def main():
# Set Tensorflow verbosity
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
# Get configuration
config = get_args()
# Optimize the parameters
if config.optimize:
print('Creating save directories...')
create_dirs([config.optimizer_path, config.model_path])
print('Initializing optimizer...')
# optimizer = HyperEngineOptimizer(config)
print('Optimizing...')
# optimizer.optimize()
# Set up test/train environment
if config.infer or config.train:
print('Building model...')
params = generate_params(config)
model = Tiramisu(params)
if config.train:
print('Creating save directories...')
create_dirs([config.model_path])
print('Initializing trainer...')
trainer = Trainer(config)
trainer.train()
if config.infer:
print('Creating save directories...')
create_dirs([config.prediction_path])
print('Initializing predicter...')
predicter = Predicter(config)
predicter.predict()
# Prepare submission
if config.rle:
print('Creating submission save directory...')
create_dirs([config.submission_path])
prepare_submission(config.prediction_path, config.submission_path,
'submission')
print('Done!')
import matplotlib.animation as manimation
from mpl_toolkits.basemap import Basemap
## import torch modules
import torch
import torch.optim as optim
# import our model and dataloader
from src.utils.parser import get_args
from src.utils.utils import createfolder, remove_file, Adam16
from src.utils.loss import Criterion, LOSS
from src.utils.visulize import plot_input, plot_target
from src.runs.GRUs import get_dataloader, get_model, get_optimizer, test
if __name__ == "__main__":
args = get_args()
pd.set_option('precision', 4)
# set seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
# Get the model
model = get_model(args=args)
# set optimizer
optimizer = get_optimizer(args=args, model=model)
param_pt = os.path.join(args.params_folder, 'params_10.pt')
checkpoint = torch.load(param_pt, map_location=args.device)
model.load_state_dict(checkpoint['model_state_dict'])
def test_io():
# Set Tensorflow verbosity
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '4'
config = get_args()
print('Creating save directories...')
create_dirs([
'./image_tests/', './image_tests/tiled/',
'./image_tests/tiled/images/', './image_tests/tiled/masks/',
'./image_tests/detiled', './image_tests/detiled/images/',
'./image_tests/detiled/masks/', config.submission_path
])
print('Initializing data loader...')
data_loader = DataGenerator(config)
# Progress bar
batches = tqdm(range(data_loader.num_iterations_debug),
total=data_loader.num_iterations_debug,
desc="Processing Batches ")
# Use session to evaluate real tensors
with tf.Session() as sess:
# Initialize data generator
data_loader.initialize(sess, 'debug')
next_element = data_loader.get_input()
for i in batches:
# Get a batch and tile it
image, mask, name = sess.run(tile_data(*next_element))
# Save tiled images
for i in range(config.batch_size):
cv2.imwrite(
os.path.join('./image_tests/tiled/images',
name[i].decode('utf-8')),
255 * image[i, :, :])
cv2.imwrite(
os.path.join('./image_tests/tiled/masks',
name[i].decode('utf-8')), 255 * mask[i, :, :])
# Convert images back to tensors because that's how the de-tiler expects them
image = tf.convert_to_tensor(image)
mask = tf.convert_to_tensor(mask)
name = tf.convert_to_tensor(name)
# De-tile images
image, mask, name = sess.run(data_crop(image, mask, name))
# Save de-tiled images
for i in range(config.batch_size):
cv2.imwrite(
os.path.join('./image_tests/detiled/images',
name[i].decode('utf-8')),
255 * image[i, :, :])
cv2.imwrite(
os.path.join('./image_tests/detiled/masks',
name[i].decode('utf-8')), 255 * mask[i, :, :])
# Close progress bar
batches.close()
# Compute RLEs for saved masks
prepare_submission('./image_tests/detiled/masks/', config.submission_path,
'test')
compare_rle('./image_tests/', './rle_tests/')
compare_iou('./rle_tests/')
print('Done!')
def compare_iou(config):
input_path = config.submission_path
# Load test images
test_path = get_max_filename(input_path, 'test', '.csv')
test = pd.read_csv(test_path)
test_images, test_image_names = csv_to_image(test)
if config.compute_iou:
# Load source images
source = pd.read_csv('./data/train.csv')
source_images, _ = csv_to_image(source)
# Compute IoU values
iou_a = mean_iou(source_images, test_images)
iou_b = vector_iou(source_images, test_images)
print('Mean IoU A: ' + str(iou_a))
print('Mean IoU B: ' + str(iou_b))
if config.rle_to_image:
create_dirs(['./converted_images/'])
save_images(test_images, test_image_names, './converted_images/')
if __name__ == '__main__':
# If executing from this folder, change the working directory to the parent to manipulate project files
if os.path.dirname(os.path.abspath(__file__)) == os.getcwd():
project_root = os.path.abspath('..')
os.chdir(project_root)
# Get args and pass them to compare_rle
config = get_args()
compare_iou(config)