def run_tensorflow(args, train_data=None, test_data=None):
# create a tensorflow session
run_config = tf.ConfigProto()
run_config.gpu_options.allow_growth = True
row_shapes = []
if test_data is not None:
row_shapes += [h.shape[0] for h in test_data]
if train_data is not None:
row_shapes += [h.shape[0] for h in train_data]
max_rows = np.max(row_shapes)
with tf.Session(config=run_config) as sess:
if args.debug:
sess = tf_debug.LocalCLIDebugWrapperSession(sess)
if not args.use_cnn2:
network = CNN(sess, args.vector_size, args.output_dim, max_rows,
args.conv_dim, args.conv_fmaps, args.hidden_units,
args.hidden_activation, args.dropout_rates,
args.batch_enabled, args.batch_size,
args.conv_activation, args.checkpoint_dir,
args.summary_dir)
else:
network = CNN2(sess,
args.vector_size,
args.output_dim,
max_rows,
enable_batch=args.batch_enabled,
batch_size=args.batch_size,
checkpoint_dir=args.checkpoint_dir,
summary_dir=args.summary_dir)
network.build_model()
print("\n".join([str(v) for v in tf.global_variables()]))
if args.train_cnn:
network.train(train_data, train_labels, args.learning_rate,
args.beta1, args.epochs, 500)
utils.logger.info("training finished")
utils.logger.info("start train data prediction test")
predicted_labels = network.predict(train_data)
utils.logger.info(
"scores: f1: %.4f, recall: %.4f, precision: %.4f" %
(sklearn.metrics.f1_score(
train_labels, predicted_labels, average='weighted'),
sklearn.metrics.recall_score(
train_labels, predicted_labels, average='weighted'),
sklearn.metrics.precision_score(
train_labels, predicted_labels, average='weighted')))
if args.validate_cnn:
if not args.train_cnn:
if args.checkpoint_file is not None:
if not network.load_file(args.checkpoint_file)[0]:
utils.logger.error(
"specified checkpoint_file is not valid")
exit(-1)
elif not network.load(args.checkpoint_dir)[0]:
utils.logger.error(
"No model available in checkpoint directory, train one first"
)
exit(-1)
utils.logger.info("start validatation on train data")
predicted_labels = network.predict(train_data)
utils.logger.info(
"scores: f1: %.4f, recall: %.4f, precision: %.4f" %
(sklearn.metrics.f1_score(
train_labels, predicted_labels, average='weighted'),
sklearn.metrics.recall_score(
train_labels, predicted_labels, average='weighted'),
sklearn.metrics.precision_score(
train_labels, predicted_labels, average='weighted')))
if args.predict_cnn:
if not args.train_cnn:
if args.checkpoint_file is not None:
if not network.load_file(args.checkpoint_file)[0]:
utils.logger.error(
"specified checkpoint_file is not valid")
exit(-1)
elif not network.load(args.checkpoint_dir)[0]:
utils.logger.error(
"No model available in checkpoint directory, train one first"
)
exit(-1)
result = network.predict(test_data)
utils.save_labels(args.o, result)
utils.logger.info("labels predicted and saved")
