DO_VALID = True # disable this to not bother with the validation set evaluation
DO_TEST = True # disable this to not generate predictions on the testset
target_filename = os.path.basename(ANALYSIS_PATH).replace(".pkl", ".npy.gz")
target_path_valid = os.path.join("predictions/final/augmented/valid",
target_filename)
target_path_test = os.path.join("predictions/final/augmented/test",
target_filename)
print("Loading model data etc.")
analysis = np.load(ANALYSIS_PATH)
input_sizes = [(69, 69), (69, 69)]
ds_transforms = [
ra.build_ds_transform(3.0, target_size=input_sizes[0]),
ra.build_ds_transform(3.0, target_size=input_sizes[1]) +
ra.build_augmentation_transform(rotation=45)
]
num_input_representations = len(ds_transforms)
# split training data into training + a small validation set
num_train = load_data.num_train
num_valid = num_train // 10 # integer division
num_train -= num_valid
num_test = load_data.num_test
valid_ids = load_data.train_ids[num_train:]
train_ids = load_data.train_ids[:num_train]
test_ids = load_data.test_ids
# DATA_TEST_PATH = "data/images_test_color_cropped33_singletf.npy.gz"
# DATA2_TEST_PATH = "data/images_test_color_8x_singletf.npy.gz"
# FEATURES_PATTERN = "features/try_convnet_chunked_ra_b3sched.%s.npy"
if continueAnalysis:
print "Loading model data"
analysis = np.load(ANALYSIS_PATH)
print "Set up data loading"
# TODO: adapt this so it loads the validation data from JPEGs and does the processing realtime
input_sizes = [(69, 69), (69, 69)]
ds_transforms = [
ra.build_ds_transform(3.0, target_size=input_sizes[0]),
ra.build_ds_transform(3.0, target_size=input_sizes[1]) + ra.build_augmentation_transform(rotation=45)
]
num_input_representations = len(ds_transforms)
augmentation_params = {
'zoom_range': (1.0 / 1.3, 1.3),
'rotation_range': (0, 360),
'shear_range': (0, 0),
'translation_range': (-4, 4),
'do_flip': True,
}
augmented_data_gen = ra.realtime_augmented_data_gen(num_chunks=NUM_CHUNKS, chunk_size=CHUNK_SIZE,
augmentation_params=augmentation_params, ds_transforms=ds_transforms,
DO_TEST = True # disable this to not generate predictions on the testset
target_filename = os.path.basename(ANALYSIS_PATH).replace(".pkl", ".npy.gz")
target_path_valid = os.path.join("predictions/final/augmented/valid", target_filename)
target_path_test = os.path.join("predictions/final/augmented/test", target_filename)
print "Loading model data etc."
analysis = np.load(ANALYSIS_PATH)
input_sizes = [(69, 69), (69, 69)]
ds_transforms = [
ra.build_ds_transform(3.0, target_size=input_sizes[0]),
ra.build_ds_transform(3.0, target_size=input_sizes[1]) + ra.build_augmentation_transform(rotation=45)]
num_input_representations = len(ds_transforms)
# split training data into training + a small validation set
num_train = load_data.num_train
num_valid = num_train // 10 # integer division
num_train -= num_valid
num_test = load_data.num_test
valid_ids = load_data.train_ids[num_train:]
train_ids = load_data.train_ids[:num_train]
test_ids = load_data.test_ids
train_indices = np.arange(num_train)
# DATA2_TRAIN_PATH = "data/images_train_color_8x_singletf.npy.gz"
# DATA_VALIDONLY_PATH = "data/images_validonly_color_cropped33_singletf.npy.gz"
# DATA2_VALIDONLY_PATH = "data/images_validonly_color_8x_singletf.npy.gz"
# DATA_TEST_PATH = "data/images_test_color_cropped33_singletf.npy.gz"
# DATA2_TEST_PATH = "data/images_test_color_8x_singletf.npy.gz"
TARGET_PATH = "predictions/final/try_convnet_cc_multirotflip_3x69r45_maxout2048.csv"
ANALYSIS_PATH = "analysis/final/try_convnet_cc_multirotflip_3x69r45_maxout2048.pkl"
# FEATURES_PATTERN = "features/try_convnet_chunked_ra_b3sched.%s.npy"
print "Set up data loading"
# TODO: adapt this so it loads the validation data from JPEGs and does the processing realtime
input_sizes = [(69, 69), (69, 69)]
ds_transforms = [ra.build_ds_transform(3.0, target_size=input_sizes[0])]
num_input_representations = len(ds_transforms)
augmentation_params = {
"zoom_range": (1.0 / 1.3, 1.3),
"rotation_range": (0, 360),
"shear_range": (0, 0),
"translation_range": (-4, 4),
"do_flip": False,
}
augmented_data_gen = ra.realtime_augmented_data_gen(
num_chunks=NUM_CHUNKS,
chunk_size=CHUNK_SIZE,
augmentation_params=augmentation_params,