D-SENO / model /DSENO.py
yushuang88's picture
Upload folder using huggingface_hub
b144968 verified
Raw History Blame Contribute Delete
3.76 kB
"""Pipe-G architecture from arXiv:2601.17407, independently implemented.
The paper defines two physical coordinate inputs, width 96, seven residual
DS blocks and the dilation order. Paper omissions (DC intermediate GELU,
zero padding and projection width 128) follow the separately audited
read-only official reference; no repository code is copied here.
"""
from __future__ import annotations
from collections.abc import Sequence
import torch
from torch import nn
class SqueezeExcitation(nn.Module):
"""Spatial mean and two pointwise convolutions produce channel gates."""
def __init__(self, channels: int, reduction: int = 1):
super().__init__()
if reduction < 1 or channels % reduction:
raise ValueError("SE reduction must divide the channel width")
hidden = channels // reduction
self.reduce = nn.Conv2d(channels, hidden, 1, bias=True)
self.activate = nn.GELU()
self.expand = nn.Conv2d(hidden, channels, 1, bias=True)
def forward(self, features: torch.Tensor) -> torch.Tensor:
pooled = features.mean(dim=(-2, -1), keepdim=True)
gates = torch.sigmoid(self.expand(self.activate(self.reduce(pooled))))
return features * gates
class DSBlock(nn.Module):
"""Two dilated convolutions, SE gating, residual addition and GELU."""
def __init__(self, channels: int, dilation: int, se_reduction: int = 1):
super().__init__()
if dilation < 1:
raise ValueError("Dilation must be positive")
self.first = nn.Conv2d(channels, channels, 3, padding=dilation,
dilation=dilation, bias=True)
self.second = nn.Conv2d(channels, channels, 3, padding=dilation,
dilation=dilation, bias=True)
self.se = SqueezeExcitation(channels, se_reduction)
self.activate = nn.GELU()
def forward(self, features: torch.Tensor) -> torch.Tensor:
transformed = self.second(self.activate(self.first(features)))
return self.activate(features + self.se(transformed))
class DSENO(nn.Module):
"""Normalized coordinates [B,H,W,2] to normalized velocity [B,H,W]."""
def __init__(self, width: int = 96,
dilations: Sequence[int] = (23, 17, 13, 9, 7, 3, 1),
projection_width: int = 128, input_channels: int = 2,
output_channels: int = 1, se_reduction: int = 1):
super().__init__()
if input_channels != 2 or output_channels != 1:
raise ValueError("Pipe requires exactly X/Y inputs and one velocity output")
if width < 1 or projection_width < 1 or not dilations:
raise ValueError("Widths and dilation sequence must be nonempty and positive")
self.input_channels = input_channels
self.lifting = nn.Conv2d(input_channels, width, 1, bias=True)
self.blocks = nn.Sequential(*(DSBlock(width, int(d), se_reduction)
for d in dilations))
self.projection = nn.Sequential(nn.Conv2d(width, projection_width, 1, bias=True),
nn.GELU(),
nn.Conv2d(projection_width, output_channels, 1,
bias=True))
def forward(self, coordinates: torch.Tensor) -> torch.Tensor:
if coordinates.ndim != 4 or coordinates.shape[-1] != self.input_channels:
raise ValueError("DSENO expects coordinates shaped [B,H,W,2]")
features = coordinates.permute(0, 3, 1, 2).contiguous(memory_format=getattr(self, "memory_format", torch.channels_last))
features = self.blocks(self.lifting(features))
return self.projection(features).squeeze(1)