OCT_Segmentation_a1 / src /models /global_context.py
farrell236's picture
Deploy native Gradio OCT inference app
8fca640 verified
Raw History Blame Contribute Delete
2.17 kB
"""Local adjacent slices conditioned on three whole-volume OCT views."""
import torch
from torch import nn
from torch.nn import functional as F
from src.models.context import ContextBoundaryNet
from src.models.unet import Block
class GlobalBoundaryNet(ContextBoundaryNet):
def __init__(self, n_boundaries, base, prior_height):
super().__init__(n_boundaries, base, prior_height)
self.global_encoder = nn.Sequential(
nn.Conv2d(3, base, 5, stride=2, padding=2, bias=False),
nn.GroupNorm(min(8, base), base),
nn.SiLU(),
Block(base, base * 2),
nn.AvgPool2d(2),
Block(base * 2, base * 4),
nn.AdaptiveAvgPool2d(1),
nn.Flatten(),
)
self.global_film = nn.Linear(base * 4, base * 12 * 2)
nn.init.zeros_(self.global_film.weight)
nn.init.zeros_(self.global_film.bias)
def forward(self, image):
if image.shape[1] != 6:
raise ValueError("Expected local3 plus global3 OCT channels")
global_image = F.interpolate(
image[:, 3:], size=(96, 128), mode="bilinear", align_corners=False
)
embedding = self.global_encoder(global_image)
scale, bias = self.global_film(embedding).chunk(2, dim=1)
x = image[:, :3].contiguous()
skips = []
for index, layer in enumerate(self.down):
if index:
x = F.avg_pool2d(x, 2)
x = layer(x)
if index == len(self.down) - 1:
x = x * (1 + scale[:, :, None, None]) + bias[:, :, None, None]
skips.append(x)
for index, layer in enumerate(self.up):
skip = skips[-2 - index]
x = F.interpolate(
x, size=skip.shape[-2:], mode="bilinear", align_corners=False
)
x = layer(torch.cat([x, skip], 1))
logits = self.head(x)
prior = self.axial_prior
if prior.shape[-2] != logits.shape[-2]:
prior = F.interpolate(
prior, size=(logits.shape[-2], 1), mode="bilinear", align_corners=False
)
return logits + prior