def _transformations_experiment(data_loader: ZTFOutlierLoader,
transformer: AbstractTransformer, dataset_name: str, class_name: str,
save_path: str):
# Todo: pass as param_dict
batch_size = 128
transform_batch_size = 1024
(x_train, y_train), (
x_val, y_val), (
x_test, y_test) = data_loader.get_outlier_detection_datasets()
mdl = TransformODModel(
data_loader=data_loader, transformer=transformer,
input_shape=x_train.shape[1:], results_folder_name=save_path)
mdl.fit(x_train=x_train, x_val=x_val, transform_batch_size=transform_batch_size,
train_batch_size=batch_size,
epochs=EPOCHS # int(np.ceil(200 / transformer.n_transforms))
)
metrics_dict = mdl.evaluate_od(
x_train, x_test, y_test, dataset_name, class_name, x_val,
transform_batch_size=transform_batch_size,
additional_score_save_path_list=[save_path, mdl.specific_model_folder],
save_hist_folder_path=mdl.specific_model_folder)
del mdl, x_train, x_val, x_test
def _transformations_experiment(data_loader: ZTFOutlierLoader,
transformer: AbstractTransformer,
dataset_name: str, class_name: str,
save_path: str):
# Todo: pass as param_dict
batch_size = 128
transform_batch_size = 1024
(x_train, y_train), (x_val, y_val), (
x_test, y_test) = data_loader.get_outlier_detection_datasets()
mdl = TransformODModel(data_loader=data_loader,
transformer=transformer,
input_shape=x_train.shape[1:],
results_folder_name=save_path)
# mdl.fit(x_train=x_train, x_val=x_val, transform_batch_size=transform_batch_size,
# train_batch_size=batch_size,
# epochs=EPOCHS # int(np.ceil(200 / transformer.n_transforms))
# )
mdl.create_specific_model_paths()
mdl.build(tuple([None] + list(x_train.shape[1:])))
weights_path = os.path.join(
PROJECT_PATH,
'results/ztf-diri-val-refact/hits/Transformer_OD_Model/PlusKernel_transformer/Transformer_OD_Model_20191127-124219/checkpoints',
'final_weights.h5')
print(os.path.abspath(weights_path))
mdl.load_weights(weights_path)
metrics_dict = mdl.evaluate_od(
x_train,
x_test,
y_test,
dataset_name,
class_name,
x_val,
transform_batch_size=transform_batch_size,
additional_score_save_path_list=[save_path, mdl.specific_model_folder],
save_hist_folder_path=mdl.specific_model_folder)
del mdl
tf.config.experimental.set_memory_growth(gpu, True)
data_loader_params = {
loader_keys.DATA_PATH:
os.path.join(PROJECT_PATH, '../datasets/ztf_v1_bogus_added.pkl'),
loader_keys.VAL_SET_INLIER_PERCENTAGE:
0.1,
loader_keys.USED_CHANNELS: [0, 1, 2],
loader_keys.CROP_SIZE:
21,
general_keys.RANDOM_SEED:
42,
loader_keys.TRANSFORMATION_INLIER_CLASS_VALUE:
1
}
ztf_od_loader = ZTFOutlierLoader(data_loader_params)
(x_train, y_train), (x_val, y_val), (
x_test, y_test) = ztf_od_loader.get_outlier_detection_datasets()
transformer = Transformer()
model = TransformODSimpleModel(data_loader=ztf_od_loader,
transformer=transformer,
input_shape=x_train.shape[1:])
model.build(tuple([None] + list(x_train.shape[1:])))
# print(model.network.model().summary())
weight_path = os.path.join(PROJECT_PATH, 'results', model.name,
'my_checkpoint_simple.h5')
if os.path.exists(weight_path):
model.load_weights(weight_path)
else:
model.fit(x_train, x_val)
N_RUNS = 1
ztf_params = {
loader_keys.DATA_PATH:
os.path.join(PROJECT_PATH, '../datasets/ztf_v1_bogus_added.pkl'),
loader_keys.VAL_SET_INLIER_PERCENTAGE:
0.1,
loader_keys.USED_CHANNELS: [0, 1, 2],
loader_keys.CROP_SIZE:
21,
general_keys.RANDOM_SEED:
42,
loader_keys.TRANSFORMATION_INLIER_CLASS_VALUE:
1
}
ztf_outlier_dataset = ZTFOutlierLoader(ztf_params)
ztf_params[loader_keys.CROP_SIZE] = None
ztf_outlier_dataset_63 = ZTFOutlierLoader(ztf_params)
hits_params = {
loader_keys.DATA_PATH:
os.path.join(PROJECT_PATH, '../datasets/HiTS2013_300k_samples.pkl'),
loader_keys.N_SAMPLES_BY_CLASS:
10000,
loader_keys.TEST_PERCENTAGE:
0.2,
loader_keys.VAL_SET_INLIER_PERCENTAGE:
0.1,
loader_keys.USED_CHANNELS: [0, 1, 2, 3],
loader_keys.CROP_SIZE:
21,
general_keys.RANDOM_SEED:
def best_score_evaluation(result_folder_name, epochs, patience=0):
trainer_params = {
param_keys.RESULTS_FOLDER_NAME: result_folder_name,
'epochs': epochs,
'patience': patience,
}
# data loaders
hits_params = {
loader_keys.DATA_PATH:
os.path.join(PROJECT_PATH, '../datasets/HiTS2013_300k_samples.pkl'),
loader_keys.N_SAMPLES_BY_CLASS:
10000,
loader_keys.TEST_PERCENTAGE:
0.2,
loader_keys.VAL_SET_INLIER_PERCENTAGE:
0.1,
loader_keys.USED_CHANNELS: [0, 1, 2, 3], # [2],#
loader_keys.CROP_SIZE:
21,
general_keys.RANDOM_SEED:
42,
loader_keys.TRANSFORMATION_INLIER_CLASS_VALUE:
1
}
hits_loader = HiTSOutlierLoader(hits_params)
ztf_params = {
loader_keys.DATA_PATH:
os.path.join(PROJECT_PATH, '../datasets/ztf_v1_bogus_added.pkl'),
loader_keys.VAL_SET_INLIER_PERCENTAGE:
0.1,
loader_keys.USED_CHANNELS: [0, 1, 2],
loader_keys.CROP_SIZE:
21,
general_keys.RANDOM_SEED:
42,
loader_keys.TRANSFORMATION_INLIER_CLASS_VALUE:
1
}
ztf_loader = ZTFOutlierLoader(ztf_params)
# transformers
transformer_72 = transformations_tf.Transformer()
trans_transformer = transformations_tf.TransTransformer()
kernel_transformer = transformations_tf.KernelTransformer()
plus_kernel_transformer = transformations_tf.PlusKernelTransformer()
# trainers
hits_trainer = Trainer(hits_loader, trainer_params)
ztf_trainer = Trainer(ztf_loader, trainer_params)
model_constructors_list = (TransformODModel, )
transformers_list = (
plus_kernel_transformer,
kernel_transformer,
transformer_72,
trans_transformer,
)
trainers_list = (hits_trainer, ) # (ztf_trainer, hits_trainer, )
trainer_model_transformer_tuples = list(
itertools.product(trainers_list, model_constructors_list,
transformers_list))
for trainer, model_constructor, transformer in trainer_model_transformer_tuples:
trainer.train_model_n_times(model_constructor,
transformer,
trainer_params,
train_times=TRAIN_TIME)
hits_trainer.create_tables_of_results_folders()
tf.config.experimental.set_memory_growth(gpu, True)
ztf_params = {
loader_keys.DATA_PATH:
os.path.join(PROJECT_PATH, '../datasets/ztf_v1_bogus_added.pkl'),
loader_keys.VAL_SET_INLIER_PERCENTAGE:
0.1,
loader_keys.USED_CHANNELS: [0, 1, 2],
loader_keys.CROP_SIZE:
21,
general_keys.RANDOM_SEED:
42,
loader_keys.TRANSFORMATION_INLIER_CLASS_VALUE:
1
}
ztf_loader = ZTFOutlierLoader(ztf_params)
(x_train,
y_train), (x_val,
y_val), (x_test,
y_test) = ztf_loader.get_outlier_detection_datasets()
plus_kernel_transformer = transformations_tf.PlusKernelTransformer()
model = TransformODModel(data_loader=ztf_loader,
transformer=plus_kernel_transformer,
input_shape=x_train.shape[1:])
weight_path = os.path.join(
PROJECT_PATH,
'results/best_scores_refact/best_scores_resnet_1_epochs/Transformer_OD_Model/PlusKernel_Transformer',
'Transformer_OD_Model_20191223-192249/checkpoints/final_weights.ckpt')
# 'Transformer_OD_Model_20191223-193141/checkpoints/final_weights.ckpt')
model.load_weights(weight_path).expect_partial()
met_dict = model.evaluate_od(x_train, x_test, y_test, ztf_loader.name,