File size: 11,103 Bytes
ab687e7 |
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 |
from argparse import ArgumentParser, Namespace
import multiprocessing
import torch
from torch import nn
import torch.nn.functional as F
from torch.utils.data import DataLoader
import torchvision.transforms as transforms
from lightning.pytorch import LightningModule, Trainer, cli_lightning_logo
from lightning.pytorch.callbacks import EarlyStopping, ModelCheckpoint
from lightning.pytorch.loggers import CSVLogger
from pytorch_caney.datasets.modis_dataset import MODISDataset
from pytorch_caney.utils import check_gpus_available
class UNet(nn.Module):
"""
Architecture based on U-Net: Convolutional Networks for
Biomedical Image Segmentation.
Link - https://arxiv.org/abs/1505.04597
>>> UNet(num_classes=2, num_layers=3) \
# doctest: +ELLIPSIS +NORMALIZE_WHITESPACE
UNet(
(layers): ModuleList(
(0): DoubleConv(...)
(1): Down(...)
(2): Down(...)
(3): Up(...)
(4): Up(...)
(5): Conv2d(64, 2, kernel_size=(1, 1), stride=(1, 1))
)
)
"""
def __init__(
self,
num_channels: int = 7,
num_classes: int = 19,
num_layers: int = 5,
features_start: int = 64,
bilinear: bool = False
):
super().__init__()
self.num_layers = num_layers
layers = [DoubleConv(num_channels, features_start)]
feats = features_start
for _ in range(num_layers - 1):
layers.append(Down(feats, feats * 2))
feats *= 2
for _ in range(num_layers - 1):
layers.append(Up(feats, feats // 2, bilinear))
feats //= 2
layers.append(nn.Conv2d(feats, num_classes, kernel_size=1))
self.layers = nn.ModuleList(layers)
def forward(self, x):
xi = [self.layers[0](x)]
# Down path
for layer in self.layers[1: self.num_layers]:
xi.append(layer(xi[-1]))
# Up path
for i, layer in enumerate(self.layers[self.num_layers: -1]):
xi[-1] = layer(xi[-1], xi[-2 - i])
return self.layers[-1](xi[-1])
class DoubleConv(nn.Module):
"""Double Convolution and BN and ReLU (3x3 conv -> BN -> ReLU) ** 2.
>>> DoubleConv(4, 4) \
# doctest: +ELLIPSIS +NORMALIZE_WHITESPACE
DoubleConv(
(net): Sequential(...)
)
"""
def __init__(self, in_ch: int, out_ch: int):
super().__init__()
self.net = nn.Sequential(
nn.Conv2d(in_ch, out_ch, kernel_size=3, padding=1),
nn.BatchNorm2d(out_ch),
nn.ReLU(inplace=True),
nn.Conv2d(out_ch, out_ch, kernel_size=3, padding=1),
nn.BatchNorm2d(out_ch),
nn.ReLU(inplace=True),
)
def forward(self, x):
return self.net(x)
class Down(nn.Module):
"""Combination of MaxPool2d and DoubleConv in series.
>>> Down(4, 8) # doctest: +ELLIPSIS +NORMALIZE_WHITESPACE
Down(
(net): Sequential(
(0): MaxPool2d(
kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(1): DoubleConv(
(net): Sequential(...)
)
)
)
"""
def __init__(self, in_ch: int, out_ch: int):
super().__init__()
self.net = nn.Sequential(
nn.MaxPool2d(kernel_size=2, stride=2), DoubleConv(in_ch, out_ch))
def forward(self, x):
return self.net(x)
class Up(nn.Module):
"""Upsampling (by either bilinear interpolation or transpose convolutions)
followed by concatenation of feature
map from contracting path, followed by double 3x3 convolution.
>>> Up(8, 4) # doctest: +ELLIPSIS +NORMALIZE_WHITESPACE
Up(
(upsample): ConvTranspose2d(8, 4, kernel_size=(2, 2), stride=(2, 2))
(conv): DoubleConv(
(net): Sequential(...)
)
)
"""
def __init__(self, in_ch: int, out_ch: int, bilinear: bool = False):
super().__init__()
self.upsample = None
if bilinear:
self.upsample = nn.Sequential(
nn.Upsample(
scale_factor=2, mode="bilinear", align_corners=True),
nn.Conv2d(
in_ch, in_ch // 2, kernel_size=1),
)
else:
self.upsample = nn.ConvTranspose2d(
in_ch, in_ch // 2, kernel_size=2, stride=2)
self.conv = DoubleConv(in_ch, out_ch)
def forward(self, x1, x2):
x1 = self.upsample(x1)
# Pad x1 to the size of x2
diff_h = x2.shape[2] - x1.shape[2]
diff_w = x2.shape[3] - x1.shape[3]
x1 = F.pad(
x1,
[
diff_w // 2, diff_w - diff_w // 2,
diff_h // 2, diff_h - diff_h // 2
])
# Concatenate along the channels axis
x = torch.cat([x2, x1], dim=1)
return self.conv(x)
class SegmentationModel(LightningModule):
def __init__(
self,
data_path: list = [],
n_classes: int = 18,
batch_size: int = 256,
lr: float = 3e-4,
num_layers: int = 5,
features_start: int = 64,
bilinear: bool = False,
**kwargs,
):
super().__init__(**kwargs)
self.data_paths = data_path
self.n_classes = n_classes
self.batch_size = batch_size
self.learning_rate = lr
self.num_layers = num_layers
self.features_start = features_start
self.bilinear = bilinear
self.validation_step_outputs = []
self.net = UNet(
num_classes=self.n_classes,
num_layers=self.num_layers,
features_start=self.features_start,
bilinear=self.bilinear
)
self.transform = transforms.Compose(
[
transforms.ToTensor(),
transforms.Normalize(
mean=[0.0173, 0.0332, 0.0088,
0.0136, 0.0381, 0.0348, 0.0249],
std=[0.0150, 0.0127, 0.0124,
0.0128, 0.0120, 0.0159, 0.0164]
),
]
)
print('> Init datasets')
self.trainset = MODISDataset(
self.data_paths, split="train", transform=self.transform)
self.validset = MODISDataset(
self.data_paths, split="valid", transform=self.transform)
print('Done init datasets')
def forward(self, x):
return self.net(x)
def training_step(self, batch, batch_nb):
img, mask = batch
img = img.float()
mask = mask.long()
out = self(img)
loss = F.cross_entropy(out, mask, ignore_index=250)
log_dict = {"train_loss": loss}
self.log_dict(log_dict)
return {"loss": loss, "log": log_dict, "progress_bar": log_dict}
def validation_step(self, batch, batch_idx):
img, mask = batch
img = img.float()
mask = mask.long()
out = self(img)
loss_val = F.cross_entropy(out, mask, ignore_index=250)
self.validation_step_outputs.append(loss_val)
return {"val_loss": loss_val}
def on_validation_epoch_end(self):
loss_val = torch.stack(self.validation_step_outputs).mean()
log_dict = {"val_loss": loss_val}
self.log("val_loss", loss_val, sync_dist=True)
self.validation_step_outputs.clear()
return {
"log": log_dict,
"val_loss": log_dict["val_loss"],
"progress_bar": log_dict
}
def configure_optimizers(self):
opt = torch.optim.Adam(self.net.parameters(), lr=self.learning_rate)
# sch = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=10)
return [opt] # , [sch]
def train_dataloader(self):
return DataLoader(
self.trainset,
batch_size=self.batch_size,
num_workers=multiprocessing.cpu_count(),
shuffle=True
)
def val_dataloader(self):
return DataLoader(
self.validset,
batch_size=self.batch_size,
num_workers=multiprocessing.cpu_count(),
shuffle=False
)
def main(hparams: Namespace):
# ------------------------
# 1 INIT LIGHTNING MODEL
# ------------------------
ngpus = int(hparams.ngpus)
# PT ligtning does not expect this, del after use
del hparams.ngpus
model = SegmentationModel(**vars(hparams))
# ------------------------
# 2 SET LOGGER
# ------------------------
# logger = True
# if hparams.log_wandb:
# logger = WandbLogger()
# # optional: log model topology
# logger.watch(model.net)
train_callbacks = [
# TQDMProgressBar(refresh_rate=20),
ModelCheckpoint(dirpath='models/',
monitor='val_loss',
save_top_k=5,
filename='{epoch}-{val_loss:.2f}.ckpt'),
EarlyStopping("val_loss", patience=10, mode='min'),
]
# See number of devices
check_gpus_available(ngpus)
# ------------------------
# 3 INIT TRAINER
# ------------------------
# trainer = Trainer(
# ------------------------
trainer = Trainer(
accelerator="gpu",
devices=ngpus,
strategy="ddp",
min_epochs=1,
max_epochs=500,
callbacks=train_callbacks,
logger=CSVLogger(save_dir="logs/"),
# precision=16 # makes loss nan, need to fix that
)
# ------------------------
# 5 START TRAINING
# ------------------------
trainer.fit(model)
trainer.save_checkpoint("best_model.ckpt")
# ------------------------
# 6 START TEST
# ------------------------
# test_set = MODISDataset(
# self.data_path, split=None, transform=self.transform)
# test_dataloader = DataLoader(...)
# trainer.test(ckpt_path="best", dataloaders=)
if __name__ == "__main__":
cli_lightning_logo()
parser = ArgumentParser()
parser.add_argument(
"--data_path", nargs='+', required=True,
help="path where dataset is stored")
parser.add_argument('--ngpus', type=int,
default=torch.cuda.device_count(),
help='number of gpus to use')
parser.add_argument(
"--n-classes", type=int, default=18, help="number of classes")
parser.add_argument(
"--batch_size", type=int, default=256, help="size of the batches")
parser.add_argument(
"--lr", type=float, default=3e-4, help="adam: learning rate")
parser.add_argument(
"--num_layers", type=int, default=5, help="number of layers on u-net")
parser.add_argument(
"--features_start", type=float, default=64,
help="number of features in first layer")
parser.add_argument(
"--bilinear", action="store_true", default=False,
help="whether to use bilinear interpolation or transposed")
# parser.add_argument(
# "--log-wandb", action="store_true", default=True,
# help="whether to use wandb as the logger")
hparams = parser.parse_args()
main(hparams)
|