def test_train_batch(data):
data.split_config = DataSplitConfig(0, 0, 0)
data.batch_size = 2
data.augment = False
assert len(data.train_batch_data(0).df.ImageId.unique()) == 2
assert len(data.train_batch_data(1).df.ImageId.unique()) == 1
def test_train_size(sample_df):
# 30% training data, 3 of 10 items
config = DataSplitConfig(0.5, 0.2, 0)
split = GroupDataSplit(sample_df, key='group_id', config=config)
print(split.train_df())
assert len(split.train_df()) == 3
def test_val_batch(data):
data.split_config = DataSplitConfig(0.99, 0, 0)
data.batch_size = 2
data.augment = True # No-op
data._split_data = None
assert len(data.val_batch_data(0).df.ImageId.unique()) == 2
assert len(data.val_batch_data(1).df.ImageId.unique()) == 1
def test_neural_network_input_output_with_augmentation(data):
data.split_config = DataSplitConfig(0, 0, 0)
data.batch_size = 2
data.augment = True
inputs, outputs = data.train_batch_data(0).to_nn((500, 500))
assert inputs.shape == (2, 500, 500, 3)
assert outputs.shape == (2, 500, 500, 1)
def test_train_batch_count_with_augmentation(data):
data.split_config = DataSplitConfig(0, 0, 0)
data.batch_size = 2
data.augment = True
assert data.train_batch_count() == 3
# Number of images has been halved since each with have an associated
# augmented counterpart
assert len(data.train_batch_data(0).df.ImageId.unique()) == 1
def model(args):
# Use Rio 3band data
config = DataConfig(DEFAULT_DATA_DIR, BAND3, RIO_REGION)
# Train (70%), Val (10%), Test (20%), Fixed seed
split_config = DataSplitConfig(0.1, 0.2,
1337) # Create data with batch size
data = Data(config,
split_config,
batch_size=args.max_examples
if hasattr(args, 'max_examples') else args.batch_size,
augment=True)
# Exclude empty polygons
data.data_filter = lambda df: df.sq_ft > 0
# Create model
return UNet(data, (args.size, args.size),
max_batches=args.max_batches,
epochs=args.epochs,
weights_path=args.weights_path,
gpu_count=args.gpu_count,
learning_rate=args.learning_rate)
def test_test_batch_count(data):
data.split_config = DataSplitConfig(0, 0.4, 0)
data.batch_size = 1
data.augment = True # No-op
assert data.test_batch_count() == 2
def test_test_batch(data):
data.split_config = DataSplitConfig(0, 0.6, 0)
data.batch_size = 1
assert len(data.test_batch_data(0).df.ImageId.unique()) == 1
assert len(data.test_batch_data(1).df.ImageId.unique()) == 1
def test_val_batch_count(data):
data.split_config = DataSplitConfig(0.99, 0, 0)
data.batch_size = 1
data.augment = True # No-op
assert data.val_batch_count() == 3
def test_train_batch_count_with_whole_last_batch(data):
data.split_config = DataSplitConfig(0, 0, 0)
data.batch_size = 3
data.augment = False
assert data.train_batch_count() == 1
def test_test_size(sample_df):
# 20% test data, 2 of 10 items
config = DataSplitConfig(0.5, 0.2, 0)
split = GroupDataSplit(sample_df, key='group_id', config=config)
assert len(split.test_df) == 2
def test_validation_size(sample_df):
# 50% validation data, 5 of 10 items
config = DataSplitConfig(0.5, 0.2, 0)
split = GroupDataSplit(sample_df, key='group_id', config=config)
assert len(split.val_df) == 5