# Use focal loss or not
focal = False
# Use centernet-like Blazeface
center = False
# # set up person only VOC dataset
# In[3]:
# load files
vocpath = os.path.join("..", "VOCdevkit", "VOC2007")
train_img_list, train_anno_list, val_img_list, val_anno_list = make_datapath_list(vocpath, cls="person")
# extend with VOC2012
vocpath = "../VOCdevkit/VOC2012"
train_img_list2, train_anno_list2, _, _ = make_datapath_list(vocpath, cls="person", VOC2012=True)
train_img_list.extend(train_img_list2)
train_anno_list.extend(train_anno_list2)
# make Dataset
voc_classes = ['person']
color_mean = (104, 117, 123) # (BGR)の色の平均値
print("trainlist: ", len(train_img_list))
print("vallist: ", len(val_img_list))
def main():
args = parser()
### setup configs ###
configfile = args.configfile
with open(configfile) as f:
configs = yaml.load(f)
## path process (path definition, make directories)
now = datetime.now().isoformat()
log_dir = Path(configs['log_dir']) / now
paths = Paths(log_dir=log_dir)
### setup logs and summary writer ###
setup_logger(logfile=paths.logfile)
writer = SummaryWriter(str(paths.summary_dir))
### setup GPU or CPU ###
if configs['n_gpus'] > 0 and torch.cuda.is_available():
logger.info('CUDA is available! using GPU...\n')
device = torch.device('cuda')
else:
logger.info('using CPU...\n')
device = torch.device('cpu')
### Dataset ###
logger.info('preparing dataset...')
dataset_name = configs['dataset']
logger.info(f'==> dataset: {dataset_name}\n')
if configs['dataset'] == 'cifar10':
transform = transforms.Compose([
transforms.Resize(configs['img_size'], configs['img_size']),
transforms.ToTensor(),
transforms.Normalize(configs['color_mean'], configs['color_std']),
])
train_dataset = datasets.CIFAR10(root=configs['data_root'],
train=True,
transform=transform,
download=True)
test_dataset = datasets.CIFAR10(root=configs['data_root'],
train=False,
transform=transform,
download=True)
elif configs['dataset'] == 'custom':
train_transform = DataTransforms(img_size=configs['img_size'],
color_mean=configs['color_mean'],
color_std=configs['color_std'],
phase='train')
test_transform = DataTransforms(img_size=configs['img_size'],
color_mean=configs['color_mean'],
color_std=configs['color_std'],
phase='test')
train_img_list, train_lbl_list, test_img_list, test_lbl_list = make_datapath_list(
root=configs['data_root'])
train_dataset = Dataset(train_img_list,
train_lbl_list,
transform=train_transform)
test_dataset = Dataset(test_img_list,
test_lbl_list,
transform=test_transform)
else:
logger.debug('dataset is not supported')
raise ValueError('dataset is not supported')
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=configs['batch_size'],
shuffle=True,
num_workers=8)
test_loader = torch.utils.data.DataLoader(test_dataset,
batch_size=configs['batch_size'],
shuffle=False,
num_workers=8)
### Network ###
logger.info('preparing network...')
network = CNNAutoEncoder(in_channels=configs['n_channels'],
n_classes=configs['n_classes'])
network = network.to(device)
cnn_criterion = nn.CrossEntropyLoss()
ae_criterion = nn.MSELoss()
optimizer = optim.Adam(network.parameters(), lr=configs['lr'])
if configs['resume']:
# Load checkpoint
logger.info('==> Resuming from checkpoint...\n')
if not Path(configs['resume']).exists():
logger.info('No checkpoint found !')
raise ValueError('No checkpoint found !')
ckpt = torch.load(configs['resume'])
network.load_state_dict(ckpt['model_state_dict'])
optimizer.load_state_dict(ckpt['optimizer_state_dict'])
start_epoch = ckpt['epoch']
loss = ckpt['loss']
else:
logger.info('==> Building model...\n')
start_epoch = 0
logger.info('model summary: ')
summary(network,
input_size=(configs['n_channels'], configs['img_size'],
configs['img_size']))
if configs["n_gpus"] > 1:
network = nn.DataParallel(network)
### Metrics ###
metrics = Metrics(n_classes=configs['n_classes'],
classes=configs['classes'],
writer=writer,
metrics_dir=paths.metrics_dir,
plot_confusion_matrix=plot_confusion_matrix)
### Train or Test ###
kwargs = {
'device': device,
'network': network,
'optimizer': optimizer,
'criterions': (cnn_criterion, ae_criterion),
'classification_loss_weight': configs['classification_loss_weight'],
'autoencoder_loss_weight': configs['autoencoder_loss_weight'],
'data_loaders': (train_loader, test_loader),
'metrics': metrics,
'writer': writer,
'n_classses': configs['n_classes'],
'save_ckpt_interval': configs['save_ckpt_interval'],
'ckpt_dir': paths.ckpt_dir,
}
cnn_classifier = CNNClassifier(**kwargs)
if args.inference:
if not configs['resume']:
logger.info('No checkpoint found for inference!')
logger.info('mode: inference\n')
cnn_classifier.test(epoch=start_epoch, inference=True)
else:
logger.info('mode: train\n')
cnn_classifier.train(n_epochs=configs['n_epochs'],
start_epoch=start_epoch)
# select from efficientnet backbone or resnet backbone
backbone = "efficientnet-b0"
scale = 1
useBiFPN = True
# scale==1: resolution 300
# scale==2: resolution 600
# ## make data.Dataset for training
# In[4]:
# load files
# set your VOCdevkit path here.
vocpath = "../VOCdevkit/VOC2007"
train_img_list, train_anno_list, val_img_list, val_anno_list = make_datapath_list(
vocpath)
vocpath = "../VOCdevkit/VOC2012"
train_img_list2, train_anno_list2, _, _ = make_datapath_list(vocpath)
train_img_list.extend(train_img_list2)
train_anno_list.extend(train_anno_list2)
print("trainlist: ", len(train_img_list))
print("vallist: ", len(val_img_list))
# make Dataset
voc_classes = [
'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat',
'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person',
'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor'
import torch.nn as nn
import torch.nn.init as init
import torch.nn.functional as F
from torch.autograd import Function
import pandas as pd
# import dataset
from utils.dataset import VOCDataset, DatasetTransform, make_datapath_list, Anno_xml2list, od_collate_fn
# # set up person only VOC dataset
# In[2]:
# load files
vocpath = os.path.join("..", "VOCdevkit", "VOC2007")
train_img_list, train_anno_list, val_img_list, val_anno_list = make_datapath_list(
vocpath, cls="person")
# make Dataset
voc_classes = ['person']
color_mean = (104, 117, 123) # (BGR)の色の平均値
input_size = 128 # 画像のinputサイズを300×300にする
## DatasetTransformを適応
transform = DatasetTransform(input_size, color_mean)
transform_anno = Anno_xml2list(voc_classes)
train_dataset = VOCDataset(train_img_list,
train_anno_list,
phase="train",
transform=transform,
transform_anno=transform_anno)
def main():
args = parser()
### setup configs ###
configfile = args.configfile
with open(configfile) as f:
configs = yaml.load(f)
## path process (path definition, make directories)
now = datetime.now().isoformat()
log_dir = Path(configs['log_dir']) / now
paths = Paths(log_dir=log_dir)
### setup logs and summary writer ###
setup_logger(logfile=paths.logfile)
writer = SummaryWriter(str(paths.summary_dir))
### setup GPU or CPU ###
if configs['n_gpus'] > 0 and torch.cuda.is_available():
logger.info('CUDA is available! using GPU...\n')
device = torch.device('cuda')
else:
logger.info('using CPU...\n')
device = torch.device('cpu')
### Dataset ###
logger.info('preparing dataset...')
train_transform = DataTransforms(img_size=configs['img_size'],
color_mean=configs['color_mean'],
color_std=configs['color_std'],
phase='train')
test_transform = DataTransforms(img_size=configs['img_size'],
color_mean=configs['color_mean'],
color_std=configs['color_std'],
phase='test')
train_img_list, test_img_list = make_datapath_list(
root=configs['data_root'])
train_dataset = Dataset(train_img_list, transform=train_transform)
test_dataset = Dataset(test_img_list, transform=test_transform)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=configs['batch_size'],
shuffle=True,
num_workers=8)
test_loader = torch.utils.data.DataLoader(test_dataset,
batch_size=configs['batch_size'],
shuffle=False,
num_workers=8)
### Network ###
logger.info('preparing network...')
network = VAE(in_channels=configs['n_channels'],
h_dim=1024,
z_dim=32,
device=device)
network = network.to(device)
criterion = nn.MSELoss()
optimizer = optim.Adam(network.parameters(), lr=configs['lr'])
if configs['resume']:
# Load checkpoint
logger.info('==> Resuming from checkpoint...\n')
if not Path(configs['resume']).exists():
logger.info('No checkpoint found !')
raise ValueError('No checkpoint found !')
ckpt = torch.load(configs['resume'])
network.load_state_dict(ckpt['model_state_dict'])
optimizer.load_state_dict(ckpt['optimizer_state_dict'])
start_epoch = ckpt['epoch']
loss = ckpt['loss']
else:
logger.info('==> Building model...\n')
start_epoch = 0
logger.info('model summary: ')
summary(network,
input_size=(configs['n_channels'], configs['img_size'],
configs['img_size']))
if configs["n_gpus"] > 1:
network = nn.DataParallel(network)
### Metrics ###
metrics = Metrics(writer=writer, metrics_dir=paths.metrics_dir)
### Train or Test ###
kwargs = {
'device': device,
'network': network,
'optimizer': optimizer,
'criterion': criterion,
'data_loaders': (train_loader, test_loader),
'metrics': metrics,
'writer': writer,
'save_ckpt_interval': configs['save_ckpt_interval'],
'ckpt_dir': paths.ckpt_dir,
'img_outdir': paths.img_outdir,
}
generalizer = Generalizer(**kwargs)
if args.inference:
if not configs['resume']:
logger.info('No checkpoint found for inference!')
logger.info('mode: inference\n')
generalizer.test(epoch=start_epoch, inference=True)
else:
logger.info('mode: train\n')
generalizer.train(n_epochs=configs['n_epochs'],
start_epoch=start_epoch)
from utils.dataset import make_datapath_list, VOCDataset, DataTransform, Anno_xml2list
from utils.dataloader import od_collate_fn
import torch.nn as nn
import torch.nn.init as init
import torch
import pandas as pd
import torch.optim as optim
import time
import random
import numpy as np
torch.manual_seed(42)
np.random.seed(42)
random.seed(42)
rootpath = "./data/VOCdevkit/VOC2012/"
train_img_list, train_anno_list, val_img_list, val_anno_list = make_datapath_list(rootpath)
voc_classes = ['aeroplane', 'bicycle', 'bird', 'boat',
'bottle', 'bus', 'car', 'cat', 'chair',
'cow', 'diningtable', 'dog', 'horse',
'motorbike', 'person', 'pottedplant',
'sheep', 'sofa', 'train', 'tvmonitor']
color_mean = (104, 117, 123)
input_size = 300
train_dataset = VOCDataset(train_img_list, train_anno_list, phase='train',
transform=DataTransform(input_size, color_mean),
transform_anno=Anno_xml2list(voc_classes))
val_dataset = VOCDataset(val_img_list, val_anno_list, phase='val',