def valid_dataloader(
train_meta_data,
input_size=128,
batch_size=64,
num_workers=4,
label_matrix=None,
transforms_i=None,
):
image_dir = os.path.join(DATASET_PATH, 'train', 'train_data', 'images')
train_label_path = os.path.join(DATASET_PATH, 'train', 'train_label')
train_meta_path = os.path.join(DATASET_PATH, 'train', 'train_data',
'train_with_valid_tags.csv')
if transforms_i is None:
transforms_i = transforms.Compose([
transforms.Resize((input_size, input_size)),
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(20),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
# transforms - batch normalization add
dataloader = DataLoader(AIRushDataset(image_dir,
train_meta_data,
label_path=train_label_path,
transform=transforms_i,
label_matrix=label_matrix),
batch_size=batch_size,
shuffle=True,
num_workers=num_workers,
pin_memory=True)
return dataloader
def infer(test_image_data_path, test_meta_data_path):
# DONOTCHANGE This Line
test_meta_data = pd.read_csv(test_meta_data_path, delimiter=',', header=0)
input_size=128 # you can change this according to your model.
batch_size=200 # you can change this. But when you use 'nsml submit --test' for test infer, there are only 200 number of data.
device = 0
dataloader = DataLoader(
AIRushDataset(test_image_data_path, test_meta_data, label_path=None,
transform=transforms.Compose([transforms.Resize((input_size, input_size)), transforms.ToTensor()])),
batch_size=batch_size,
shuffle=False,
num_workers=0,
pin_memory=True)
model_nsml.to(device)
model_nsml.eval()
predict_list = []
with torch.no_grad(): # No need for torch.backward() - no gradient calculation
for batch_idx, image in enumerate(dataloader):
image = image.to(device)
output = model_nsml(image).double()
output_prob = F.softmax(output, dim=1)
predict = np.argmax(to_np(output_prob), axis=1)
predict_list.append(predict)
predict_vector = np.concatenate(predict_list, axis=0)
return predict_vector # this return type should be a numpy array which has shape of (138343, 1)
def infer(test_image_data_path, test_meta_data_path):
# DONOTCHANGE This Line
test_meta_data = pd.read_csv(test_meta_data_path,
delimiter=',',
header=0)
input_size = 224 # you can change this according to your model.
batch_size = _BATCH_SIZE // 4 # you can change this. But when you use 'nsml submit --test' for test infer, there are only 200 number of data.
# test time gpu memory error, so i reduce batch size when test time
device = 0
dataloader = DataLoader(
AIRushDataset(
test_image_data_path,
test_meta_data,
label=None,
transform=transforms.Compose([
ObjROI(),
transforms.Resize((input_size, input_size)),
transforms.ToTensor()
#,transforms.Normalize(mean=mean_v, std=std_v)
])),
batch_size=batch_size,
shuffle=False,
num_workers=0,
pin_memory=True)
model_nsml.to(device)
model_nsml.eval()
predict_list = []
for batch_idx, image in enumerate(dataloader):
image = image.to(device)
image_lr = image.flip(3) #batch, ch, h, w
output_org = model_nsml(image).double()
output_lr = model_nsml(image_lr).double()
output = output_org + output_lr
output_prob = F.softmax(output, dim=1)
predict = np.argmax(to_np(output_prob), axis=1)
predict_list.append(predict)
predict_vector = np.concatenate(predict_list, axis=0)
return predict_vector # this return type should be a numpy array which has shape of (138343, 1)
def infer(test_image_data_path, test_meta_data_path):
# DONOTCHANGE This Line
test_meta_data = pd.read_csv(test_meta_data_path, delimiter=',', header=0)
input_size=224 # you can change this according to your model.
batch_size=200 # you can change this. But when you use 'nsml submit --test' for test infer, there are only 200 number of data.
device = 0
we = 0.25
ensemble = [['team_62/airush1/320', '02'],['team_62/airush1/320','12'],['team_62/airush1/320','22'],['team_62/airush1/320','32']]
#ensemble = [['team_62/airush1/415', '03'],['team_62/airush1/415','13'],['team_62/airush1/415','23'],['team_62/airush1/415','33']]
predict_list = []
for i in range(4):
dataloader = DataLoader(
AIRushDataset(test_image_data_path, test_meta_data, label_path=None,
transform=transforms.Compose([transforms.Resize((input_size, input_size)), transforms.RandomRotation(20),transforms.ToTensor(),transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])),
batch_size=batch_size,
shuffle=False,
num_workers=4,
pin_memory=True)
# Let's do ensemble!!!
nsml.load(checkpoint=str(ensemble[i][1]),session=str(ensemble[i][0]))
# model load
model_nsml.to(device)
model_nsml.eval()
predict_output_list = []
with torch.no_grad():
for batch_idx, image in enumerate(dataloader):
image = image.to(device)
output = model_nsml(image).double()
output_prob = to_np(F.softmax(output, dim=1))
predict_output_list.append(output_prob * we)
predict_output_list = np.concatenate(predict_output_list,axis=0)
predict_list.append(predict_output_list)
predict_vector = np.argmax(np.sum(predict_list,axis=0), axis=1)
return predict_vector # this return type should be a numpy array which has shape of (138343)
def infer(test_image_data_path, test_meta_data_path):
# DONOTCHANGE This Line
test_meta_data = pd.read_csv(test_meta_data_path,
delimiter=',',
header=0)
batch_size = infer_batch_size # you can change this. But when you use 'nsml submit --test' for test infer, there are only 200 number of data.
device = 0
dataloader = DataLoader(AIRushDataset(
test_image_data_path,
test_meta_data,
label_path=None,
transform=transforms.Compose(infer_transform_list)),
batch_size=batch_size,
shuffle=False,
num_workers=0,
pin_memory=True)
model_nsml.to(device)
model_nsml.eval()
predict_list = []
for batch_idx, image in enumerate(dataloader):
image = image.to(device)
output = model_nsml(image).double()
output_prob = output
if args.loss_type == "cross_entropy":
output_prob = F.softmax(output, dim=1)
elif args.loss_type == "bce":
output_prob = torch.sigmoid(output)
predict = np.argmax(to_np(output_prob), axis=1)
predict_list.append(predict)
predict_vector = np.concatenate(predict_list, axis=0)
return predict_vector # this return type should be a numpy array which has shape of (138343)
def infer(test_image_data_path, test_meta_data_path):
# DONOTCHANGE This Line
test_meta_data = pd.read_csv(test_meta_data_path,
delimiter=',',
header=0)
device = 0
models = args.models.split(",")
model_weights = [float(w) for w in args.model_weights.split(",")]
nsml_sessionss = args.nsml_sessionss.split(",")
nsml_checkpoints = args.nsml_checkpoints.split(",")
loss_types = args.loss_types.split(",")
transform_random_crop = args.transform_random_crop.split(",")
transform_random_sized_crop = args.transform_random_sized_crop.split(
",")
transform_norm = args.transform_norm.split(",")
infer_transform_center_crop = args.infer_transform_center_crop.split(
",")
total_output_probs = None
for i, model_name in enumerate(models):
batch_size = batch_size_map[model_name] // 2
infer_transform_list = []
if infer_transform_center_crop[i] == "True":
infer_transform_list.append(transforms.Resize((248, 248)))
infer_transform_list.append(
transforms.CenterCrop((args.input_size, args.input_size)))
infer_transform_list.append(transforms.ToTensor())
if transform_norm[i] == "True":
infer_transform_list.append(
transforms.Normalize(
[0.44097832, 0.44847423, 0.42528335],
[0.25748107, 0.26744914, 0.30532702]))
else:
if transform_random_crop[i] == "True":
infer_transform_list.append(transforms.Resize((256, 256)))
infer_transform_list.append(
transforms.CenterCrop(
(args.input_size, args.input_size)))
elif transform_random_sized_crop[i] == "True":
infer_transform_list.append(transforms.Resize((256, 256)))
infer_transform_list.append(
transforms.CenterCrop(
(args.input_size, args.input_size)))
else:
infer_transform_list.append(
transforms.Resize((args.input_size, args.input_size)))
infer_transform_list.append(transforms.ToTensor())
if transform_norm[i] == "True":
infer_transform_list.append(
transforms.Normalize(
[0.44097832, 0.44847423, 0.42528335],
[0.25748107, 0.26744914, 0.30532702]))
print("transform", infer_transform_list)
dataloader = DataLoader(
AIRushDataset(
test_image_data_path,
test_meta_data,
label_path=None,
transform=transforms.Compose(infer_transform_list)
), #[transforms.Resize((args.input_size, args.input_size)), transforms.ToTensor()])),
batch_size=batch_size,
shuffle=False,
num_workers=0,
pin_memory=True)
if model_name == "Resnet18":
model = Resnet18(args.output_size)
elif model_name == "Resnet152":
model = Resnet152(args.output_size)
elif model_name == "baseline":
model = Baseline(args.hidden_size, args.output_size)
elif model_name.split("-")[0] == "efficientnet":
model = EfficientNet.from_pretrained(args.model,
args.output_size)
else:
raise Exception("model type is invalid : " + args.model)
model.to(device)
def load_fn(dir_name):
save_state_path = os.path.join(dir_name, 'state_dict.pkl')
state = torch.load(save_state_path)
model.load_state_dict(state['model'])
print("model loaded", dir_name)
model.eval()
nsml.load(checkpoint=nsml_checkpoints[i],
load_fn=load_fn,
session="team_13/airush1/" + nsml_sessionss[i])
output_probs = None
for batch_idx, image in enumerate(dataloader):
image = image.to(device)
output = model(image).double()
if loss_types[i] == "cross_entropy":
output_prob = F.softmax(output, dim=1)
else:
output_prob = torch.sigmoid(output)
if output_probs is None:
output_probs = to_np(output_prob)
else:
output_probs = np.concatenate(
[output_probs, to_np(output_prob)], axis=0)
if total_output_probs is None:
total_output_probs = output_probs * model_weights[i]
else:
total_output_probs += (output_probs * model_weights[i])
predict = np.argmax(total_output_probs, axis=1)
return predict # this return type should be a numpy array which has shape of (138343)
def infer(test_image_data_path, test_meta_data_path):
# DONOTCHANGE This Line
test_meta_data = pd.read_csv(test_meta_data_path,
delimiter=',',
header=0)
# dropout ratio
ensemble0 = [['team_62/airush1/320', '02'],
['team_62/airush1/320', '12'],
['team_62/airush1/320', '22'],
['team_62/airush1/98', '4']] # effi
ensemble1 = [['team_62/airush1/415', '03'],
['team_62/airush1/415', '13'],
['team_62/airush1/415', '23'],
['team_62/airush1/415', '33']] # effi
ensemble2 = [['team_62/airush1/678', '02'],
['team_62/airush1/678', '12'],
['team_62/airush1/185',
'17']] #[['team_62/airush1/185','17']] # resnet50
ensemble3 = [['team_62/airush1/683', '02'],
['team_62/airush1/683',
'12']] # oct ['team_62/airush1/409','18']
#ensemble4 = [['team_62/airush1/605','8']] # SKNet # transforms 에서 normalize 반드시 뺄 것
input_size = 224 # you can change this according to your model.
batch_size = 512 # you can change this. But when you use 'nsml submit --test' for test infer, there are only 200 number of data.
device = 0
w0 = 0.125
w2 = 0.166
w3 = 0.25
predict_list = []
for i in range(4): # ensemble 개수
#print('i th inference')
dataloader = DataLoader(AIRushDataset(
test_image_data_path,
test_meta_data,
label_path=None,
transform=transforms.Compose([
transforms.Resize((input_size, input_size)),
transforms.RandomRotation(20),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
])),
batch_size=batch_size,
shuffle=False,
num_workers=4,
pin_memory=True)
# 9:10 결과보고 뺄지 말지 결정
# Let's do ensemble!!!
if (i == 0):
# 'efficientNet_b0 : ensemble 4 - fold'
for j in range(4):
model_name = 'efficientnet-b0'
model = EfficientNet.from_name(model_name)
bind_model(model)
nsml.load(checkpoint=str(ensemble0[j][1]),
session=str(ensemble0[j][0]))
model.to(device)
model.eval()
predict_output_list = []
with torch.no_grad():
for batch_idx, image in enumerate(dataloader):
image = image.to(device)
output = model(image).double()
output_prob = to_np(F.softmax(output, dim=1))
predict_output_list.append(output_prob * w0)
predict_output_list = np.concatenate(predict_output_list,
axis=0)
predict_list.append(predict_output_list)
elif (i == 1):
# resnet50
for j in range(3):
model = resnext50(
num_classes=args.output_size) # 모델에 맞게 수정
bind_model(model)
nsml.load(checkpoint=str(ensemble2[j][1]),
session=str(ensemble2[j][0])) # 모델에 맞게 수정
model.to(device)
model.eval()
predict_output_list = []
with torch.no_grad():
for batch_idx, image in enumerate(dataloader):
image = image.to(device)
output = model(image).double()
output_prob = to_np(F.softmax(output, dim=1))
#print(output_prob)
predict_output_list.append(output_prob * w2)
predict_output_list = np.concatenate(predict_output_list,
axis=0)
predict_list.append(predict_output_list)
#print('resnet model')
elif (i == 2):
# resnet50
for j in range(2):
model = OctResNet(
Bottleneck, [3, 4, 6, 3],
num_classes=args.output_size) # 모델에 맞게 수정
bind_model(model)
nsml.load(checkpoint=str(ensemble3[j][1]),
session=str(ensemble3[j][0])) # 모델에 맞게 수정
model.to(device)
model.eval()
predict_output_list = []
with torch.no_grad():
for batch_idx, image in enumerate(dataloader):
image = image.to(device)
output = model(image).double()
output_prob = to_np(F.softmax(output, dim=1))
#print(output_prob)
predict_output_list.append(output_prob * w3) # 수정
predict_output_list = np.concatenate(predict_output_list,
axis=0)
predict_list.append(predict_output_list)
#print('resnet model')
# ensemble 추가
# 마지막 SENet 추가
predict_vector = np.argmax(np.sum(predict_list, axis=0), axis=1)
return predict_vector # this return type should be a numpy array which has shape of (138343)
def infer(test_image_data_path, test_meta_data_path):
print('ensemble dekitennzyane?')
# DONOTCHANGE This Line
test_meta_data = pd.read_csv(test_meta_data_path,
delimiter=',',
header=0)
input_size = 128 # you can change this according to your model.
input_size_resnet = 224
input_size_inception = 299
# you can change this. But when you use 'nsml submit --test' for test
# infer, there are only 200 number of data.
batch_size = 62
device = 0
# ここのローディングもいくつかバリエーション持たせてtransformの変換でうまいことできるかも
dataloader_resnet = DataLoader(AIRushDataset(
test_image_data_path,
test_meta_data,
label_path=None,
transform=transforms.Compose([
transforms.Resize((input_size_resnet, input_size_resnet)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])),
batch_size=batch_size,
shuffle=False,
num_workers=0,
pin_memory=True)
dataloader_inception = DataLoader(AIRushDataset(
test_image_data_path,
test_meta_data,
label_path=None,
transform=transforms.Compose([
transforms.Resize(
(input_size_inception, input_size_inception)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])),
batch_size=batch_size,
shuffle=False,
num_workers=0,
pin_memory=True)
dataloader_resnet_crop = DataLoader(AIRushDataset(
test_image_data_path,
test_meta_data,
label_path=None,
transform=transforms.Compose([
transforms.Resize(
(input_size_resnet + 20, input_size_resnet + 20)),
transforms.RandomCrop(input_size_resnet),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])),
batch_size=batch_size,
shuffle=False,
num_workers=0,
pin_memory=True)
for __modelnum__ in range(MODELNUM):
model_list[__modelnum__].to(device)
model_list[__modelnum__].eval()
predict_list = []
for batch_idx, (image_resnet, image_inception,
image_resnet_crop) in enumerate(
zip(dataloader_resnet, dataloader_inception,
dataloader_resnet_crop)):
image_resnet = image_resnet.to(device)
image_inception = image_inception.to(device)
image_resnet_crop = image_resnet_crop.to(device)
output0 = to_np(model_list[0](image_resnet).double()) # VGG
output1 = to_np(model_list[1](image_inception).squeeze(-1).squeeze(
-1).double()) # Inception
output2 = to_np(model_list[2](image_resnet).double()) # Resnet
output3 = to_np(
model_list[3](image_resnet_crop).double()) # Resnet
print(output0.shape)
print(output1.shape)
print(output2.shape)
print(output3.shape)
print(np.sum(output0, axis=1))
print(np.sum(output1, axis=1))
print(np.sum(output2, axis=1))
print(np.sum(output3, axis=1))
output0_sum = np.abs(np.sum(output0, axis=1))
output1_sum = np.abs(np.sum(output1, axis=1))
output2_sum = np.abs(np.sum(output2, axis=1))
output3_sum = np.abs(np.sum(output3, axis=1))
print(output0_sum.shape)
print(output1_sum.shape)
print(output2_sum.shape)
print(output3_sum.shape)
output0 /= output0_sum[:, np.newaxis]
output1 /= output1_sum[:, np.newaxis]
output2 /= output2_sum[:, np.newaxis]
output3 /= output3_sum[:, np.newaxis]
print(np.sum(output0[0, :]))
print(np.sum(output1[0, :]))
print(np.sum(output2[0, :]))
print(np.sum(output3[0, :]))
output2 *= 0.666 #出力が負なので、低い方が重みついてる
output = output0 + output1 + output2 + output3
predict = np.argmax(output, axis=1)
predict_list.append(predict)
predict_vector = np.concatenate(predict_list, axis=0)
# this return type should be a numpy array which has shape of (138343)
return predict_vector