)
Typical T_think: 2-8 steps (learned, not fixed)
"""
def __init__(self, d_model: int, d_reason: int = 128, max_steps: int = 8):
super().__init__()
self.d_reason = d_reason
self.max_steps = max_steps
# Initialize reasoning state from input
self.state_init = nn.Sequential(
nn.Linear(d_model, d_reason * 2),
nn.GELU(),
nn.Linear(d_reason * 2, d_reason)
)
# Lightweight reasoning block (intentionally small)
self.reason_ssm = SelectiveStateSpace(d_reason, d_state=8, d_conv=3)
self.reason_ffn = nn.Sequential(
nn.Linear(d_reason, d_reason * 2),
nn.GELU(),
nn.Linear(d_reason * 2, d_reason)
)
self.reason_norm = nn.LayerNorm(d_reason)
# Discard gate: reject bad updates
self.discard_gate = nn.Sequential(
nn.Linear(d_reason * 2, d_reason),
nn.Sigmoid()
)
# Stop gate: halt when converged
self.stop_gate = nn.Sequential(
nn.Linear(d_reason, 1),
nn.Sigmoid()
)
self.stop_threshold = 0.8 # learnable threshold
# Project reasoning state back to condition the main network
self.reason_proj = nn.Linear(d_reason, d_model)
def forward(self, x: torch.Tensor, return_steps: bool = False) -> Tuple[torch.Tensor, dict]:
"""
x: (B, L, D) - input features (latent tokens + conditioning)
Returns: (B, D_model) reasoning conditioning vector, info dict
"""
B = x.shape[0]
# Initialize reasoning state from global average of input
x_global = x.mean(dim=1) # (B, D)
r = self.state_init(x_global) # (B, d_reason)
info = {'steps': [], 'discard_rates': [], 'stop_values': []}
# Iterative reasoning loop
total_steps = 0
for step in range(self.max_steps):
# Expand reasoning state and process with SSM
r_expanded = r.unsqueeze(1).expand(-1, x.shape[1], -1) # (B, L, d_reason)
# Lightweight processing
r_processed = self.reason_ssm(self.reason_norm(r_expanded))
r_proposal = self.reason_ffn(r_processed.mean(dim=1)) # (B, d_reason)
# Discard gate
d = self.discard_gate(torch.cat([r, r_proposal], dim=-1))
r_new = d * r + (1 - d) * r_proposal
# Stop gate
s = self.stop_gate(r_new).squeeze(-1) # (B,)
info['discard_rates'].append(d.mean().item())
info['stop_values'].append(s.mean().item())
r = r_new
total_steps += 1
# In inference, stop if all batch elements want to stop
if not self.training and (s > self.stop_threshold).all():
break
info['total_steps'] = total_steps
# Project to conditioning dimension
cond = self.reason_proj(r) # (B, D_model)
return cond, info
# ============================================================================
# Timestep + Text Embedding
# ============================================================================
class TimestepEmbedding(nn.Module):
"""
Sinusoidal timestep embedding with MLP projection.
Standard approach from DDPM, with the addition of frequency scaling
for better coverage of the continuous [0,1] range used in flow matching.
"""
def __init__(self, d_model: int, max_period: int = 10000):
super().__init__()
self.d_model = d_model
self.max_period = max_period
self.mlp = nn.Sequential(
nn.Linear(d_model, d_model * 4),
nn.SiLU(),
nn.Linear(d_model * 4, d_model)
)
def forward(self, t: torch.Tensor) -> torch.Tensor:
"""
t: (B,) timestep values in [0, 1]
Returns: (B, d_model)
"""
half_dim = self.d_model // 2
freqs = torch.exp(
-math.log(self.max_period) * torch.arange(half_dim, device=t.device).float() / half_dim
)
args = t.unsqueeze(1) * freqs.unsqueeze(0) * 1000 # Scale for better range
embedding = torch.cat([torch.sin(args), torch.cos(args)], dim=-1)
if self.d_model % 2:
embedding = F.pad(embedding, (0, 1))
return self.mlp(embedding)
class TextProjection(nn.Module):
"""
Projects text encoder outputs to model dimension.
Supports variable-length text with a pooled global + per-token output.
"""
def __init__(self, d_text: int, d_model: int):
super().__init__()
self.proj = nn.Linear(d_text, d_model)
self.pool_proj = nn.Linear(d_text, d_model)
self.norm = nn.LayerNorm(d_model)
def forward(self, text_features: torch.Tensor, text_mask: Optional[torch.Tensor] = None):
"""
text_features: (B, M, D_text)
text_mask: (B, M) boolean mask
Returns: per_token (B, M, D), pooled (B, D)
"""
per_token = self.norm(self.proj(text_features))
if text_mask is not None:
# Masked mean pooling
mask = text_mask.unsqueeze(-1).float()
pooled = (text_features * mask).sum(1) / mask.sum(1).clamp(min=1)
else:
pooled = text_features.mean(dim=1)
pooled = self.pool_proj(pooled)
return per_token, pooled