test_file_list = [f.split('/')[-1].split('\\')[-1][:-4] for f in test_file_list]
#five fold generator
fold = KFold(n_splits = 5, shuffle = True, random_state = 42)
for j, idx in enumerate(fold.split(file_list)):
j = 2
train_idx = idx[0]
val_idx = idx[1]
#20 % train/val split
file_list_val = list(map(file_list.__getitem__,val_idx))
file_list_train = list(map(file_list.__getitem__,train_idx))
#define dataset iterators
dataset = TGSSaltDataset(train_path, file_list_train, augmentation = True)
dataset_val = TGSSaltDataset(train_path, file_list_val)
test_dataset = TGSSaltDataset(test_path, test_file_list, is_test = True)
if '34' in parameter_path:
model = get_model(num_classes = 1, num_filters = 32, pretrained = True)
else:
model = get_model(encoder_depth = 101, num_classes = 1, num_filters=32, dropout_2d=0.2, pretrained=True, is_deconv=True)
model.load_state_dict(torch.load('../torch_parameters/' + parameter_path + '/model-' + str(j) + '.pt'))
model.train(False)
#training parameters
epoch = 100
learning_rate = 1e-3
test_file_list = glob.glob(os.path.join(test_path, 'images', '*.png'))
test_file_list = [f.split('/')[-1].split('\\')[-1][:-4] for f in test_file_list]
#five fold generator
fold = StratifiedKFold(n_splits = 10, shuffle = True, random_state = 42)
folds = list(fold.split(file_list,mask_class))
train_idx = folds[0][0]
val_idx = folds[0][1]
#20 % train/val split
file_list_val = list(map(file_list.__getitem__,val_idx))
file_list_train = list(map(file_list.__getitem__,train_idx))
#define dataset iterators
dataset = TGSSaltDataset(train_path, file_list_train)
dataset_val = TGSSaltDataset(train_path, file_list_val)
test_dataset = TGSSaltDataset(test_path, test_file_list, is_test = True)
class ConvBnRelu(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size):
super().__init__()
self.conv = nn.Sequential(nn.Conv2d(in_channels, out_channels, kernel_size = kernel_size, padding=1),
nn.BatchNorm2d(out_channels),
nn.ReLU(inplace=True)
)
def forward(self, x):
return self.conv(x)
class StackingFCN(nn.Module):
'CV5_resnext50_weighted_loss_no_drop_low_pixels_two_stage_SE_stratified_on_plateau_adam_hyper_decoder_gated'
]
submission_name = 'CV5_resnext50-101_weighted_loss_no_drop_low_pixels_two_stage_SE_stratified_on_plateau_adam_hyper_decoder_gated_finetuned_v2-tta.csv'
modelss = [[0, 1, 2, 3, 4], [0, 1, 2, 3, 4]]
weights = [0.7, 0.3]
model_predictions = []
test_path = '../test'
#test files
test_file_list = glob.glob(os.path.join(test_path, 'images', '*.png'))
test_file_list = [
f.split('/')[-1].split('\\')[-1][:-4] for f in test_file_list
]
test_dataset = TGSSaltDataset(test_path, test_file_list, is_test=True)
sigmoid = nn.Sigmoid()
for k, parameter_path in enumerate(parameter_path_list):
models = modelss[k]
for j in models:
#load best model
if '101' in parameter_path:
model = get_model_152(num_classes=1)
else:
model = get_model_101(num_classes=1)
model.load_state_dict(
torch.load('../torch_parameters/' + parameter_path + '/model-' +
str(j) + '.pt'))
#20 % train/val split
file_list_val = list(map(file_list.__getitem__, val_idx))
file_list_train = list(map(file_list.__getitem__, train_idx))
#redo mask list
mask_paths = [
train_path + '/masks/' + str(file) + '.png' for file in file_list_train
]
mask_class = [
get_mask_type(np.transpose(cv2.imread(mask), (2, 0, 1))[0, :, :] / 255)
for mask in mask_paths
]
#define dataset iterators
dataset = TGSSaltDataset(train_path,
file_list_train,
augmentation=True,
classes=mask_class)
dataset_val = TGSSaltDataset(train_path, file_list_val)
#define resnext50
model_50 = get_model_50(num_classes=1)
model_50.load_state_dict(
torch.load('../torch_parameters/' + parameter_path_50 + '/model-' +
str(fold_to_test) + '.pt'))
#define resnext101
model_101 = get_model_101(num_classes=1)
model_101.load_state_dict(
torch.load('../torch_parameters/' + parameter_path_101 + '/model-' +
str(fold_to_test) + '.pt'))
#redo file list
#file_list = [file for file in file_list if file not in removed_images]
#10 % train/val split
file_list_val = file_list[::10] #every 10th image into validation
file_list_train = [f for f in file_list if f not in file_list_val]
#test files
test_file_list = glob.glob(os.path.join(test_path, 'images', '*.png'))
test_file_list = [
f.split('/')[-1].split('\\')[-1][:-4] for f in test_file_list
]
#define dataset iterators
dataset = TGSSaltDataset(train_path, file_list_train, augmentation=True)
dataset_val = TGSSaltDataset(train_path, file_list_val)
test_dataset = TGSSaltDataset(test_path, test_file_list, is_test=True)
#load best model
model = get_model(num_classes=1, num_filters=32, pretrained=True)
model.load_state_dict(
torch.load('../torch_parameters/best_model_albunet_drop_pretrained.pt'))
model.train(False)
image = test_dataset.__getitem__(555)
with torch.no_grad():
y_pred = model(image[0].unsqueeze(0).cuda()).cpu().data.numpy()
y_pred_flipped = to_tensor(
hflip(
to_pil_image(