def train_vae():
# Instantiate and Train Autoencoder
vae_kwargs = {}
vae_kwargs["latent_dim"] = 2
vae_kwargs["hidden_dim"] = [15, 7]
vae_kwargs["epochs"] = 8
vae_kwargs["batch_size"] = 128
# vae_kwargs["model_path"] = ae_model_path
vae = VAEModel(in_train.shape[1], **vae_kwargs)
vae.train(in_train, in_test)
vae.save_model()
inlier_scores = vae.compute_anomaly_score(in_test)
outlier_scores = vae.compute_anomaly_score(out_test)
print(inlier_scores)
print(outlier_scores)
metrics = eval_utils.evaluate_model(
inlier_scores, outlier_scores, model_name="vae", show_plot=False)
print(metrics)
return metrics
try:
os.mkdir(log_dir)
logging.info('============================================================')
logging.info(' Logging Directory: %s' %log_dir)
logging.info('============================================================\n')
except OSError:
pass
# ==================================================================================================================================================================
# ============== LOAD THE MODELS ===================================================================================================================================
# ==================================================================================================================================================================
vae_network = VariationalAutoencoder
model = VAEModel(vae_network, config, model_name, log_dir)
# Load the vae model as our baseline
path = os.path.join(project_code_root, config["model_directory"])
model.load_from_path(path, config["model_name"] , config["latest_model_epoch"])
# ==================================================================================================================================================================
# ============== LOAD THE DATA =====================================================================================================================================
# ==================================================================================================================================================================
if preprocess_enabled == "slice":
logging.info('============================================================')
logging.info('Loading training data from: ' + project_data_root)
data_tr = data_freiburg_numpy_to_preprocessed_hdf5.load_cropped_data_sliced(basepath = project_data_root,
idx_start = config['train_data_start_idx'],
idx_end = config['train_data_end_idx'],
from models.vae import VAEModel
from .bert import BERTModel
MODELS = {
BERTModel.code(): BERTModel,
VAEModel.code(): VAEModel,
}
def model_factory(args, dataset_meta):
model = MODELS[args.model_code]
return model(args, dataset_meta)
end = one_subject.shape[0] - (one_subject.shape[0] -
num_of_y_slices) / 2
temp_reshape[i * num_of_y_slices:(i + 1) *
num_of_y_slices, :, :, :, :] = one_subject[
int(start):int(end), :, :, :, :]
images_vl = temp_reshape
# ====================================================================================
# Initialize the network architecture, training parameters, model_name, and logging directory
# ====================================================================================
#Initialize the VAE
vae_network = VariationalAutoencoder
model = VAEModel(vae_network, config, model_name, log_dir)
model.initialize()
model.summarize()
# Should we continue training an existing model or is it a new model
if continue_training:
path = os.path.join(project_code_root, config["model_directory"])
model.load_from_path(path, config["model_name"],
config["latest_model_epoch"])
already_completed_epochs = config["latest_model_epoch"]
else:
model.summarize()
already_completed_epochs = 0