def check_schedule(self, path, gt_scheduler, gt_lr):
logger = TrainingManager(path)
w = Variable(torch.zeros((1, )))
optimizer = Adam([w])
lr_scheduler = logger.lr_scheduler(optimizer)
for epoch, (gold_class, gold_lr) in enumerate(zip(gt_scheduler,
gt_lr)):
self.assertEqual(lr_scheduler.current_scheduler.__class__,
gold_class)
self.assertEqual(lr_scheduler.get_lr(), [gold_lr])
lr_scheduler.step()
def check_schedule_groups(self, path, gt_scheduler, gt_lr):
logger = TrainingManager(path)
w1 = Variable(torch.zeros((1, )))
w2 = Variable(torch.zeros((1, )))
optimizer = Adam([{
"params": [w1],
"lr": 0.0
}, {
"params": [w2],
"lr": 0.0
}])
lr_scheduler = logger.lr_scheduler(optimizer)
for epoch, (gold_class, gold_lr) in enumerate(zip(gt_scheduler,
gt_lr)):
print(epoch)
self.assertEqual(lr_scheduler.current_scheduler.__class__,
gold_class)
self.assertEqual(lr_scheduler.get_lr(), gold_lr)
lr_scheduler.step()
def check_optimizer(self, path):
logger = TrainingManager(path)
model = logger.model()
params = logger.parameters(model)
optimizer = logger.optimizer(params)
scheduler = logger.lr_scheduler(optimizer)
return optimizer
from manager import TrainingManager
from torch.utils.tensorboard import SummaryWriter
from albumentations import *
import cv2
import json
from pycocotools.cocoeval import COCOeval
try:
from apex import amp
APEX = True
except ModuleNotFoundError:
APEX = False
if __name__ == '__main__':
opt = opts().parse()
logger = TrainingManager(opt.save_dir)
history = History(opt.save_dir, opt.resume)
writer = SummaryWriter()
torch.backends.cudnn.benchmark = True
print(opt)
transforms = {
"train":
Compose(
[
ShiftScaleRotate(rotate_limit=90,
scale_limit=(-0.35, 0.3),
border_mode=cv2.BORDER_CONSTANT),
PadIfNeeded(min_height=512,
min_width=512,
border_mode=cv2.BORDER_CONSTANT,
always_apply=True),
import os
import torch.utils.data
from albumentations import *
from torch import device
from history import History
from models import load_model, save_model
from opts import opts
from trainers.segmentation_trainer import SegmentationTrainer as Trainer
from manager import TrainingManager
if __name__ == '__main__':
opt = opts().parse()
logger = TrainingManager(opt.save_dir)
history = History(opt.save_dir, opt.resume)
# writer = SummaryWriter()
torch.backends.cudnn.benchmark = True
print(opt)
transforms = {
"train": Compose([
ToGray(),
HorizontalFlip(),
OneOf([
Compose([
OneOf([
ElasticTransform(alpha=200, sigma=35, p=0.5, border_mode=cv2.BORDER_WRAP),
OpticalDistortion(border_mode=cv2.BORDER_WRAP)
], p=0.5),
RandomSunFlare(flare_roi=(0, 0, 1, 0.25), num_flare_circles_lower=1, num_flare_circles_upper=2,
src_radius=150, p=0.5),
from manager import TrainingManager
if __name__ == '__main__':
opt = opts().parse()
video = cv2.VideoCapture(opt.video)
sc = 0.6
writer = cv2.VideoWriter(opt.video.replace(".avi", "result.avi"),
cv2.VideoWriter_fourcc(*'MJPG'), 25,
(int(int(video.get(3) * sc) // 32 * 32),
int(int(video.get(4) * sc) // 32 * 32)))
writer_map = cv2.VideoWriter(opt.video.replace(".avi", "result_map.avi"),
cv2.VideoWriter_fourcc("M", "J", "P",
"G"), 25,
(int(int(video.get(3) * sc) // 32 * 32),
int(int(video.get(4) * sc) // 32 * 32 * 2)))
logger = TrainingManager(opt.save_dir)
history = History(opt.save_dir, opt.resume)
torch.backends.cudnn.benchmark = True
print(opt)
os.environ['CUDA_VISIBLE_DEVICES'] = opt.gpus_str
opt.device = device('cuda' if opt.gpus[0] >= 0 else 'cpu')
losses, loss_weights = logger.loss
model = logger.model
params = logger.parameters(model)
optimizer = logger.optimizer(params)
lr_schedule = logger.lr_scheduler(optimizer)
model, optimizer, start_epoch, best = load_model(model, opt.load_model,
optimizer, opt.resume)
metrics = logger.metric
trainer = Trainer(model,
losses,
import numpy as np
import torch.utils.data
from albumentations import *
from sklearn.metrics import roc_auc_score
from torch import device
from history import History
from models.__init__ import load_model, save_model
from opts import opts
from trainers.classify_trainer import ClassifyTrainer as Trainer
from manager import TrainingManager
from transforms.random_lines import random_microscope, AdvancedHairAugmentation
if __name__ == '__main__':
opt = opts().parse()
logger = TrainingManager(opt.save_dir)
history = History(opt.save_dir, opt.resume)
torch.backends.cudnn.benchmark = True
print(opt)
transforms = {
"train":
Compose([
IAAAffine(shear=12, p=0.7),
ShiftScaleRotate(rotate_limit=45, scale_limit=(-0.5, 0.5)),
Flip(),
Transpose(),
ElasticTransform(alpha=100, sigma=25, p=0.5),
AdvancedHairAugmentation(hairs=10, hairs_folder="hairs"),
random_microscope(),
CoarseDropout(min_holes=8, max_width=16, max_height=16, p=0.75),
OneOf([