def tta_predict(network, path_model, args):
# load model
if network.load(path_model) is not True:
return 0
tta_preprocess = [
get_transforms_det(network.size_input),
get_transforms_hflip(network.size_input),
get_transforms_gray(network.size_input),
get_transforms_aug(network.size_input),
get_transforms_aug(network.size_input),
]
dataloaders = []
for transform in tta_preprocess:
# test dataset
data = gdata.Dataset(
data=factory.FactoryDataset.factory(pathname=args.data,
name=args.name_dataset,
subset=factory.validation,
download=True),
num_channels=network.num_input_channels,
transform=transform,
)
dataloader = DataLoader(test_data,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers)
dataloaders.append(dataloader)
# print neural net class
print('NeuralNet: {}'.format(datetime.datetime.now()), flush=True)
#print(network)
pathnameout = args.out
print('dir: {}'.format(pathnameout))
files = [
f for f in sorted(os.listdir(pathnameout)) if f.split('.')[-1] == 'csv'
]
for i, data in enumerate(dataloaders):
Yhat, Y = network.test(data)
df = pd.DataFrame(np.concatenate((Y, Yhat), axis=1))
df.to_csv(os.path.join(pathnameout,
'val_dp{}.csv'.format(i + len(files))),
index=False,
encoding='utf-8')
print('DONE !!!')
def predict( network, path_model, args ):
# load model
if network.load( args.path_model ) is not True:
return 0
data = IMaterialistImageDataset(
pathname=args.data,
ext='jpg',
num_channels=network.num_input_channels,
transform=get_transforms_det(network.size_input)
)
dataloader = DataLoader(test_data, batch_size=args.batch_size, shuffle=False, num_workers=args.workers )
dataloaders.append(dataloader)
# print neural net class
print('NeuralNet: {}'.format(datetime.datetime.now()), flush=True )
#print(network)
Ids, Yhat = network.predict( data )
return Id, pred
def main():
parser = arg_parser();
args = parser.parse_args();
# Configuration
project = args.project
projectname = args.projectname
pathnamedataset = args.pathdataset
pathnamemodel = args.model
pathproject = os.path.join( project, projectname )
pathnameout = args.pathnameout
filename = args.filename
pathname = os.path.expanduser( '~/.datasets' )
num_workers=0
no_cuda=False
parallel=False
gpu=0
seed=1
# experiments
experiments = [
{ 'name': 'ferp', 'subset': FactoryDataset.test, 'real': True },
{ 'name': 'affectnet', 'subset': FactoryDataset.validation, 'real': True },
{ 'name': 'ck', 'subset': FactoryDataset.training, 'real': True },
{ 'name': 'jaffe', 'subset': FactoryDataset.training, 'real': True },
{ 'name': 'bu3dfe', 'subset': FactoryDataset.training, 'real': True },
]
# Load models
network = NeuralNetClassifier(
patchproject=project,
nameproject=projectname,
no_cuda=no_cuda,
seed=seed,
gpu=gpu
)
cudnn.benchmark = True
# load model
if network.load( pathnamemodel ) is not True:
print('>>Error!!! load model')
assert(False)
size_input = network.size_input
tuplas=[]
for i, experiment in enumerate(experiments):
name_dataset = experiment['name']
subset = experiment['subset']
breal = experiment['real']
dataset = []
# real dataset
dataset = Dataset(
data=FactoryDataset.factory(
pathname=pathnamedataset,
name=name_dataset,
subset=subset,
#idenselect=idenselect,
download=True
),
num_channels=3,
transform=get_transforms_det( network.size_input ),
)
dataloader = DataLoader(dataset, batch_size=100, shuffle=False, num_workers=num_workers )
Yhat, Y = network.test( dataloader )
df = pd.DataFrame( np.concatenate((Yhat, Y), axis=1) )
df.to_csv( os.path.join(pathproject , '{}_{}_{}_dp.csv'.format(subset,projectname,name_dataset)), index=False, encoding='utf-8')
yhat = np.argmax( Yhat, axis=1 )
y = Y
acc = metrics.accuracy_score(y, yhat)
precision = metrics.precision_score(y, yhat, average='macro')
recall = metrics.recall_score(y, yhat, average='macro')
f1_score = 2*precision*recall/(precision+recall)
#|Name|Dataset|Cls|Acc| ...
tupla = {
#'Name':projectname,
'Dataset': '{}({})'.format( name_dataset, subset ),
'Accuracy': acc*100,
'Precision': precision*100,
'Recall': recall*100,
'F1 score': f1_score*100,
}
tuplas.append(tupla)
df = pd.DataFrame(tuplas)
df.to_csv( os.path.join( pathproject, 'experiments_cls.csv' ) , index=False, encoding='utf-8')
print('save experiments class ...')
print('DONE!!!')
def main():
# parameters
parser = arg_parser();
args = parser.parse_args();
random.seed(0)
print('Baseline clasification {}!!!'.format(datetime.datetime.now()))
print('\nArgs:')
[ print('\t* {}: {}'.format(k,v) ) for k,v in vars(args).items() ]
print('')
network = NeuralNetClassifier(
patchproject=args.project,
nameproject=args.name,
no_cuda=args.no_cuda,
parallel=args.parallel,
seed=args.seed,
print_freq=args.print_freq,
gpu=args.gpu
)
network.create(
arch=args.arch,
num_output_channels=args.num_classes,
num_input_channels=args.channels,
loss=args.loss,
lr=args.lr,
momentum=args.momentum,
optimizer=args.opt,
lrsch=args.scheduler,
pretrained=args.finetuning,
th=0.5,
size_input=args.image_size,
)
cudnn.benchmark = True
# resume model
if args.resume:
network.resume( os.path.join(network.pathmodels, args.resume ) )
# print neural net class
print('Load model: ')
print(network)
# datasets
# training dataset
train_data_kaggle = ATLASDataset(
path=args.data,
train=True,
folders_images='train', #train, cloud/train_full
metadata='train.csv',
ext='png', #png, tif
#count=20000,
num_channels=args.channels,
transform=get_transforms_aug( network.size_input ), #get_transforms_aug
)
# train_data_external = ATLASDataset(
# path=args.data,
# train=True,
# folders_images='train_external',
# metadata='train_external.csv',
# #count=20000,
# num_channels=args.channels,
# transform=get_transforms_aug( network.size_input ), #get_transforms_aug
# )
# train_data = torch.utils.data.ConcatDataset( [train_data_kaggle, train_data_external] )
train_data = train_data_kaggle
frec = np.array([ x for x in train_data_kaggle.data.data['Target'] ]).sum(axis=0)
weights = 1 / frec
# target_external = train_data_external.data.data['Target']
target_kaggle = train_data_kaggle.data.data['Target']
# target = np.concatenate( ( target_kaggle, target_external ) )
target = target_kaggle
samples_weights = np.array([ weights[ np.array(x).astype( np.uint8 ) ].max() for x in target ])
num_train = len(train_data)
#sampler = SubsetRandomSampler(np.random.permutation( num_train ) )
sampler = WeightedRandomSampler( weights=samples_weights, num_samples=len(samples_weights) , replacement=True )
train_loader = DataLoader(train_data, batch_size=args.batch_size,
sampler=sampler, num_workers=args.workers, pin_memory=network.cuda, drop_last=True)
# validate dataset
val_data = ATLASDataset(
path=args.data,
train=True,
folders_images='train', #train, cloud/train_full
metadata='train.csv',
ext='png', #png, tif
count=5000,
num_channels=args.channels,
transform=get_transforms_det( network.size_input ),
)
num_val = len(val_data)
val_loader = DataLoader(val_data, batch_size=80,
shuffle=False, num_workers=args.workers, pin_memory=network.cuda, drop_last=False)
print('Load datset')
print('Train: ', len(train_data))
print('Val: ', len(val_data))
# training neural net
network.fit( train_loader, val_loader, args.epochs, args.snapshot )
print("Optimization Finished!")
print("DONE!!!")
def main():
# parameters
parser = arg_parser()
args = parser.parse_args()
random.seed(0)
print('Baseline clasification {}!!!'.format(datetime.datetime.now()))
print('\nArgs:')
[print('\t* {}: {}'.format(k, v)) for k, v in vars(args).items()]
print('')
network = NeuralNetClassifier(patchproject=args.project,
nameproject=args.name,
no_cuda=args.no_cuda,
parallel=args.parallel,
seed=args.seed,
print_freq=args.print_freq,
gpu=args.gpu)
network.create(
arch=args.arch,
num_output_channels=args.num_classes,
num_input_channels=args.channels,
loss=args.loss,
lr=args.lr,
momentum=args.momentum,
optimizer=args.opt,
lrsch=args.scheduler,
pretrained=args.finetuning,
topk=(
1,
2,
),
)
cudnn.benchmark = True
# resume model
if args.resume:
network.resume(os.path.join(network.pathmodels, args.resume))
# print neural net class
print('Load model: ')
print(network)
# datasets
# training dataset
train_data = Dataset(
data=FactoryDataset.factory(pathname=args.data,
name=args.name_dataset,
subset=FactoryDataset.training,
download=True),
#count=20000,
num_channels=network.num_input_channels,
transform=get_transforms_aug(network.size_input),
)
num_train = len(train_data)
sampler = SubsetRandomSampler(np.random.permutation(num_train))
train_loader = DataLoader(train_data,
batch_size=args.batch_size,
sampler=sampler,
num_workers=args.workers,
pin_memory=network.cuda,
drop_last=True)
# validate dataset
val_data = Dataset(
data=FactoryDataset.factory(pathname=args.data,
name=args.name_dataset,
subset=FactoryDataset.validation,
download=True),
num_channels=network.num_input_channels,
transform=get_transforms_det(network.size_input),
)
num_val = len(val_data)
val_loader = DataLoader(val_data,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=network.cuda,
drop_last=False)
print('Load datset')
print('Train: ', len(train_data))
print('Val: ', len(val_data))
# training neural net
network.fit(train_loader, val_loader, args.epochs, args.snapshot)
print("Optimization Finished!")
print("DONE!!!")