def generate_graphs(eval_args):
"""
Generate graphs (networkx format) given a trained generative model
and save them to a directory
:param eval_args: ArgsEvaluate object
"""
train_args = eval_args.train_args
if train_args.note == 'GraphRNN':
gen_graphs = gen_graphs_graph_rnn(eval_args)
elif train_args.note == 'DFScodeRNN':
gen_graphs = gen_graphs_dfscode_rnn(eval_args)
elif train_args.note == 'DGMG':
gen_graphs = gen_graphs_dgmg(eval_args)
if os.path.isdir(eval_args.current_graphs_save_path):
shutil.rmtree(eval_args.current_graphs_save_path)
os.makedirs(eval_args.current_graphs_save_path)
save_graphs(eval_args.current_graphs_save_path, gen_graphs)
def train_model(train_dir, model_name, show_val):
prepare_data(train_dir)
if model_name == 'custom':
model, history, val_pred_probs = train_custom_model(100, 64)
elif model_name == 'bottleneck':
override_keras_directory_iterator_next()
generate_and_save_bottleneck_features(64)
model, history, val_pred_probs = train_bottleneck_model(100, 64)
elif model_name == 'finetune':
override_keras_directory_iterator_next()
model, history, val_pred_probs = train_fine_tune_model(30, 32)
min_loss, min_loss_acc = find_min_loss_acc(history)
postfix = '_{0:.2f}l_{1:.2f}a'.format(min_loss, min_loss_acc)
# Rename model weights
try:
print("Renaming model weights.\n")
tmp_path = cfg.model_weights_path.replace('.h5',
'{}.h5'.format(postfix))
os.rename(cfg.model_weights_path, tmp_path)
except Exception as e:
print("Renaming model weights failed: {}".format(e))
# Save model architecture
try:
print("Saving model architecture.\n")
save_model(model, postfix)
except Exception as e:
print("Saving model architecture failed: {}".format(e))
# Save loss and accuracy graphs
try:
print("Saving loss and accuracy graphs.\n")
save_graphs(history, postfix)
except Exception as e:
print("Saving loss and accuracy graphs failed: {}".format(e))
# Save validation results
try:
print("Saving validation results.\n")
# Get validation images
val_imgs, val_labels, val_img_paths = load_train_dir(cfg.val_data_dir)
# Reorder predictions probabilities if binary class
if cfg.nb_classes == 2:
val_pred_probs = np.append(1.0 - val_pred_probs,
val_pred_probs,
axis=1)
# Get predictions labels
val_pred_labels = np.argmax(val_pred_probs, axis=1)
# Save validation results
write_val_results(postfix, val_img_paths, val_labels, val_pred_probs,
val_pred_labels)
# Show misclassified validation images
if show_val:
show_mc_val_images(train_dir, val_labels, val_img_paths,
val_pred_probs, val_pred_labels)
except Exception as e:
print("Saving validation results failed: {}".format(e))