def run_test(path_model, path_csv, path_output):
model_name = 'SENET154'
batch_size = 16
img_size = 256
crop_size = 224
target_mean = np.array([0.485, 0.456, 0.406])
target_std = np.array([0.229, 0.224, 0.225])
data_transform = DataTransform(no_bg=True, pad=True)
data_transform_valid = data_transform.get_test(img_size=img_size,
crop_size=crop_size,
target_mean=target_mean,
target_std=target_std,
positions=[0, 2, 9])
device = torch.device("cuda:0")
output_lst = predict.predict(path_csv=path_csv,
path_model=path_model,
model_name=model_name,
batch_size=batch_size,
device=device,
transform=data_transform_valid)
pandas.DataFrame(output_lst).to_csv(path_output, header=False, index=False)
def run_test(path_model, path_csv, path_output):
model_name = 'DENSENET161-LARGE3'
batch_size = 16
img_size = 366
crop_size = 320
target_mean = 0.456
target_std = 0.225
data_transform = DataTransform(no_bg=True, pad=True)
data_transform_valid = data_transform.get_test(img_size=img_size,
crop_size=crop_size,
target_mean=target_mean,
target_std=target_std,
positions=[6, 3, 4])
device = torch.device("cuda:0")
output_lst = predict.predict(path_csv=path_csv,
path_model=path_model,
model_name=model_name,
batch_size=batch_size,
device=device,
transform=data_transform_valid)
pandas.DataFrame(output_lst).to_csv(path_output, header=False, index=False)
def run_test(path_model, path_csv, path_output):
model_name = 'VGG16-BN'
batch_size = 16
img_size = 256
crop_size = 224
target_mean = 0.0
target_std = 1.0
data_transform = DataTransform(no_bg=True, pad=True)
data_transform_valid = data_transform.get_test(img_size=img_size,
crop_size=crop_size,
target_mean=target_mean,
target_std=target_std,
positions=[5, 6, 4])
device = torch.device("cuda:0")
output_lst = predict.predict(path_csv=path_csv,
path_model=path_model,
model_name=model_name,
batch_size=batch_size,
device=device,
transform=data_transform_valid)
pandas.DataFrame(output_lst).to_csv(path_output, header=False, index=False)
def run_test(path_model):
model_name = 'PNASNET'
model_pretrained = True
path_data = '../../MURA-v1.0/'
path_root = '../../'
batch_size = 6
img_size = 354
crop_size = 331
target_mean = 0.5
target_std = 0.5
data_transform = DataTransform(no_bg=True, pad=True)
data_transform_valid = data_transform.get_valid(img_size=img_size,
crop_size=crop_size,
target_mean=target_mean,
target_std=target_std)
device = None
opts, _ = getopt.getopt(sys.argv[1:], "d:", ["device="])
for opt, arg in opts:
if opt in ("-d", "--device") and torch.cuda.is_available():
device = torch.device("cuda:" + str(arg))
if device is None:
print("GPU not found! Using CPU!")
device = torch.device("cpu")
print('NN architecture = ', model_name)
print("using data transforms: " + str(data_transform))
if os.path.exists(path_model + "-L.pth.tar"):
print('Testing the model with best valid-loss')
train_multiview.test(path_data=path_data,
path_root=path_root,
path_model=path_model + "-L",
model_name=model_name,
model_pretrained=model_pretrained,
batch_size=batch_size,
device=device,
transform=data_transform_valid)
if os.path.exists(path_model + "-A.pth.tar"):
print('Testing the model with best valid-auroc')
train_multiview.test(path_data=path_data,
path_root=path_root,
path_model=path_model + "-A",
model_name=model_name,
model_pretrained=model_pretrained,
batch_size=batch_size,
device=device,
transform=data_transform_valid)
def run_test(path_model):
model_name = 'DENSENET201'
model_pretrained = True
path_data_valid = '../../MURA_valid1_keras'
batch_size = 16
img_size = 256
crop_size = 224
target_mean = 0.0
target_std = 1.0
data_transform = DataTransform(revised=False,
no_bg=True,
pad=True,
to_rgb=False)
data_transform_valid = data_transform.get_valid(img_size=img_size,
crop_size=crop_size,
target_mean=target_mean,
target_std=target_std)
device = None
opts, _ = getopt.getopt(sys.argv[1:], "d:", ["device="])
for opt, arg in opts:
if opt in ("-d", "--device") and torch.cuda.is_available():
device = torch.device("cuda:" + str(arg))
if device is None:
print("GPU not found! Using CPU!")
device = torch.device("cpu")
print('NN architecture = ', model_name)
print("using data transforms: " + str(data_transform))
print('Testing the model with best valid-loss')
train.test(path_data=path_data_valid,
path_model=path_model + "-L",
model_name=model_name,
model_pretrained=model_pretrained,
batch_size=batch_size,
device=device,
transform=data_transform_valid)
print('Testing the model with best valid-auroc')
train.test(path_data=path_data_valid,
path_model=path_model + "-A",
model_name=model_name,
model_pretrained=model_pretrained,
batch_size=batch_size,
device=device,
transform=data_transform_valid)
def run_test(path_model, path_csv, path_output):
model_name = 'DUALPATHNET107_5k'
batch_size = 16
img_size = 256
crop_size = 224
target_mean = np.array([124 / 255, 117 / 255, 104 / 255])
target_std = 1 / (.0167 * 255)
data_transform = DataTransform(no_bg=True, pad=True)
data_transform_valid = data_transform.get_test(img_size=img_size, crop_size=crop_size, target_mean=target_mean,
target_std=target_std, positions=[0, 1, 8])
device = torch.device("cuda:0")
output_lst = predict.predict(
path_csv=path_csv,
path_model=path_model,
model_name=model_name,
batch_size=batch_size,
device=device,
transform=data_transform_valid
)
pandas.DataFrame(output_lst).to_csv(path_output, header=False, index=False)
def run_test(path_model, path_csv, path_output):
model_name = 'INCEPTIONV4-LARGE'
batch_size = 16
img_size = 378
crop_size = 331
target_mean = 0.5
target_std = 0.5
data_transform = DataTransform(no_bg=True, pad=True)
data_transform_valid = data_transform.get_test(img_size=img_size, crop_size=crop_size, target_mean=target_mean,
target_std=target_std, positions=[1, 2, 9])
device = torch.device("cuda:0")
output_lst = predict.predict(
path_csv=path_csv,
path_model=path_model,
model_name=model_name,
batch_size=batch_size,
device=device,
transform=data_transform_valid
)
pandas.DataFrame(output_lst).to_csv(path_output, header=False, index=False)
def main():
configs = [
Config(model_name="DUALPATHNET107_5k",
file_model="m-03916-09049",
batch_size=32,
img_size=256,
crop_size=224,
target_mean=np.array([124 / 255, 117 / 255, 104 / 255]),
target_std=1 / (.0167 * 255),
positions=[0, 1, 8],
data_transform=DataTransform(no_bg=True, pad=True)),
Config(model_name="SENET154",
file_model="m-03924-09074",
batch_size=32,
img_size=256,
crop_size=224,
target_mean=np.array([0.485, 0.456, 0.406]),
target_std=np.array([0.229, 0.224, 0.225]),
positions=[0, 2, 9],
data_transform=DataTransform(no_bg=True, pad=True)),
Config(model_name="INCEPTIONV4-LARGE",
file_model="m-03879-09080",
batch_size=32,
img_size=378,
crop_size=331,
target_mean=0.5,
target_std=0.5,
positions=[1, 2, 9],
data_transform=DataTransform(no_bg=True, pad=True)),
Config(model_name="VGG16-BN",
file_model="m-04051-09053",
batch_size=48,
img_size=256,
crop_size=224,
target_mean=0.0,
target_std=1.0,
positions=[5, 6, 4],
data_transform=DataTransform(no_bg=True, pad=True)),
Config(model_name="DENSENET201-LARGE3",
file_model="m-04001-09065",
batch_size=48,
img_size=366,
crop_size=320,
target_mean=0.456,
target_std=0.225,
positions=[5, 7, 3],
data_transform=DataTransform(no_bg=True, pad=True)),
Config(model_name="DENSENET161-LARGE3",
file_model="m-03988-09084",
batch_size=32,
img_size=366,
crop_size=320,
target_mean=0.456,
target_std=0.225,
positions=[6, 3, 4],
data_transform=DataTransform(no_bg=True, pad=True)),
Config(model_name="NASNETALARGE",
file_model="m-03815-09099",
batch_size=8,
img_size=354,
crop_size=331,
target_mean=0.5,
target_std=0.5,
positions=[7, 8, 4],
data_transform=DataTransform(no_bg=True, pad=True))
]
keys, results = [], []
for config in configs:
print(config.model_name)
keys, result = run_test(sys.argv[1], config)
results.append(result)
results = np.concatenate(results, axis=1)
score = np.mean(results, axis=1)
label = np.array(score >= 0.5, dtype=np.int32)
pandas.DataFrame(label, index=keys).to_csv(sys.argv[2], header=False)
def run_train():
timestamp = time.strftime("%Y%m%d") + '-' + time.strftime("%H%M%S")
model_name = 'INCEPTIONV4-LARGE'
model_pretrained = True
path_data_train = '../../MURA_trainval_keras'
path_data_valid = '../../MURA_valid1_keras'
path_log = '../../trained_models/' + timestamp + '/tb'
batch_size = 16
epoch_num = 80
img_size = 331
crop_size = 331
target_mean = 0.5
target_std = 0.5
path_model = '../../trained_models/' + timestamp + '/m-' + timestamp
data_transform = DataTransform(aug="rot30",
no_bg=True,
pad=True,
no_crop=True)
data_transform_train = data_transform.get_train(img_size=img_size,
crop_size=crop_size,
target_mean=target_mean,
target_std=target_std)
data_transform_valid = data_transform.get_valid(img_size=img_size,
crop_size=crop_size,
target_mean=target_mean,
target_std=target_std)
device = None
opts, _ = getopt.getopt(sys.argv[1:], "d:", ["device="])
for opt, arg in opts:
if opt in ("-d", "--device") and torch.cuda.is_available():
device = torch.device("cuda:" + str(arg))
if device is None:
print("GPU not found! Using CPU!")
device = torch.device("cpu")
print('Training NN architecture = ', model_name)
train.train(path_data_train=path_data_train,
path_data_valid=path_data_valid,
path_log=path_log,
path_model=path_model,
model_name=model_name,
model_pretrained=model_pretrained,
batch_size=batch_size,
epoch_num=epoch_num,
checkpoint=None,
device=device,
transform_train=data_transform_train,
transform_valid=data_transform_valid,
optimizer_fn=optimizers.adam_optimizers)
print('NN architecture = ', model_name)
print("using data transforms: " + str(data_transform))
print('Testing the model with best valid-loss')
print('timestamp = ' + timestamp)
train.test(path_data=path_data_valid,
path_model=path_model + "-L",
model_name=model_name,
model_pretrained=model_pretrained,
batch_size=batch_size,
device=device,
transform=data_transform_valid)
print('Testing the model with best valid-auroc')
print('timestamp = ' + timestamp)
train.test(path_data=path_data_valid,
path_model=path_model + "-A",
model_name=model_name,
model_pretrained=model_pretrained,
batch_size=batch_size,
device=device,
transform=data_transform_valid)
def run_train():
timestamp = time.strftime("%Y%m%d") + '-' + time.strftime("%H%M%S")
model_name = 'INCEPTIONRESNETV2'
model_pretrained = True
path_data = '../../MURA-v1.0/'
path_root = '../../'
path_log = '../../trained_models/' + timestamp + '/tb'
batch_size = 24
epoch_num = 15
img_size = 299
crop_size = 299
target_mean = 0.5
target_std = 0.5
checkpoint = "../../trained_models/04037-09061-inceptionresnetv2-adam-nobgpad/m-20180518-015825-L.pth.tar"
path_model = '../../trained_models/' + timestamp + '/m-' + timestamp
data_transform = DataTransform(no_bg=True, pad=True)
data_transform_train = data_transform.get_train(img_size=img_size,
crop_size=crop_size,
target_mean=target_mean,
target_std=target_std)
data_transform_valid = data_transform.get_valid(img_size=img_size,
crop_size=crop_size,
target_mean=target_mean,
target_std=target_std)
device = None
opts, _ = getopt.getopt(sys.argv[1:], "d:", ["device="])
for opt, arg in opts:
if opt in ("-d", "--device") and torch.cuda.is_available():
device = torch.device("cuda:" + str(arg))
if device is None:
print("GPU not found! Using CPU!")
device = torch.device("cpu")
print('Training NN architecture = ', model_name)
print('Training Multiview')
train_multiview.train(path_data=path_data,
path_root=path_root,
path_log=path_log,
path_model=path_model,
model_name=model_name,
model_pretrained=model_pretrained,
batch_size=batch_size,
epoch_num=epoch_num,
checkpoint=checkpoint,
device=device,
transform_train=data_transform_train,
transform_valid=data_transform_valid,
optimizer_fn=optimizers.adam_optimizers_small)
print('NN architecture = ', model_name)
print("using data transforms: " + str(data_transform))
print('Testing the model with best valid-loss')
print('timestamp = ' + timestamp)
train_multiview.test(path_data=path_data,
path_root=path_root,
path_model=path_model + "-L",
model_name=model_name,
model_pretrained=model_pretrained,
batch_size=batch_size,
device=device,
transform=data_transform_valid)
print('Testing the model with best valid-auroc')
print('timestamp = ' + timestamp)
train_multiview.test(path_data=path_data,
path_root=path_root,
path_model=path_model + "-A",
model_name=model_name,
model_pretrained=model_pretrained,
batch_size=batch_size,
device=device,
transform=data_transform_valid)
def main():
configs = [
Config(model_name="SENET154",
file_model="senet154",
batch_size=32,
img_size=256,
crop_size=224,
target_mean=np.array([0.485, 0.456, 0.406]),
target_std=np.array([0.229, 0.224, 0.225]),
data_transform=DataTransform(no_bg=True, pad=True),
positions=[0, 6, 7, 3, 4]),
Config(model_name="SENET154-LARGE",
file_model="senet154large",
batch_size=24,
img_size=293,
crop_size=256,
target_mean=np.array([0.485, 0.456, 0.406]),
target_std=np.array([0.229, 0.224, 0.225]),
data_transform=DataTransform(no_bg=True, pad=True),
positions=[5, 1, 2, 3, 9]),
Config(model_name="INCEPTIONV4",
file_model="inceptionv4",
batch_size=32,
img_size=341,
crop_size=299,
target_mean=0.5,
target_std=0.5,
data_transform=DataTransform(no_bg=True, pad=True),
positions=[5, 6, 2, 8, 4]),
Config(model_name="INCEPTIONV4-LARGE",
file_model="inceptionv4large",
batch_size=16,
img_size=378,
crop_size=331,
target_mean=0.5,
target_std=0.5,
data_transform=DataTransform(no_bg=True, pad=True),
positions=[0, 6, 2, 3, 9]),
Config(model_name="INCEPTIONRESNETV2",
file_model="inceptionresnetv2",
batch_size=8,
img_size=341,
crop_size=299,
target_mean=0.5,
target_std=0.5,
data_transform=DataTransform(no_bg=True, pad=True),
positions=[0, 1, 7, 8, 9]),
Config(model_name="DENSENET201-LARGE3",
file_model="densenet201",
batch_size=24,
img_size=366,
crop_size=320,
target_mean=0.456,
target_std=0.225,
data_transform=DataTransform(no_bg=True, pad=True),
positions=[5, 1, 7, 8, 4]),
Config(model_name="DENSENET161-LARGE3",
file_model="densenet161",
batch_size=16,
img_size=366,
crop_size=320,
target_mean=0.456,
target_std=0.225,
data_transform=DataTransform(no_bg=True, pad=True),
positions=[0, 6, 7, 8, 4]),
Config(model_name="DENSENET169-LARGE3",
file_model="densenet169",
batch_size=24,
img_size=366,
crop_size=320,
target_mean=0.456,
target_std=0.225,
data_transform=DataTransform(no_bg=True, pad=True),
positions=[5, 6, 2, 3, 4]),
Config(model_name="NASNETALARGE",
file_model="nasnetalarge",
batch_size=8,
img_size=354,
crop_size=331,
target_mean=0.5,
target_std=0.5,
data_transform=DataTransform(no_bg=True, pad=True),
positions=[5, 1, 2, 8, 9]),
Config(model_name="PNASNET",
file_model="pnasnet",
batch_size=8,
img_size=354,
crop_size=331,
target_mean=0.5,
target_std=0.5,
data_transform=DataTransform(no_bg=True, pad=True),
positions=[0, 1, 7, 3, 9])
]
keys, results = [], []
for config in configs:
print(config.model_name)
keys, result = run_test(sys.argv[1], config)
results.append(result)
results = np.concatenate(results, axis=1)
score = np.mean(results, axis=1)
label = np.array(score >= 0.5, dtype=np.int32)
pandas.DataFrame(label, index=keys).to_csv(sys.argv[2], header=False)
import yaml
import numpy as np
from data_augmentation import DataTransform
if __name__ == "__main__":
config_train = {
"transform":
DataTransform(no_bg=True,
pad=True,
aug_rotate=np.random.uniform(15.0, 30.0),
aug_shear=np.random.uniform(0.0, 10.0),
flip_h="random",
crop_mode="random",
random_crop_factor=1.0 / np.random.uniform(7.5, 15.0)),
"batch_size":
16,
"optimizer_name":
"adam",
"learning_rate":
1e-6 * (0.01 / 1e-6)**(26 / 80),
"differential_lr":
1,
"weight_decay":
np.power(10.0, np.random.uniform(-7.0, -4.0)).item(),
"is_nesterov":
False,
"beta1":
0.9,
"beta2":
0.999,
def run_train():
timestamp = time.strftime("%Y%m%d") + '-' + time.strftime("%H%M%S")
model_name = 'SENET154-LARGE'
model_pretrained = True
path_data = '../../MURA-v1.0/'
path_root = '../../'
path_log = '../../trained_models/' + timestamp + '/tb'
batch_size = 16
epoch_num = 10
img_size = 256
crop_size = 256
target_mean = np.array([0.485, 0.456, 0.406])
target_std = np.array([0.229, 0.224, 0.225])
checkpoint = "../../trained_models/03926-09052-senet154large-adam-nobgpad/m-20180626-081034-L.pth.tar"
path_model = '../../trained_models/' + timestamp + '/m-' + timestamp
data_transform = DataTransform(no_bg=True, pad=True)
data_transform_train = data_transform.get_train(img_size=img_size,
crop_size=crop_size,
target_mean=target_mean,
target_std=target_std)
data_transform_valid = data_transform.get_valid(img_size=img_size,
crop_size=crop_size,
target_mean=target_mean,
target_std=target_std)
device = None
opts, _ = getopt.getopt(sys.argv[1:], "d:", ["device="])
for opt, arg in opts:
if opt in ("-d", "--device") and torch.cuda.is_available():
device = torch.device("cuda:" + str(arg))
if device is None:
print("GPU not found! Using CPU!")
device = torch.device("cpu")
print('Training NN architecture = ', model_name)
print('Training Multiview')
train_multiview.train(path_data=path_data,
path_root=path_root,
path_log=path_log,
path_model=path_model,
model_name=model_name,
model_pretrained=model_pretrained,
batch_size=batch_size,
epoch_num=epoch_num,
checkpoint=checkpoint,
device=device,
transform_train=data_transform_train,
transform_valid=data_transform_valid,
optimizer_fn=optimizers.adam_optimizers_small)
print('NN architecture = ', model_name)
print("using data transforms: " + str(data_transform))
print('Testing the model with best valid-loss')
print('timestamp = ' + timestamp)
train_multiview.test(path_data=path_data,
path_root=path_root,
path_model=path_model + "-L",
model_name=model_name,
model_pretrained=model_pretrained,
batch_size=batch_size,
device=device,
transform=data_transform_valid)
print('Testing the model with best valid-auroc')
print('timestamp = ' + timestamp)
train_multiview.test(path_data=path_data,
path_root=path_root,
path_model=path_model + "-A",
model_name=model_name,
model_pretrained=model_pretrained,
batch_size=batch_size,
device=device,
transform=data_transform_valid)
import yaml
import numpy as np
from data_augmentation import DataTransform
if __name__ == "__main__":
config_train = {
"transform": DataTransform(
no_bg=True,
pad=True,
aug_rotate=20,
aug_shear=10,
flip_h="random",
crop_mode="random",
random_crop_factor=0.08
),
"batch_size": 20,
"optimizer_name": "adam",
"differential_lr": 10,
"is_nesterov": False,
"beta1": 0.9,
"beta2": 0.999,
"epoch_num": 80
}
config_valid = {
"model_name": "InceptionV4",
"img_size": 378,
"crop_size": 331
}
import yaml
from data_augmentation import DataTransform
if __name__ == "__main__":
config = {
"model_name":
"NASNet",
"img_size":
354,
"crop_size":
331,
"transform":
DataTransform(no_bg=True,
pad=True,
crop_mode="ten",
ten_crop_positions=[7, 8, 4],
normalize=True),
"batch_size":
8,
"path_model":
"models/NASNet.pt"
}
with open("test_configs/config07.yaml", "w") as f:
yaml.dump(config, f, default_flow_style=False)