def __init__(self, args, img_paths, labels, mode=None):
assert mode == 'train' or mode == 'val'
self.args = args
self.img_size = args.input_size
self.img_paths = img_paths
self.labels = labels
self.mode = mode
self.transform = None
self.interval_rate = 0.1
# 在 custom_service.py 文件中一定要记得相应的更改
if self.mode == 'train':
self.transform = transforms.Compose([
# tr.CenterCrop(self.img_size, self.interval_rate), # 0.1, random = 0.5
tr.MaxResize(self.img_size, mode='train'), # 随机切割, rate=1.2
tr.RandomHorizontalFlip(), # 随机水平翻转 0.5
# tr.RandomGaussianBlur(), # 随机高斯模糊 # 0.5 random.random()
# tr.Normalize(mean=(0.485,0.456,0.406), std=(0.229,0.224,0.225)), # 标准化 mean= (0.485,0.456,0.406),std= (0.229,0.224,0.225)
tr.ToTensor(), # 2 tensor
])
elif self.mode == 'val':
self.transform = transforms.Compose([
tr.MaxResize(self.img_size, mode='val'),
# tr.Normalize(mean=(0.485,0.456,0.406), std=(0.229,0.224,0.225)), # 标准化
tr.ToTensor(), # 2 tensor
])
def train():
model = Dilated_Net().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = t.optim.Adam(model.parameters(), lr=1e-5, weight_decay=5e-4)
frame = MyFrame(model, criterion, K)
transform = transforms.Compose([tr.RandomHorizontalFlip(), tr.ToTensor()])
dataset = DAVIS2016(db_root_dir='/root/DAVIS-2016',
train=True,
transform=transform)
dataloader = DataLoader(dataset, batch_size=1, shuffle=True, num_workers=1)
loss_meter = meter.AverageValueMeter()
iou_meter = meter.AverageValueMeter()
max_iou = 0
for i in range(1000):
loss_meter.reset()
iou_meter.reset()
for ii, data in enumerate(dataloader):
img_1 = data['image_1'].to(device)
img_t = data['image_t'].to(device)
gt_1 = data['gt_1'].to(device)
gt_t = data['gt_t'].to(device)
if gt_1.sum() == 0:
continue
frame.set_data(img_1, img_t, gt_1, gt_t)
optimizer.zero_grad()
loss, IoU, pred = frame.forward()
loss.backward()
optimizer.step()
vis.img(name='train',
img_=overlay_mask(im_normalize(tens2image(img_t.cpu())),
tens2image(pred.cpu())).transpose(
2, 0, 1))
# loss_temp = loss
# loss_temp = loss_temp.detach().numpy
loss_meter.add(loss.item())
iou_meter.add(IoU.item())
print("train epoch:{}, loss:{} IoU:{}".format(i,
loss_meter.value()[0],
iou_meter.value()[0]))
vis.log(
"train epoch:{}, loss:{} IoU:{}".format(i,
loss_meter.value()[0],
iou_meter.value()[0]),
'loss_iou')
vis.plot('train_loss', loss_meter.value()[0])
vis.plot('train_iou', iou_meter.value()[0])
if (i + 1) % 5 == 0:
with torch.no_grad():
iou_val = evaluation(frame=frame)
if max_iou < iou_val:
max_iou = iou_val
frame.save_net('train11_28.pth')
def prepare_dataset(self):
cfg = self.cfg
self.print_log('Process dataset...')
composed_transforms = transforms.Compose([
tr.RandomScale(cfg.DATA_MIN_SCALE_FACTOR,
cfg.DATA_MAX_SCALE_FACTOR, cfg.DATA_SHORT_EDGE_LEN),
tr.RandomCrop(cfg.DATA_RANDOMCROP),
tr.RandomHorizontalFlip(cfg.DATA_RANDOMFLIP),
tr.Resize(cfg.DATA_RANDOMCROP),
tr.ToTensor()
])
train_datasets = []
if 'vecolor' in cfg.DATASETS:
train_vecolor_dataset = VECOLOR(root=cfg.DIR_VECOLOR,
phase='train',
transform=composed_transforms)
train_datasets.append(train_vecolor_dataset)
val_transforms = transforms.Compose(
[tr.Resize(cfg.DATA_RANDOMCROP),
tr.ToTensor()])
val_vecolor_dataset = VECOLOR(root=cfg.DIR_VECOLOR,
phase='val',
transform=val_transforms)
if len(train_datasets) > 1:
train_dataset = torch.utils.data.ConcatDataset(train_datasets)
elif len(train_datasets) == 1:
train_dataset = train_datasets[0]
else:
self.print_log('No dataset!')
exit(0)
val_dataset = val_vecolor_dataset
if cfg.DIST_ENABLE:
self.train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
self.val_sampler = torch.utils.data.distributed.DistributedSampler(
val_dataset)
else:
self.train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset, num_replicas=1, rank=0)
self.val_sampler = torch.utils.data.distributed.DistributedSampler(
val_dataset, num_replicas=1, rank=0)
self.trainloader = DataLoader(train_dataset,
batch_size=int(cfg.TRAIN_BATCH_SIZE /
cfg.TRAIN_GPUS),
shuffle=False,
num_workers=cfg.DATA_WORKERS,
pin_memory=True,
sampler=self.train_sampler)
self.valloader = DataLoader(val_dataset,
batch_size=1,
shuffle=False,
num_workers=0,
pin_memory=True,
sampler=self.val_sampler)
self.print_log('Done!')
gt_1 = data['gt_1'].to(device)
gt_t = data['gt_t'].to(device)
if (gt_1.sum() == 0):
continue
frame.set_data(img_1, img_t, gt_1, gt_t)
loss_val, IoU_val = frame.forward()
loss_meter_val.add(loss_val.item())
iou_meter_val.add(IoU_val.item())
return loss_meter_val.value()[0], iou_meter_val.value()[0]
model = Dilated_Net().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = t.optim.Adam(model.parameters(), lr=1e-5, weight_decay=5e-4)
frame = MyFrame(model, criterion, K)
transforms = transforms.Compose([tr.RandomHorizontalFlip(), tr.ToTensor()])
dataset = DAVIS_OVER_FIT_TEST1(db_root_dir='/root/DAVIS-2016',
train=True,
transform=transforms)
dataloader = DataLoader(dataset, batch_size=1, shuffle=True, num_workers=1)
loss_meter = meter.AverageValueMeter()
iou_meter = meter.AverageValueMeter()
max_iou = 0
for i in range(1000):
loss_meter.reset()
iou_meter.reset()
for ii, data in enumerate(dataloader):
img_1 = data['image_1'].to(device)
img_t = data['image_t'].to(device)
gt_1 = data['gt_1'].to(device)
gt_t = data['gt_t'].to(device)
exclude_db = [str(db) for db in self.excluded]
return 'Included datasets:' + str(
include_db) + '\n' + 'Excluded datasets:' + str(exclude_db)
if __name__ == "__main__":
import matplotlib.pyplot as plt
from dataloaders import pascal
from dataloaders import sbd
import torch
import numpy as np
import dataset.custom_transforms as tr
from torchvision import transforms
composed_transforms_tr = transforms.Compose([
tr.RandomHorizontalFlip(),
tr.ScaleNRotate(rots=(-15, 15), scales=(.75, 1.25)),
tr.FixedResize(resolutions={
'image': (450, 450),
'gt': (450, 450)
}),
tr.DistanceMap(v=0.15, elem='gt'),
tr.ConcatInputs(elems=('image', 'distance_map')),
tr.ToTensor()
])
composed_transforms_ts = transforms.Compose([
tr.FixedResize(resolutions={
'image': (450, 450),
'gt': (450, 450)
}),
def evaluation(frame=None):
with torch.no_grad():
if frame is None:
from models.dilated_net import Dilated_Net
model = Dilated_Net(pretrained=False)
model.load_state_dict(
t.load(
"/root/PycharmProjects/VideoMatch/checkpoint/train.pth"))
model = model.to(device)
criterion = nn.CrossEntropyLoss()
frame = MyFrame(model, criterion, K)
loss_meter_val = meter.AverageValueMeter()
iou_meter_val = meter.AverageValueMeter()
db_root_dir = '/root/DAVIS-2016'
fname = 'val_seqs'
with open(os.path.join(db_root_dir, fname + '.txt')) as f:
seqs = f.readlines()
val_files_len = len(seqs)
for val_index in range(val_files_len):
transform = transforms.Compose(
[tr.RandomHorizontalFlip(),
tr.ToTensor()])
dataset_val = DAVIS2016(db_root_dir='/root/DAVIS-2016',
train=False,
transform=transform,
val_index=val_index)
dataloader_val = DataLoader(dataset_val,
batch_size=1,
shuffle=True,
num_workers=1)
for ii, data in enumerate(dataloader_val):
img_1 = data['image_1'].to(device)
img_t = data['image_t'].to(device)
gt_1 = data['gt_1'].to(device)
gt_t = data['gt_t'].to(device)
if (gt_1.sum() == 0):
continue
frame.set_data(img_1, img_t, gt_1, gt_t)
loss_val, iou_val, pred = frame.forward()
vis.img(name='val',
img_=overlay_mask(
im_normalize(tens2image(img_t.cpu())),
tens2image(pred.cpu())).transpose(2, 0, 1))
loss_meter_val.add(loss_val.item())
iou_meter_val.add(iou_val.item())
print("evaluation on sequence:{}, loss: {}, IoU: {}".format(
seqs[val_index],
loss_meter_val.value()[0],
iou_meter_val.value()[0]))
vis.log(
"evaluation on sequence:{}, loss: {}, IoU: {}".format(
seqs[val_index],
loss_meter_val.value()[0],
iou_meter_val.value()[0]), 'loss_iou')
vis.plot('val_loss', loss_meter_val.value()[0])
vis.plot('val_iou', iou_meter_val.value()[0])
vis.log("avg of all sequences:{}".format(iou_meter_val.value()[0]),
'loss_iou')
if frame is not None:
return iou_meter_val.value()[0]