def load_model(out_channels=2, enc_type='efficientnet', dec_type='unet', pretrained=True, output_stride=8):
# Network
global net_type
if 'unet' in dec_type:
net_type = 'unet'
if 'efficient' in enc_type:
model = EfficientUnet(enc_type, out_channels=out_channels, concat_input=True,
pretrained=pretrained)#, model_name=enc_type)
else:
model = EncoderDecoderNet(**net_config)
else:
net_type = 'deeplab'
model = SPPNet(output_channels=out_channels, enc_type=enc_type, dec_type=dec_type, output_stride=output_stride)
return model
train_config = config['Train']
loss_config = config['Loss']
opt_config = config['Optimizer']
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
t_max = opt_config['t_max']
max_epoch = train_config['max_epoch']
batch_size = train_config['batch_size']
fp16 = train_config['fp16']
resume = train_config['resume']
pretrained_path = train_config['pretrained_path']
# Network
if 'unet' in net_config['dec_type']:
net_type = 'unet'
model = EncoderDecoderNet(**net_config)
else:
net_type = 'deeplab'
model = SPPNet(**net_config)
dataset = data_config['dataset']
if dataset == 'pascal':
from dataset.pascal_voc import PascalVocDataset as Dataset
net_config['output_channels'] = 21
classes = np.arange(1, 21)
elif dataset == 'cityscapes':
from dataset.cityscapes import CityscapesDataset as Dataset
net_config['output_channels'] = 19
classes = np.arange(1, 19)
elif dataset == 'sherbrooke':
from dataset.sherbrooke import SherbrookeDataset as Dataset
def process():
input_size = 224
timestamp = datetime.timestamp(datetime.now())
print("timestamp =", datetime.fromtimestamp(timestamp))
output_dir = Path(
os.path.join(
ROOT_DIR,
f'model/{model_name}_{datetime.fromtimestamp(timestamp)}'))
output_dir.mkdir(exist_ok=True)
# Opening JSON file
with open(
'/home/sfoucher/DEV/geoimagenet/dataset_test/deepglobe_classif/train/meta.json'
) as json_file:
meta = json.load(json_file)
dataset = {
'path': DATASET_ROOT / 'train',
DATASET_DATA_KEY: meta['patches']
}
with open(
'/home/sfoucher/DEV/geoimagenet/b76e46a7-4bea-4281-b802-2e296f97b960/meta.json'
) as json_file:
meta = json.load(json_file)
dataset = {
'path':
'/home/sfoucher/DEV/geoimagenet/b76e46a7-4bea-4281-b802-2e296f97b960',
DATASET_DATA_KEY: meta['patches']
}
with open(
'/home/sfoucher/DEV/geoimagenet/dataset_test/deepglobe_classif/val/meta.json'
) as json_file:
meta = json.load(json_file)
dataset_val = {
'path': DATASET_ROOT / 'val',
DATASET_DATA_KEY: meta['patches']
}
class_mapping = [('AgriculturalLand', 222), ('BarrenLand', 251),
('ForestLand', 232), ('RangeLand', 228),
('UrbanLand', 199), ('Water', 238)]
#model_class_map= {v: k k, v in class_mapping}
model_class_map = {
223: 'AgriculturalLand',
252: 'BarrenLand',
233: 'ForestLand',
229: 'RangeLand',
200: 'UrbanLand',
239: 'Water',
}
model_class_map = {
223: 0,
252: 1,
233: 2,
229: 3,
200: 4,
239: 5,
}
class GINParser(torch.utils.data.Dataset):
def __init__(self, dataset=None, transforms=None):
if not (isinstance(dataset, dict) and len(dataset)):
raise ValueError(
"Expected dataset parameters as configuration input.")
# thelper.data.Dataset.__init__(self, transforms=transforms, deepcopy=False)
self.root = Path(dataset["path"])
self.transforms = transforms
# keys matching dataset config for easy loading and referencing to same fields
self.image_key = IMAGE_DATA_KEY # key employed by loader to extract image data (pixel values)
self.label_key = IMAGE_LABEL_KEY # class id from API mapped to match model task
self.path_key = "path" # actual file path of the patch
self.idx_key = "index" # increment for __getitem__
self.mask_key = 'mask' # actual mask path of the patch
self.meta_keys = [
self.path_key, self.idx_key, DATASET_DATA_PATCH_CROPS_KEY,
DATASET_DATA_PATCH_IMAGE_KEY, DATASET_DATA_PATCH_FEATURE_KEY
]
# model_class_map = dataset[DATASET_DATA_KEY][DATASET_DATA_MAPPING_KEY]
sample_class_ids = set()
samples = []
for patch_info in dataset['data']:
# convert the dataset class ID into the model class ID using mapping, drop sample if not found
class_name = model_class_map.get(patch_info.get('class'))
class_name = str(patch_info.get('class'))
if class_name is not None:
sample_class_ids.add(class_name)
samples.append(deepcopy(patch_info))
samples[-1][
self.
path_key] = self.root / patch_info['crops'][0]['path']
samples[-1][self.label_key] = class_name
mask_name = patch_info['crops'][0].get('mask', None)
if mask_name:
samples[-1][self.mask_key] = self.root / mask_name
if not len(sample_class_ids):
raise ValueError(
"No patch/class could be retrieved from batch loading for specific model task."
)
self.samples = samples
self.sample_class_ids = sample_class_ids
def __len__(self):
return len(self.samples)
def __getitem__(self, idx):
if torch.is_tensor(idx):
idx = idx.tolist()
sample = self.samples[idx]
img_name = sample[self.path_key]._str
# image = cv.imread(img_name._str)
image = Image.open(img_name)
assert image is not None, "could not load image '%s' via opencv" % sample[
self.path_key]
#tensor_to_PIL = torchvision.transforms.ToPILImage(mode='RGB')
#image = tensor_to_PIL(image)
sample = {
self.image_key: image,
self.path_key: sample[self.path_key],
self.label_key: sample[self.label_key],
self.mask_key: sample[self.mask_key],
self.idx_key: idx
}
if self.transforms:
image = self.transforms(image)
return image, sample[self.label_key]
# Initialize the model for this run
# model_ft, input_size = initialize_model(model_name, num_classes, feature_extract, use_pretrained=True)
from models.net import EncoderDecoderNet
net_type = 'unet'
net_config = dict()
net_config['enc_type'] = 'resnet18'
net_config['dec_type'] = 'unet_seibn'
net_config['num_filters'] = 16
net_config['output_channels'] = 7
net_config['pretrained'] = True
model = EncoderDecoderNet(**net_config)
# Data augmentation and normalization for training
# Just normalization for validation
data_transforms = {
'train':
transforms.Compose([
transforms.Resize(input_size),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
'val':
transforms.Compose([
transforms.Resize(input_size),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
}
image_datasets = dict()
dataloaders_dict = dict()
image_datasets['train'] = GINParser(dataset=dataset,
transforms=data_transforms['train'])
image_datasets['val'] = GINParser(dataset=dataset_val,
transforms=data_transforms['val'])
dataloaders_dict['train'] = torch.utils.data.DataLoader(
image_datasets['train'],
batch_size=batch_size,
shuffle=True,
num_workers=4)
dataloaders_dict['val'] = torch.utils.data.DataLoader(
image_datasets['val'],
batch_size=batch_size,
shuffle=False,
num_workers=4)
# Detect if we have a GPU available
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
with tqdm(dataloaders_dict['train']) as _tqdm:
for i, (inputs, labels) in enumerate(_tqdm):
#for inputs, labels in dataloaders[phase]:
inputs = inputs.to(device)
labels = labels.to(device)
# Print the model we just instantiated
print(model_ft)
# Send the model to GPU
model_ft = model_ft.to(device)
# Gather the parameters to be optimized/updated in this run. If we are
# finetuning we will be updating all parameters. However, if we are
# doing feature extract method, we will only update the parameters
# that we have just initialized, i.e. the parameters with requires_grad
# is True.
params_to_update = model_ft.parameters()
# Observe that all parameters are being optimized
optimizer_ft = optim.SGD(params_to_update, lr=lr, momentum=0.9)
print("Params to learn:")
if feature_extract:
params_to_update = []
for name, param in model_ft.named_parameters():
if param.requires_grad == True:
params_to_update.append(param)
print("\t", name)
optimizer_ft = optim.SGD(params_to_update, lr=lr, momentum=0.9)
else:
n_param = 0
name_block = dict()
lr_blcok = dict()
for l, (name, param) in enumerate(model_ft.named_parameters()):
n_param += 1
for l, (name, param) in enumerate(model_ft.named_parameters()):
blcok = name.split('.')[0]
name_block[l] = blcok
lr_blcok[blcok] = lr * 10**(2 * (l / n_param - 1))
params_to_update = []
#name_block = set(name_block)
for l, (name, param) in enumerate(model_ft.named_parameters()):
if l < int(n_param) - 8:
param.requires_grad = False
else:
if 'fc.' not in name:
params_to_update.append({
"params": param,
"lr": lr_blcok[name_block[l]],
})
else:
params_to_update.append({
"params": param,
"lr": lr,
})
if param.requires_grad == True:
print("\t", name)
optimizer_ft = optim.SGD(params_to_update,
momentum=0.9,
weight_decay=0.001)
# Setup the loss fxn
criterion = nn.CrossEntropyLoss()
def train_model(model,
dataloaders,
criterion,
optimizer,
num_epochs=25,
is_inception=False):
since = time.time()
val_acc_history = []
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=3,
gamma=0.1)
best_model_wts = copy.deepcopy(model.state_dict())
best_acc = 0.0
ma_loss = 0.0
ma_iou = 0.0
i_iter = 0
for epoch in range(num_epochs):
print('Epoch {}/{}'.format(epoch, num_epochs - 1))
print('-' * 10)
# Each epoch has a training and validation phase
for phase in dataloaders.keys():
if phase == 'train':
model.train() # Set model to training mode
else:
model.eval() # Set model to evaluate mode
running_loss = 0.0
running_corrects = 0
# Iterate over data.
with tqdm(dataloaders[phase]) as _tqdm:
for i, (inputs, labels) in enumerate(_tqdm):
#for inputs, labels in dataloaders[phase]:
inputs = inputs.to(device)
labels = labels.to(device)
# zero the parameter gradients
optimizer.zero_grad()
# forward
# track history if only in train
with torch.set_grad_enabled(phase == 'train'):
# Get model outputs and calculate loss
# Special case for inception because in training it has an auxiliary output. In train
# mode we calculate the loss by summing the final output and the auxiliary output
# but in testing we only consider the final output.
if is_inception and phase == 'train':
# From https://discuss.pytorch.org/t/how-to-optimize-inception-model-with-auxiliary-classifiers/7958
outputs, aux_outputs = model(inputs)
loss1 = criterion(outputs, labels)
loss2 = criterion(aux_outputs, labels)
loss = loss1 + 0.4 * loss2
else:
outputs = model(inputs)
loss = criterion(outputs, labels)
_, preds = torch.max(outputs, 1)
acc = np.logical_and(
preds.cpu().numpy(),
labels.cpu().numpy()).sum() / len(labels)
_tqdm.set_postfix(
OrderedDict(seg_loss=f'{loss.item():.5f}',
acc=f'{acc*100:.3f}'))
# backward + optimize only if in training phase
if phase == 'train':
loss.backward()
optimizer.step()
ma_loss = 0.01 * loss.item() + 0.99 * ma_loss
ma_iou = 0.01 * acc + 0.99 * ma_iou
plotter.plot(
'loss', 'train', 'iteration Loss', i_iter,
loss.item(
)) # y-axis, name serie, name chart
plotter.plot('acc', 'train', 'iteration acc',
i_iter, acc)
plotter.plot('loss', 'ma_loss',
'iteration Loss', i_iter, ma_loss)
plotter.plot('acc', 'ma_acc', 'iteration acc',
i_iter, ma_iou)
i_iter += 1
else:
i_iter = i_iter
# statistics
running_loss += loss.item() * inputs.size(0)
running_corrects += torch.sum(preds == labels.data)
epoch_loss = running_loss / len(dataloaders[phase].dataset)
epoch_acc = running_corrects.double().item() / len(
dataloaders[phase].dataset)
print('{} Loss: {:.4f} Acc: {:.4f}'.format(
phase, epoch_loss, epoch_acc))
if phase == 'train':
plotter.plot('loss-epoch', 'train', 'epoch Loss', epoch,
epoch_loss)
plotter.plot('acc-epoch', 'train', 'epoch acc', epoch,
epoch_acc)
else:
plotter.plot('loss-epoch', 'valid', 'epoch Loss', epoch,
epoch_loss)
plotter.plot('acc-epoch', 'valid', 'epoch acc', epoch,
epoch_acc)
# deep copy the model
if phase == 'val' and epoch_acc > best_acc:
print('Best Epoch!')
best_acc = epoch_acc
best_model_wts = copy.deepcopy(model.state_dict())
torch.save(model.state_dict(),
output_dir.joinpath('model.pth'))
torch.save(optimizer.state_dict(),
output_dir.joinpath('opt.pth'))
if phase == 'val':
val_acc_history.append(epoch_acc)
print()
# scheduler.step()
time_elapsed = time.time() - since
print('Training complete in {:.0f}m {:.0f}s'.format(
time_elapsed // 60, time_elapsed % 60))
print('Best val Acc: {:4f}'.format(best_acc))
# load best model weights
model.load_state_dict(best_model_wts)
return model, val_acc_history
# Train and evaluate
model_ft, hist = train_model(model_ft,
dataloaders_dict,
criterion,
optimizer_ft,
num_epochs=num_epochs,
is_inception=(model_name == "inception"))
output_dir = Path(train_config['output_dir'])
output_dir.mkdir(exist_ok=True, parents=True)
shutil.copyfile(str(config_path), str(output_dir / config_path.name))
log_dir = Path(train_config['log_dir'])
log_dir.mkdir(exist_ok=True, parents=True)
eval_every_n_epochs = train_config['eval_every_n_epochs']
vis_flag = train_config['vis_flag']
include_bg = loss_config['include_bg']
del loss_config['include_bg']
# Network
if 'unet' in net_config['dec_type']:
net_type = 'unet'
model = EncoderDecoderNet(**net_config)
else:
net_type = 'deeplab'
model = SPPNet(**net_config)
dataset = data_config['dataset']
if dataset == 'pascal':
from dataset.pascal_voc import PascalVocDataset as Dataset
elif dataset == 'cityscapes':
from dataset.cityscapes import CityscapesDataset as Dataset
else:
raise NotImplementedError
if include_bg:
classes = np.arange(0, data_config['num_classes'])
else:
classes = np.arange(1, data_config['num_classes'])
parser.add_argument("--vis", action="store_true")
args = parser.parse_args()
config_path = Path(args.config_path)
tta_flag = args.tta
vis_flag = args.vis
config = yaml.load(open(config_path))
net_config = config["Net"]
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
modelname = config_path.stem
model_path = Path("../model") / modelname / "model.pth"
if "unet" in net_config["dec_type"]:
net_type = "unet"
model = EncoderDecoderNet(**net_config)
else:
net_type = "deeplab"
model = SPPNet(**net_config)
model.to(device)
model.update_bn_eps()
param = torch.load(model_path)
model.load_state_dict(param)
del param
model.eval()
batch_size = 1
scales = [0.25, 0.75, 1, 1.25]
train_config = config['Train']
loss_config = config['Loss']
opt_config = config['Optimizer']
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
t_max = opt_config['t_max']
max_epoch = train_config['max_epoch']
batch_size = train_config['batch_size']
fp16 = train_config['fp16']
resume = train_config['resume']
pretrained_path = train_config['pretrained_path']
# Network
if 'unet' in net_config['dec_type']:
net_type = 'unet'
model = EncoderDecoderNet(**net_config)
else:
net_type = 'deeplab'
model = SPPNet(**net_config)
dataset = data_config['dataset']
if dataset == 'pascal':
from dataset.pascal_voc import PascalVocDataset as Dataset
net_config['output_channels'] = 21
classes = np.arange(1, 21)
elif dataset == 'cityscapes':
from dataset.cityscapes import CityscapesDataset as Dataset
net_config['output_channels'] = 19
classes = np.arange(1, 19)
else:
raise NotImplementedError
output_dir.mkdir(exist_ok=True, parents=True)
log_dir = Path('../logs').joinpath(modelname)
log_dir.mkdir(exist_ok=True, parents=True)
logger = debug_logger(log_dir)
logger.info(f'Device: {device}')
logger.info(f'Max Epoch: {max_epoch}')
del data_config['dataset']
train_dataset = Dataset(split='train', **data_config)
valid_dataset = Dataset(split='valid', **data_config)
train_loader = DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4,
pin_memory=True)
valid_loader = DataLoader(valid_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=4,
pin_memory=True)
if 'unet' in net_config['dec_type']:
model = EncoderDecoderNet(**net_config).to(device)
else:
model = SPPNet(**net_config).to(device)
loss_fn = CrossEntropy2d(**loss_config).to(device)
optimizer, scheduler = create_optimizer(model=model, **opt_config)
train()
def __init__(
self,
model_path='../model/deepglobe_deeplabv3_weights-cityscapes_19-outputs/model.pth',
dataset='deepglobe',
output_channels=19,
split='valid',
net_type='deeplab',
batch_size=1,
shuffle=True):
"""
Initializes the tester by loading the model with the good parameters.
:param model_path: Path to model weights
:param dataset: dataset used amongst {'deepglobe', 'pascal', 'cityscapes'}
:param output_channels: num of output channels of model
:param split: split to be used amongst {'train', 'valid'}
:param net_type: model type to be used amongst {'deeplab', 'unet'}
:param batch_size: batch size when loading images (always 1 here)
:param shuffle: when loading images from dataset
"""
model_path = '/home/sfoucher/DEV/pytorch-segmentation/model/my_pascal_unet_res18_scse/model.pth'
dataset_dir = '/home/sfoucher/DEV/pytorch-segmentation/data/deepglobe_as_pascalvoc/VOCdevkit/VOC2012'
output_channels = 8
net_type = 'unet'
print('[Tester] [Init] Initializing tester...')
self.dataset = dataset
self.model_path = model_path
# Load model
print('[Tester] [Init] Loading model ' + model_path + ' with ' +
str(output_channels) + ' output channels...')
self.device = torch.device(
'cuda:0' if torch.cuda.is_available() else 'cpu')
if net_type == 'unet':
self.model = EncoderDecoderNet(output_channels=8,
enc_type='resnet18',
dec_type='unet_scse',
num_filters=8)
else:
self.model = SPPNet(output_channels=output_channels).to(
self.device)
param = torch.load(model_path)
self.model.load_state_dict(param)
del param
# Create data loader depending on dataset, split and net type
if dataset == 'pascal':
self.valid_dataset = PascalVocDataset(split=split,
net_type=net_type)
elif dataset == 'cityscapes':
self.valid_dataset = CityscapesDataset(split=split,
net_type=net_type)
elif dataset == 'deepglobe':
self.valid_dataset = DeepGlobeDataset(base_dir=dataset_dir,
target_size=(64, 64),
split=split,
net_type=net_type)
else:
raise NotImplementedError
self.valid_loader = DataLoader(self.valid_dataset,
batch_size=batch_size,
shuffle=shuffle)
print('[Tester] [Init] ...done!')
print('[Tester] [Init] Tester created.')
def process(config_path):
gc.collect()
torch.cuda.empty_cache()
config = yaml.load(open(config_path))
net_config = config['Net']
data_config = config['Data']
train_config = config['Train']
loss_config = config['Loss']
opt_config = config['Optimizer']
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
t_max = opt_config['t_max']
# Collect training parameters
max_epoch = train_config['max_epoch']
batch_size = train_config['batch_size']
fp16 = train_config['fp16']
resume = train_config['resume']
pretrained_path = train_config['pretrained_path']
freeze_enabled = train_config['freeze']
seed_enabled = train_config['seed']
#########################################
# Deterministic training
if seed_enabled:
seed = 100
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed=seed)
import random
random.seed(a=100)
#########################################
# Network
if 'unet' in net_config['dec_type']:
net_type = 'unet'
model = EncoderDecoderNet(**net_config)
else:
net_type = 'deeplab'
net_config['output_channels'] = 19
model = SPPNet(**net_config)
dataset = data_config['dataset']
if dataset == 'deepglobe-dynamic':
from dataset.deepglobe_dynamic import DeepGlobeDatasetDynamic as Dataset
net_config['output_channels'] = 7
classes = np.arange(0, 7)
else:
raise NotImplementedError
del data_config['dataset']
modelname = config_path.stem
timestamp = datetime.timestamp(datetime.now())
print("timestamp =", datetime.fromtimestamp(timestamp))
output_dir = Path(os.path.join(ROOT_DIR, f'model/{modelname}_{datetime.fromtimestamp(timestamp)}') )
output_dir.mkdir(exist_ok=True)
log_dir = Path(os.path.join(ROOT_DIR, f'logs/{modelname}_{datetime.fromtimestamp(timestamp)}') )
log_dir.mkdir(exist_ok=True)
dataset_dir= '/home/sfoucher/DEV/pytorch-segmentation/data/deepglobe_as_pascalvoc/VOCdevkit/VOC2012'
logger = debug_logger(log_dir)
logger.debug(config)
logger.info(f'Device: {device}')
logger.info(f'Max Epoch: {max_epoch}')
# Loss
loss_fn = MultiClassCriterion(**loss_config).to(device)
params = model.parameters()
optimizer, scheduler = create_optimizer(params, **opt_config)
# history
if resume:
with open(log_dir.joinpath('history.pkl'), 'rb') as f:
history_dict = pickle.load(f)
best_metrics = history_dict['best_metrics']
loss_history = history_dict['loss']
iou_history = history_dict['iou']
start_epoch = len(iou_history)
for _ in range(start_epoch):
scheduler.step()
else:
start_epoch = 0
best_metrics = 0
loss_history = []
iou_history = []
affine_augmenter = albu.Compose([albu.HorizontalFlip(p=.5),albu.VerticalFlip(p=.5)
# Rotate(5, p=.5)
])
# image_augmenter = albu.Compose([albu.GaussNoise(p=.5),
# albu.RandomBrightnessContrast(p=.5)])
image_augmenter = None
# This has been put in the loop for the dynamic training
"""
# Dataset
train_dataset = Dataset(affine_augmenter=affine_augmenter, image_augmenter=image_augmenter,
net_type=net_type, **data_config)
valid_dataset = Dataset(split='valid', net_type=net_type, **data_config)
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=4,
pin_memory=True, drop_last=True)
valid_loader = DataLoader(valid_dataset, batch_size=1, shuffle=False, num_workers=4, pin_memory=True)
"""
# Pretrained model
if pretrained_path:
logger.info(f'Resume from {pretrained_path}')
param = torch.load(pretrained_path)
model.load_state_dict(param)
model.logits = torch.nn.Conv2d(256, net_config['output_channels'], 1)
del param
# To device
model = model.to(device)
#########################################
if freeze_enabled:
# Code de Rémi
# Freeze layers
for param_index in range(int((len(optimizer.param_groups[0]['params']))*0.5)):
optimizer.param_groups[0]['params'][param_index].requires_grad = False
#########################################
params_to_update = model.parameters()
print("Params to learn:")
if freeze_enabled:
params_to_update = []
for name,param in model.named_parameters():
if param.requires_grad == True:
params_to_update.append(param)
print("\t",name)
optimizer, scheduler = create_optimizer(params_to_update, **opt_config)
# fp16
if fp16:
# I only took the necessary files because I don't need the C backend of apex,
# which is broken and can't be installed
# from apex import fp16_utils
from utils.apex.apex.fp16_utils.fp16util import BN_convert_float
from utils.apex.apex.fp16_utils.fp16_optimizer import FP16_Optimizer
# model = fp16_utils.BN_convert_float(model.half())
model = BN_convert_float(model.half())
# optimizer = fp16_utils.FP16_Optimizer(optimizer, verbose=False, dynamic_loss_scale=True)
optimizer = FP16_Optimizer(optimizer, verbose=False, dynamic_loss_scale=True)
logger.info('Apply fp16')
# Restore model
if resume:
model_path = output_dir.joinpath(f'model_tmp.pth')
logger.info(f'Resume from {model_path}')
param = torch.load(model_path)
model.load_state_dict(param)
del param
opt_path = output_dir.joinpath(f'opt_tmp.pth')
param = torch.load(opt_path)
optimizer.load_state_dict(param)
del param
i_iter = 0
ma_loss= 0
ma_iou= 0
# Train
for i_epoch in range(start_epoch, max_epoch):
logger.info(f'Epoch: {i_epoch}')
logger.info(f'Learning rate: {optimizer.param_groups[0]["lr"]}')
train_losses = []
train_ious = []
model.train()
# Initialize randomized but balanced datasets
train_dataset = Dataset(base_dir = dataset_dir,
affine_augmenter=affine_augmenter, image_augmenter=image_augmenter,
net_type=net_type, **data_config)
valid_dataset = Dataset(base_dir = dataset_dir,
split='valid', net_type=net_type, **data_config)
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=4,
pin_memory=True, drop_last=True)
valid_loader = DataLoader(valid_dataset, batch_size=1, shuffle=False, num_workers=4, pin_memory=True)
with tqdm(train_loader) as _tqdm:
for i, batched in enumerate(_tqdm):
images, labels = batched
if fp16:
images = images.half()
images, labels = images.to(device), labels.to(device)
optimizer.zero_grad()
preds = model(images)
if net_type == 'deeplab':
preds = F.interpolate(preds, size=labels.shape[1:], mode='bilinear', align_corners=True)
if fp16:
loss = loss_fn(preds.float(), labels)
else:
loss = loss_fn(preds, labels)
preds_np = preds.detach().cpu().numpy()
labels_np = labels.detach().cpu().numpy()
iou = compute_iou_batch(np.argmax(preds_np, axis=1), labels_np, classes)
_tqdm.set_postfix(OrderedDict(seg_loss=f'{loss.item():.5f}', iou=f'{iou:.3f}'))
train_losses.append(loss.item())
train_ious.append(iou)
ma_loss= 0.01*loss.item() + 0.99 * ma_loss
ma_iou= 0.01*iou + 0.99 * ma_iou
plotter.plot('loss', 'train', 'iteration Loss', i_iter, loss.item())
plotter.plot('iou', 'train', 'iteration iou', i_iter, iou)
plotter.plot('loss', 'ma_loss', 'iteration Loss', i_iter, ma_loss)
plotter.plot('iou', 'ma_iou', 'iteration iou', i_iter, ma_iou)
if fp16:
optimizer.backward(loss)
else:
loss.backward()
optimizer.step()
i_iter += 1
scheduler.step()
train_loss = np.mean(train_losses)
train_iou = np.nanmean(train_ious)
logger.info(f'train loss: {train_loss}')
logger.info(f'train iou: {train_iou}')
plotter.plot('loss-epoch', 'train', 'iteration Loss', i_epoch, train_loss)
plotter.plot('iou-epoch', 'train', 'iteration iou', i_epoch, train_iou)
torch.save(model.state_dict(), output_dir.joinpath('model_tmp.pth'))
torch.save(optimizer.state_dict(), output_dir.joinpath('opt_tmp.pth'))
valid_losses = []
valid_ious = []
model.eval()
with torch.no_grad():
with tqdm(valid_loader) as _tqdm:
for batched in _tqdm:
images, labels = batched
if fp16:
images = images.half()
images, labels = images.to(device), labels.to(device)
preds = model.tta(images, net_type=net_type)
if fp16:
loss = loss_fn(preds.float(), labels)
else:
loss = loss_fn(preds, labels)
preds_np = preds.detach().cpu().numpy()
labels_np = labels.detach().cpu().numpy()
# I changed a parameter in the compute_iou method to prevent it from yielding nans
iou = compute_iou_batch(np.argmax(preds_np, axis=1), labels_np, classes)
_tqdm.set_postfix(OrderedDict(seg_loss=f'{loss.item():.5f}', iou=f'{iou:.3f}'))
valid_losses.append(loss.item())
valid_ious.append(iou)
valid_loss = np.mean(valid_losses)
valid_iou = np.mean(valid_ious)
logger.info(f'valid seg loss: {valid_loss}')
logger.info(f'valid iou: {valid_iou}')
plotter.plot('loss-epoch', 'valid', 'iteration Loss', i_epoch, valid_loss)
plotter.plot('iou-epoch', 'valid', 'iteration iou', i_epoch, valid_iou)
if best_metrics < valid_iou:
best_metrics = valid_iou
logger.info('Best Model!')
torch.save(model.state_dict(), output_dir.joinpath('model.pth'))
torch.save(optimizer.state_dict(), output_dir.joinpath('opt.pth'))
loss_history.append([train_loss, valid_loss])
iou_history.append([train_iou, valid_iou])
history_ploter(loss_history, log_dir.joinpath('loss.png'))
history_ploter(iou_history, log_dir.joinpath('iou.png'))
history_dict = {'loss': loss_history,
'iou': iou_history,
'best_metrics': best_metrics}
with open(log_dir.joinpath('history.pkl'), 'wb') as f:
pickle.dump(history_dict, f)