File size: 9,818 Bytes
57746f1 |
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 |
import types
import time
import random
import clip
import torch
import torch.nn as nn
import torchvision.transforms as transforms
from argparse import ArgumentParser
import pytorch_lightning as pl
from data import get_dataset, get_available_datasets
from encoding.models import get_segmentation_model
from encoding.nn import SegmentationLosses
from encoding.utils import batch_pix_accuracy, batch_intersection_union
# add mixed precision
import torch.cuda.amp as amp
import numpy as np
from encoding.utils import SegmentationMetric
class LSegmentationModule(pl.LightningModule):
def __init__(self, data_path, dataset, batch_size, base_lr, max_epochs, **kwargs):
super().__init__()
self.data_path = data_path
self.batch_size = batch_size
self.base_lr = base_lr / 16 * batch_size
self.lr = self.base_lr
self.epochs = max_epochs
self.other_kwargs = kwargs
self.enabled = False #True mixed precision will make things complicated and leading to NAN error
self.scaler = amp.GradScaler(enabled=self.enabled)
def forward(self, x):
return self.net(x)
def evaluate(self, x, target=None):
pred = self.net.forward(x)
if isinstance(pred, (tuple, list)):
pred = pred[0]
if target is None:
return pred
correct, labeled = batch_pix_accuracy(pred.data, target.data)
inter, union = batch_intersection_union(pred.data, target.data, self.nclass)
return correct, labeled, inter, union
def evaluate_random(self, x, labelset, target=None):
pred = self.net.forward(x, labelset)
if isinstance(pred, (tuple, list)):
pred = pred[0]
if target is None:
return pred
correct, labeled = batch_pix_accuracy(pred.data, target.data)
inter, union = batch_intersection_union(pred.data, target.data, self.nclass)
return correct, labeled, inter, union
def training_step(self, batch, batch_nb):
img, target = batch
with amp.autocast(enabled=self.enabled):
out = self(img)
multi_loss = isinstance(out, tuple)
if multi_loss:
loss = self.criterion(*out, target)
else:
loss = self.criterion(out, target)
loss = self.scaler.scale(loss)
final_output = out[0] if multi_loss else out
train_pred, train_gt = self._filter_invalid(final_output, target)
if train_gt.nelement() != 0:
self.train_accuracy(train_pred, train_gt)
self.log("train_loss", loss)
return loss
def training_epoch_end(self, outs):
self.log("train_acc_epoch", self.train_accuracy.compute())
def validation_step(self, batch, batch_nb):
img, target = batch
out = self(img)
multi_loss = isinstance(out, tuple)
if multi_loss:
val_loss = self.criterion(*out, target)
else:
val_loss = self.criterion(out, target)
final_output = out[0] if multi_loss else out
valid_pred, valid_gt = self._filter_invalid(final_output, target)
self.val_iou.update(target, final_output)
pixAcc, iou = self.val_iou.get()
self.log("val_loss_step", val_loss)
self.log("pix_acc_step", pixAcc)
self.log(
"val_acc_step",
self.val_accuracy(valid_pred, valid_gt),
)
self.log("val_iou", iou)
def validation_epoch_end(self, outs):
pixAcc, iou = self.val_iou.get()
self.log("val_acc_epoch", self.val_accuracy.compute())
self.log("val_iou_epoch", iou)
self.log("pix_acc_epoch", pixAcc)
self.val_iou.reset()
def _filter_invalid(self, pred, target):
valid = target != self.other_kwargs["ignore_index"]
_, mx = torch.max(pred, dim=1)
return mx[valid], target[valid]
def configure_optimizers(self):
params_list = [
{"params": self.net.pretrained.parameters(), "lr": self.base_lr},
]
if hasattr(self.net, "scratch"):
print("Found output scratch")
params_list.append(
{"params": self.net.scratch.parameters(), "lr": self.base_lr * 10}
)
if hasattr(self.net, "auxlayer"):
print("Found auxlayer")
params_list.append(
{"params": self.net.auxlayer.parameters(), "lr": self.base_lr * 10}
)
if hasattr(self.net, "scale_inv_conv"):
print(self.net.scale_inv_conv)
print("Found scaleinv layers")
params_list.append(
{
"params": self.net.scale_inv_conv.parameters(),
"lr": self.base_lr * 10,
}
)
params_list.append(
{"params": self.net.scale2_conv.parameters(), "lr": self.base_lr * 10}
)
params_list.append(
{"params": self.net.scale3_conv.parameters(), "lr": self.base_lr * 10}
)
params_list.append(
{"params": self.net.scale4_conv.parameters(), "lr": self.base_lr * 10}
)
if self.other_kwargs["midasproto"]:
print("Using midas optimization protocol")
opt = torch.optim.Adam(
params_list,
lr=self.base_lr,
betas=(0.9, 0.999),
weight_decay=self.other_kwargs["weight_decay"],
)
sch = torch.optim.lr_scheduler.LambdaLR(
opt, lambda x: pow(1.0 - x / self.epochs, 0.9)
)
else:
opt = torch.optim.SGD(
params_list,
lr=self.base_lr,
momentum=0.9,
weight_decay=self.other_kwargs["weight_decay"],
)
sch = torch.optim.lr_scheduler.LambdaLR(
opt, lambda x: pow(1.0 - x / self.epochs, 0.9)
)
return [opt], [sch]
def train_dataloader(self):
return torch.utils.data.DataLoader(
self.trainset,
batch_size=self.batch_size,
shuffle=True,
num_workers=16,
worker_init_fn=lambda x: random.seed(time.time() + x),
)
def val_dataloader(self):
return torch.utils.data.DataLoader(
self.valset,
batch_size=self.batch_size,
shuffle=False,
num_workers=16,
)
def get_trainset(self, dset, augment=False, **kwargs):
print(kwargs)
if augment == True:
mode = "train_x"
else:
mode = "train"
print(mode)
dset = get_dataset(
dset,
root=self.data_path,
split="train",
mode=mode,
transform=self.train_transform,
**kwargs
)
self.num_classes = dset.num_class
self.train_accuracy = pl.metrics.Accuracy()
return dset
def get_valset(self, dset, augment=False, **kwargs):
self.val_accuracy = pl.metrics.Accuracy()
self.val_iou = SegmentationMetric(self.num_classes)
if augment == True:
mode = "val_x"
else:
mode = "val"
print(mode)
return get_dataset(
dset,
root=self.data_path,
split="val",
mode=mode,
transform=self.val_transform,
**kwargs
)
def get_criterion(self, **kwargs):
return SegmentationLosses(
se_loss=kwargs["se_loss"],
aux=kwargs["aux"],
nclass=self.num_classes,
se_weight=kwargs["se_weight"],
aux_weight=kwargs["aux_weight"],
ignore_index=kwargs["ignore_index"],
)
@staticmethod
def add_model_specific_args(parent_parser):
parser = ArgumentParser(parents=[parent_parser], add_help=False)
parser.add_argument(
"--data_path", type=str, help="path where dataset is stored"
)
parser.add_argument(
"--dataset",
choices=get_available_datasets(),
default="ade20k",
help="dataset to train on",
)
parser.add_argument(
"--batch_size", type=int, default=16, help="size of the batches"
)
parser.add_argument(
"--base_lr", type=float, default=0.004, help="learning rate"
)
parser.add_argument("--momentum", type=float, default=0.9, help="SGD momentum")
parser.add_argument(
"--weight_decay", type=float, default=1e-4, help="weight_decay"
)
parser.add_argument(
"--aux", action="store_true", default=False, help="Auxilary Loss"
)
parser.add_argument(
"--aux-weight",
type=float,
default=0.2,
help="Auxilary loss weight (default: 0.2)",
)
parser.add_argument(
"--se-loss",
action="store_true",
default=False,
help="Semantic Encoding Loss SE-loss",
)
parser.add_argument(
"--se-weight", type=float, default=0.2, help="SE-loss weight (default: 0.2)"
)
parser.add_argument(
"--midasproto", action="store_true", default=False, help="midasprotocol"
)
parser.add_argument(
"--ignore_index",
type=int,
default=-1,
help="numeric value of ignore label in gt",
)
parser.add_argument(
"--augment",
action="store_true",
default=False,
help="Use extended augmentations",
)
return parser
|