Spaces:
Running
Running
File size: 12,950 Bytes
ac59957 |
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 |
import argparse
import torch
import torch.optim as optim
import pytorch_lightning as pl
from torch.utils import data
import torch.nn.functional as F
from core import datasets
from core.memfof import MEMFOF
from core.datasets import fetch_dataloader
from core.utils.utils import load_ckpt
from core.loss import sequence_loss
class MEMFOFLit(pl.LightningModule):
"""PyTorch Lightning module for MEMFOF optical flow estimation.
This class implements the training and validation logic for the MEMFOF model
using PyTorch Lightning framework.
Parameters
----------
args : argparse.Namespace
Configuration parameters for the model and training process.
"""
def __init__(self, args: argparse.Namespace):
super().__init__()
self.args = args
self.model = MEMFOF(
backbone=self.args.pretrain,
dim=self.args.dim,
corr_radius=self.args.radius,
num_blocks=self.args.num_blocks,
use_var=self.args.use_var,
var_min=self.args.var_min,
var_max=self.args.var_max,
)
if self.args.restore_ckpt is not None:
load_ckpt(self, self.args.restore_ckpt)
print(f"restore ckpt from {self.args.restore_ckpt}")
self.log_kwargs = {"sync_dist": True, "add_dataloader_idx": False}
def training_step(
self, data_blob: tuple[torch.Tensor, torch.Tensor, torch.Tensor]
) -> torch.Tensor:
"""Perform a single training step.
Parameters
----------
data_blob : tuple[torch.Tensor, torch.Tensor, torch.Tensor]
Tuple containing (images, flow_gts, valids) tensors.
- images: Input image sequence of shape (B, 3, 3, H, W)
- flow_gts: Ground truth flow fields of shape (B, 2, 2, H, W)
- valids: Validity masks of shape (B, 2, H, W)
Returns
-------
torch.Tensor
Scalar loss value.
"""
images, flow_gts, valids = data_blob
outputs = self.model(images, flow_gts=flow_gts, iters=self.args.iters)
loss = sequence_loss(outputs, flow_gts, valids, self.args.gamma)
self.log("train-sequence-loss", loss, **self.log_kwargs)
return loss
def backward_flow(self, images: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
"""Calculate backward optical flow.
Parameters
----------
images : torch.Tensor
Input image sequence of shape (B, 3, 3, H, W)
Returns
-------
Tuple[torch.Tensor, torch.Tensor]
Tuple containing (flow, info) tensors.
- flow: Backward flow field of shape (B, 2, H, W)
- info: Additional information tensor of shape (B, 4, H, W)
"""
output = self.model(images, iters=self.args.iters)
flow_final = output["flow"][-1][:, 0]
info_final = output["info"][-1][:, 0]
return flow_final, info_final
def forward_flow(self, images: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
"""Calculate forward optical flow.
Parameters
----------
images : torch.Tensor
Input image sequence of shape (B, 3, 3, H, W)
Returns
-------
Tuple[torch.Tensor, torch.Tensor]
Tuple containing (flow, info) tensors.
- flow: Forward flow field of shape (B, 2, H, W)
- info: Additional information tensor of shape (B, 4, H, W)
"""
output = self.model(images, iters=self.args.iters)
flow_final = output["flow"][-1][:, 1]
info_final = output["info"][-1][:, 1]
return flow_final, info_final
def get_val_scale(self, data_name: str) -> int:
"""Get validation scale factor for different datasets.
Parameters
----------
data_name : str
Name of the validation dataset.
Returns
-------
int
Scale factor for validation. 0 means no scaling, 1 means 2x scaling, etc.
"""
return {"spring": 0}.get(data_name, 1)
def scale_and_forward_flow(
self, images: torch.Tensor, scale: float
) -> tuple[torch.Tensor, torch.Tensor]:
"""Calculate optical flow with specified scale.
Parameters
----------
images : torch.Tensor
Input image sequence of shape (B, 3, 3, H, W)
scale : float
Images will be scaled 2^scale times
Returns
-------
Tuple[torch.Tensor, torch.Tensor]
Tuple containing (flow, info) tensors.
- flow: Optical flow field of shape (B, 2, H, W)
- info: Additional information tensor of shape (B, 4, H, W)
"""
imgs = []
for i in range(3):
imgs.append(
F.interpolate(
images[:, i],
scale_factor=2**scale,
mode="bilinear",
align_corners=False,
)
)
imgs = torch.stack(imgs, dim=1)
flow, info = self.forward_flow(imgs)
flow_down = F.interpolate(
flow, scale_factor=0.5**scale, mode="bilinear", align_corners=False
) * (0.5**scale)
info_down = F.interpolate(info, scale_factor=0.5**scale, mode="area")
return flow_down, info_down
def validation_step_1(
self, data_blob: tuple[torch.Tensor, torch.Tensor, torch.Tensor], data_name: str
) -> None:
"""Validation step for chairs, sintel, and spring datasets.
Parameters
----------
data_blob : tuple[torch.Tensor, torch.Tensor, torch.Tensor]
Tuple containing (images, flow_gts, valids) tensors.
data_name : str
Name of the validation dataset.
"""
images, flow_gt, _ = data_blob
flow_gt = flow_gt.squeeze(dim=1)
flow, _ = self.scale_and_forward_flow(images, self.get_val_scale(data_name))
epe = torch.sum((flow - flow_gt) ** 2, dim=1).sqrt()
px1 = (epe < 1.0).float().mean(dim=[1, 2])
px3 = (epe < 3.0).float().mean(dim=[1, 2])
px5 = (epe < 5.0).float().mean(dim=[1, 2])
epe = epe.mean(dim=[1, 2])
self.log(
f"val-{data_name}-px1",
100 * (1 - px1).mean(),
**self.log_kwargs,
)
self.log(
f"val-{data_name}-px3",
100 * (1 - px3).mean(),
**self.log_kwargs,
)
self.log(
f"val-{data_name}-px5",
100 * (1 - px5).mean(),
**self.log_kwargs,
)
self.log(f"val-{data_name}-epe", epe.mean(), **self.log_kwargs)
def validation_step_2(
self, data_blob: tuple[torch.Tensor, torch.Tensor, torch.Tensor], data_name: str
) -> None:
"""Validation step for KITTI dataset.
Parameters
----------
data_blob : Tuple[torch.Tensor, torch.Tensor, torch.Tensor]
Tuple containing (images, flows_gt, valids_gt) tensors.
data_name : str
Name of the validation dataset.
"""
images, flow_gt, valid_gt = data_blob
flow_gt = flow_gt.squeeze(dim=1)
valid_gt = valid_gt.squeeze(dim=1)
flow, _ = self.scale_and_forward_flow(images, self.get_val_scale(data_name))
epe = torch.sum((flow - flow_gt) ** 2, dim=1).sqrt()
mag = torch.sum(flow_gt**2, dim=1).sqrt()
val = valid_gt >= 0.5
out = ((epe > 3.0) & ((epe / mag) > 0.05)).float()
epe_list = []
out_valid_pixels = 0
num_valid_pixels = 0
for b in range(out.shape[0]):
epe_list.append(epe[b][val[b]].mean())
out_valid_pixels += out[b][val[b]].sum()
num_valid_pixels += val[b].sum()
epe = torch.mean(torch.tensor(epe_list, device=self.device))
f1 = 100 * out_valid_pixels / num_valid_pixels
self.log(f"val-{data_name}-epe", epe, **self.log_kwargs)
self.log(f"val-{data_name}-f1", f1, **self.log_kwargs)
def validation_step(
self,
data_blob: tuple[torch.Tensor, torch.Tensor, torch.Tensor],
batch_idx: int,
dataloader_idx: int = 0,
) -> None:
"""Main validation step that routes to specific validation methods.
Parameters
----------
data_blob : Tuple[torch.Tensor, torch.Tensor, torch.Tensor]
Tuple containing validation data tensors.
batch_idx : int
Index of the current batch.
dataloader_idx : int, optional
Index of the current dataloader, by default 0
"""
if not self.args.val_datasets:
return
data_name = self.args.val_datasets[dataloader_idx]
if data_name in (
"chairs",
"sintel",
"sintel-clean",
"sintel-final",
"spring",
"spring-1080",
):
self.validation_step_1(data_blob, data_name)
elif data_name in ("kitti",):
self.validation_step_2(data_blob, data_name)
def configure_optimizers(self) -> dict:
"""Configure optimizers and learning rate schedulers.
Returns
-------
Dict[str, Any]
Dictionary containing optimizer and scheduler configurations.
"""
optimizer = optim.AdamW(
self.model.parameters(),
lr=self.args.lr,
weight_decay=self.args.wdecay,
eps=self.args.epsilon,
)
scheduler = optim.lr_scheduler.OneCycleLR(
optimizer,
self.args.lr,
self.args.num_steps + 100,
pct_start=0.05,
cycle_momentum=False,
anneal_strategy="linear",
)
lr_scheduler_dict = {"scheduler": scheduler, "interval": "step"}
return {"optimizer": optimizer, "lr_scheduler": lr_scheduler_dict}
class DataModule(pl.LightningDataModule):
"""PyTorch Lightning DataModule for MEMFOF training and validation.
Parameters
----------
args : argparse.Namespace
Configuration parameters for data loading.
"""
def __init__(self, args: argparse.Namespace) -> None:
super().__init__()
self.args = args
def train_dataloader(self) -> data.DataLoader:
"""Get training dataloader.
Returns
-------
data.DataLoader
Training dataloader instance.
"""
return fetch_dataloader(self.args)
def val_dataloader(self) -> list[data.DataLoader]:
"""Get validation dataloaders for different datasets.
Returns
-------
List[data.DataLoader]
List of validation dataloaders.
"""
kwargs = {
"pin_memory": False,
"shuffle": False,
"num_workers": self.args.num_workers,
"drop_last": False,
}
val_dataloaders = []
for val_dataset in self.args.val_datasets:
if val_dataset == "sintel":
clean = datasets.three_frame_wrapper_val(
datasets.MpiSintel, {"split": "val", "dstype": "clean"}
)
final = datasets.three_frame_wrapper_val(
datasets.MpiSintel, {"split": "val", "dstype": "final"}
)
loader = data.DataLoader(clean + final, batch_size=8, **kwargs)
elif val_dataset == "sintel-clean":
clean = datasets.three_frame_wrapper_val(
datasets.MpiSintel, {"split": "val", "dstype": "clean"}
)
loader = data.DataLoader(clean, batch_size=8, **kwargs)
elif val_dataset == "sintel-final":
final = datasets.three_frame_wrapper_val(
datasets.MpiSintel, {"split": "val", "dstype": "final"}
)
loader = data.DataLoader(final, batch_size=8, **kwargs)
elif val_dataset == "kitti":
kitti = datasets.three_frame_wrapper_val(
datasets.KITTI, {"split": "val"}
)
loader = data.DataLoader(kitti, batch_size=1, **kwargs)
elif val_dataset == "spring":
spring = datasets.three_frame_wrapper_val(
datasets.SpringFlowDataset, {"split": "val"}
)
loader = data.DataLoader(spring, batch_size=4, **kwargs)
elif val_dataset == "spring-1080":
spring = datasets.three_frame_wrapper_val(
datasets.SpringFlowDataset, {"split": "val"}
)
loader = data.DataLoader(spring, batch_size=4, **kwargs)
else:
raise ValueError(f"Unknown validation dataset: {val_dataset}")
val_dataloaders.append(loader)
return val_dataloaders
|