File size: 11,897 Bytes
8966f37 1d82623 8966f37 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 | from copy import deepcopy
from abc import ABC, abstractmethod
import torch
import torch.nn.functional as F
from safetensors.torch import load_file
from torch import nn
from torchvision.transforms import v2
import wandb
class BaseUnet(nn.Module, ABC):
def __init__(
self,
in_channels: int,
num_classes: int,
n_organs: int,
size: int = 32,
depth: int = 3,
**kwargs
):
super().__init__()
self.in_channels = in_channels
self.out_channels = num_classes
self.n_organs = n_organs
self.size = size
self.depth = depth
self.distill = False
self.distill_unet = False
self.distill_model = None
self.distill_adapter = None
self.distill_loss = None
self._build_model(**kwargs)
@staticmethod
def _pad_to_2d(x: torch.Tensor, stride: int):
h, w = x.shape[-2:]
new_h = h if h % stride == 0 else h + stride - (h % stride)
new_w = w if w % stride == 0 else w + stride - (w % stride)
top = (new_h - h) // 2
bottom = (new_h - h) - top
left = (new_w - w) // 2
right = (new_w - w) - left
pads = (left, right, top, bottom)
x_pad = F.pad(x, pads, mode="constant", value=0)
return x_pad, pads
@staticmethod
def _unpad_2d(x: torch.Tensor, pads):
left, right, top, bottom = pads
if top or bottom:
end_h = -bottom if bottom > 0 else None
x = x[:, :, top:end_h, :]
if left or right:
end_w = -right if right > 0 else None
x = x[:, :, :, left:end_w]
return x
@abstractmethod
def _build_model(self, **kwargs):
pass
def _prepare_forward(
self,
*,
pixel_values: torch.Tensor,
organ_id: torch.Tensor | None = None,
**kwargs,
) -> dict:
return {}
@abstractmethod
def _encode(
self,
layer: nn.Module,
x: torch.Tensor,
organ_id: torch.Tensor | None = None,
forward_ctx: dict | None = None,
):
pass
@abstractmethod
def _bottleneck(
self,
x: torch.Tensor,
organ_id: torch.Tensor | None = None,
forward_ctx: dict | None = None,
):
pass
@abstractmethod
def _decode(
self,
layer: nn.Module,
x: torch.Tensor,
organ_id: torch.Tensor | None = None,
forward_ctx: dict | None = None,
):
pass
def encode(
self,
x: torch.Tensor,
organ_id: torch.Tensor | None = None,
forward_ctx: dict | None = None,
):
feat_list = []
pads = None
pre_padding = (
(x.size(-1) % 2**self.depth != 0)
or (x.size(-2) % 2**self.depth != 0)
or (x.size(-3) % 2**self.depth != 0)
)
if pre_padding:
x, pads = self._pad_to_2d(x, 2**self.depth)
out, feat = self._encode(
self.encoder["0"], x, organ_id=organ_id, forward_ctx=forward_ctx
)
feat_list.append(feat)
for key in list(self.encoder.keys())[1:]:
out, feat = self._encode(
self.encoder[key], out, organ_id=organ_id, forward_ctx=forward_ctx
)
feat_list.append(feat)
out = self._bottleneck(out, organ_id=organ_id, forward_ctx=forward_ctx)
return out, feat_list, pads
def decode(
self,
out: torch.Tensor,
feat_list: list[torch.Tensor],
pads,
organ_id: torch.Tensor | None = None,
forward_ctx: dict | None = None,
):
for key in self.decoder:
out = self._decode(
self.decoder[key],
torch.cat((out, feat_list[int(key)]), dim=1),
organ_id=organ_id,
forward_ctx=forward_ctx,
)
del feat_list[int(key)]
out = self.out_layer(out)
if pads is not None:
out = self._unpad_2d(out, pads).squeeze(1)
return out
def _apply_auxiliary_losses(
self,
*,
loss: torch.Tensor | float,
logits: torch.Tensor,
masks: torch.Tensor | None,
organ_id: torch.Tensor | None = None,
forward_ctx: dict | None = None,
**kwargs,
):
return loss
def _init_distillation(
self,
*,
distill: bool = False,
distill_unet: bool = False,
medsam_teacher_ckpt: str = "/work/phd_ultrasounds/UUSIC_new/checkpoints/medsam_unfreezed/model.safetensors",
unet_teacher_ckpt: str = "/work/phd_ultrasounds/UUSIC_new/checkpoints/unet5_attn_distilled/model.safetensors",
unet_teacher_kwargs: dict | None = None,
):
if distill and distill_unet:
raise ValueError("distill and distill_unet cannot both be enabled.")
self.distill = bool(distill)
self.distill_unet = bool(distill_unet)
self.distill_model = None
self.distill_adapter = None
self.distill_loss = None
if not self.distill and not self.distill_unet:
return
from .segm_net import DistillationLoss, MedSAM
student_channels = (2048 // (32 // self.size)) // (2 ** (5 - self.depth))
if self.distill:
self.distill_adapter = nn.Conv2d(student_channels, 256, kernel_size=1)
else:
self.distill_adapter = nn.Conv2d(student_channels, 2048, kernel_size=1)
if self.distill:
import importlib
sam_model_registry = importlib.import_module("segment_anything").sam_model_registry
from utils.paths import MEDSAM_BASE_WEIGHTS
sam_model = sam_model_registry["vit_b"](checkpoint=MEDSAM_BASE_WEIGHTS)
self.distill_model = MedSAM(
image_encoder=deepcopy(sam_model.image_encoder),
mask_decoder=deepcopy(sam_model.mask_decoder),
prompt_encoder=deepcopy(sam_model.prompt_encoder),
predict_bboxes=True,
freeze_image_encoder=0,
)
state_dict = load_file(medsam_teacher_ckpt)
load_result = self.distill_model.load_state_dict(state_dict)
print(f"Loaded MedSam teacher model and loaded weights:\n{load_result}")
else:
from .unet_attn import UNet2DAttn
teacher_kwargs = {
"in_channels": 3,
"num_classes": 1,
"n_organs": 10,
"size": 32,
"depth": 5,
"attn_start": 0,
"use_attn": True,
"img_size": 512,
"patch_size": 8,
"emb_dim": 768,
"n_heads": 8,
"distill": False,
"distill_unet": False,
"use_dwt": False,
"wavelet": "haar",
"use_shape": False,
"shape_res": 64,
}
if unet_teacher_kwargs is not None:
teacher_kwargs.update(unet_teacher_kwargs)
self.distill_model = UNet2DAttn(**teacher_kwargs)
state_dict = load_file(unet_teacher_ckpt)
state_dict = {k: v for k, v in state_dict.items() if "distill" not in k}
load_result = self.distill_model.load_state_dict(state_dict)
print(f"Loaded UNet teacher model and loaded weights:\n{load_result}")
for p in self.distill_model.parameters():
p.requires_grad = False
self.distill_model.eval()
self.distill_loss = DistillationLoss()
def _forward_distillation(
self,
*,
student_logits: torch.Tensor,
student_bottleneck: torch.Tensor,
pixel_values: torch.Tensor,
organ_id: torch.Tensor | None = None,
pixel_values_medsam: torch.Tensor | None = None,
) -> torch.Tensor | None:
if not self.distill and not self.distill_unet:
return None
if self.distill:
if pixel_values_medsam is None:
raise ValueError("pixel_values_medsam is required when distill=True.")
with torch.no_grad():
up_pixel_values = v2.functional.resize(
pixel_values_medsam, 1024, v2.InterpolationMode.BICUBIC
)
image_embedding = self.distill_model.image_encoder(up_pixel_values)
student_resized = F.interpolate(
student_bottleneck,
size=(image_embedding.shape[-1], image_embedding.shape[-1]),
mode="bilinear",
align_corners=False,
)
up_feat = self.distill_adapter(student_resized)
distill_loss_emb = self.distill_loss(
student_logits=up_feat,
teacher_logits=image_embedding.detach(),
)
if wandb.run is not None:
wandb.log(
{
"distill_loss_emb": distill_loss_emb["loss"].item(),
},
commit=False,
)
return distill_loss_emb["loss"]
with torch.no_grad():
forward_ctx = self._prepare_forward(
pixel_values=pixel_values,
organ_id=organ_id,
)
teacher_embedding, _, _ = self.encode(
pixel_values, organ_id=organ_id, forward_ctx=forward_ctx
)
student_resized = F.interpolate(
student_bottleneck,
size=(teacher_embedding.shape[-1], teacher_embedding.shape[-1]),
mode="bilinear",
align_corners=False,
)
up_feat = self.distill_adapter(student_resized)
distill_loss_emb = self.distill_loss(
student_logits=student_bottleneck,
teacher_logits=teacher_embedding.detach(),
)
if wandb.run is not None:
wandb.log(
{"distill_loss_logits": distill_loss_emb["loss"].item()},
commit=False,
)
return distill_loss_emb["loss"]
def forward(
self,
pixel_values,
organ_id=None,
labels=None,
masks=None,
bbox_coords=None,
organ_id_metric=None,
teacher_embedding=None,
teacher_mask=None,
pixel_values_medsam=None,
**kwargs,
):
forward_ctx = self._prepare_forward(
pixel_values=pixel_values,
organ_id=organ_id,
masks=masks,
**kwargs,
)
out_bottleneck, feat_list, pads = self.encode(
pixel_values, organ_id=organ_id, forward_ctx=forward_ctx
)
out = self.decode(
out_bottleneck,
feat_list,
pads,
organ_id=organ_id,
forward_ctx=forward_ctx,
)
if masks is not None:
loss = self.criterion(out, masks)
else:
loss = 0.0
distill_loss = self._forward_distillation(
student_logits=out,
student_bottleneck=out_bottleneck,
pixel_values=pixel_values,
organ_id=organ_id,
pixel_values_medsam=pixel_values_medsam,
)
if distill_loss is not None:
loss = loss + distill_loss
loss = self._apply_auxiliary_losses(
loss=loss,
logits=out,
masks=masks,
organ_id=organ_id,
forward_ctx=forward_ctx,
pixel_values=pixel_values,
**kwargs,
)
return {
"loss": loss,
"logits": out,
"labels": masks,
"organ_id": organ_id,
"organ_id_metric": organ_id_metric,
}
|