def convert_model():
# Getting anchors and labels for the prediction
class_names = get_classes(config.classes_path)
anchors = read_anchors(config.anchors_path)
num_classes = config.num_classes
num_anchors = config.num_anchors
# Retriving the input shape of the model i.e. (608x608), (416x416), (320x320)
input_shape = (config.input_shape, config.input_shape)
# Defining placeholder for passing the image data onto the model
image_tensor = tf.placeholder(dtype=tf.float32,
shape=[None, None, None, 3],
name='input_image')
image_shape = tf.placeholder(dtype=tf.int32, shape=[2], name='input_shape')
output_nodes = yolo(input_images=image_tensor,
is_training=False,
config_path=config.yolov3_cfg_path,
num_classes=config.num_classes)
print(output_nodes)
sess = tf.Session()
scale_1, scale_2, scale3 = tf.identity(
output_nodes[0], name='scale_1'), tf.identity(
output_nodes[1], name='scale_2'), tf.identity(output_nodes[2],
name='scale_3')
ckpt_path = config.model_dir
exponential_moving_average_obj = tf.train.ExponentialMovingAverage(
config.weight_decay)
saver = tf.train.Saver(
exponential_moving_average_obj.variables_to_restore())
ckpt = tf.train.get_checkpoint_state(ckpt_path)
# chkp.print_tensors_in_checkpoint_file(checkmate.get_best_checkpoint(ckpt_path), tensor_name='', all_tensors=True)
# exit()
if config.pre_train is True:
load_ops = load_weights(tf.global_variables(),
config.yolov3_weights_path)
sess.run(load_ops)
elif ckpt and tf.train.checkpoint_exists(ckpt.model_checkpoint_path):
print('Restoring model ', checkmate.get_best_checkpoint(ckpt_path))
saver.restore(sess, checkmate.get_best_checkpoint(ckpt_path))
print('Model Loaded!')
else:
print("No appropriate weights found for creating protobuf file")
if not os.path.exists(config.model_export_path.split('/')[1]):
os.mkdir(config.model_export_path.split('/')[1])
freeze_graph(sess, config.model_export_path)
sess.close()
def predict_new(x):
saver = tf.train.Saver()
with tf.Session() as session:
ckpt = get_best_checkpoint(model.ckpt_dir)
saver.restore(session, ckpt)
if model_name == 'm2':
pred = session.run([model.predictions], {model.x: x})
else:
y_ = model.q_y_x_model(model.x)
pred = session.run([y_], {model.x: x})
return pred
def test():
logger.info("Loading Data...")
logger.info("Data processing...")
test_data = dh.load_data_and_labels(args.test_file, args.word2vec_file)
logger.info("Data padding...")
test_dataset = dh.MyData(test_data, args.pad_seq_len, device)
test_loader = DataLoader(test_dataset,
batch_size=args.batch_size,
shuffle=False)
VOCAB_SIZE, EMBEDDING_SIZE, pretrained_word2vec_matrix = dh.load_word2vec_matrix(
args.word2vec_file)
criterion = Loss()
net = HMIDP(args, VOCAB_SIZE, EMBEDDING_SIZE,
pretrained_word2vec_matrix).to(device)
checkpoint_file = cm.get_best_checkpoint(CPT_DIR,
select_maximum_value=False)
checkpoint = torch.load(checkpoint_file)
net.load_state_dict(checkpoint['model_state_dict'])
net.eval()
logger.info("Scoring...")
true_labels, predicted_scores = [], []
batches = trange(len(test_loader), desc="Batches", leave=True)
for batch_cnt, batch in zip(batches, test_loader):
x_test_fb_content, x_test_fb_question, x_test_fb_option, \
x_test_fb_clens, x_test_fb_qlens, x_test_fb_olens, y_test_fb = batch
logits, scores = net(x_test_fb_content, x_test_fb_question,
x_test_fb_option)
for i in y_test_fb[0].tolist():
true_labels.append(i)
for j in scores[0].tolist():
predicted_scores.append(j)
# Calculate the Metrics
test_rmse = mean_squared_error(true_labels, predicted_scores)**0.5
test_r2 = r2_score(true_labels, predicted_scores)
test_pcc, test_doa = dh.evaluation(true_labels, predicted_scores)
logger.info(
"All Test set: PCC {0:.4f} | DOA {1:.4f} | RMSE {2:.4f} | R2 {3:.4f}".
format(test_pcc, test_doa, test_rmse, test_r2))
logger.info('Test Finished.')
logger.info('Creating the prediction file...')
dh.create_prediction_file(save_dir=SAVE_DIR,
identifiers=test_data['f_id'],
predictions=predicted_scores)
logger.info('All Finished.')
def test():
logger.info("Loading Data...")
logger.info("Data processing...")
test_data = dh.load_data_and_labels(args.test_file)
test_dataset = dh.MyData(test_data.activity, test_data.timestep, test_data.labels)
test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, collate_fn=dh.collate_fn)
# Load word2vec model
COURSE_SIZE = dh.course2vec(args.course2vec_file)
criterion = Loss()
net = MOOCNet(args, COURSE_SIZE).to(device)
checkpoint_file = cm.get_best_checkpoint(CPT_DIR, select_maximum_value=False)
checkpoint = torch.load(checkpoint_file)
net.load_state_dict(checkpoint['model_state_dict'])
net.eval()
logger.info("Scoring...")
true_labels, predicted_scores, predicted_labels = [], [], []
batches = trange(len(test_loader), desc="Batches", leave=True)
for batch_cnt, batch in zip(batches, test_loader):
x_test, tsp_test, y_test = create_input_data(batch)
logits, scores = net(x_test, tsp_test)
for i in y_test.tolist():
true_labels.append(i)
for j in scores.tolist():
predicted_scores.append(j)
if j >= 0.5:
predicted_labels.append(1)
else:
predicted_labels.append(0)
# Calculate the Metrics
logger.info('Test Finished.')
logger.info('Creating the prediction file...')
dh.create_prediction_file(save_dir=SAVE_DIR, identifiers=test_data.id, predictions=predicted_labels)
logger.info('All Finished.')
def test_harnn():
"""Test HARNN model."""
# Print parameters used for the model
dh.tab_printer(args, logger)
# Load data
logger.info("Loading data...")
logger.info("Data processing...")
test_data = dh.load_data_and_labels(args.test_file, args.num_classes_list, args.total_classes,
args.word2vec_file, data_aug_flag=False)
logger.info("Data padding...")
x_test, y_test, y_test_tuple = dh.pad_data(test_data, args.pad_seq_len)
y_test_labels = test_data.labels
# Load harnn model
OPTION = dh._option(pattern=1)
if OPTION == 'B':
logger.info("Loading best model...")
checkpoint_file = cm.get_best_checkpoint(BEST_CPT_DIR, select_maximum_value=True)
else:
logger.info("Loading latest model...")
checkpoint_file = tf.train.latest_checkpoint(CPT_DIR)
logger.info(checkpoint_file)
graph = tf.Graph()
with graph.as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=args.allow_soft_placement,
log_device_placement=args.log_device_placement)
session_conf.gpu_options.allow_growth = args.gpu_options_allow_growth
sess = tf.Session(config=session_conf)
with sess.as_default():
# Load the saved meta graph and restore variables
saver = tf.train.import_meta_graph("{0}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
# Get the placeholders from the graph by name
input_x = graph.get_operation_by_name("input_x").outputs[0]
input_y_first = graph.get_operation_by_name("input_y_first").outputs[0]
input_y_second = graph.get_operation_by_name("input_y_second").outputs[0]
input_y_third = graph.get_operation_by_name("input_y_third").outputs[0]
input_y_fourth = graph.get_operation_by_name("input_y_fourth").outputs[0]
input_y = graph.get_operation_by_name("input_y").outputs[0]
dropout_keep_prob = graph.get_operation_by_name("dropout_keep_prob").outputs[0]
beta = graph.get_operation_by_name("beta").outputs[0]
is_training = graph.get_operation_by_name("is_training").outputs[0]
# Tensors we want to evaluate
first_scores = graph.get_operation_by_name("first-output/scores").outputs[0]
second_scores = graph.get_operation_by_name("second-output/scores").outputs[0]
third_scores = graph.get_operation_by_name("third-output/scores").outputs[0]
fourth_scores = graph.get_operation_by_name("fourth-output/scores").outputs[0]
scores = graph.get_operation_by_name("output/scores").outputs[0]
loss = graph.get_operation_by_name("loss/loss").outputs[0]
# Split the output nodes name by '|' if you have several output nodes
output_node_names = "first-output/scores|second-output/scores|third-output/scores|fourth-output/scores|output/scores"
# Save the .pb model file
output_graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def,
output_node_names.split("|"))
tf.train.write_graph(output_graph_def, "graph", "graph-harnn-{0}.pb".format(MODEL), as_text=False)
# Generate batches for one epoch
batches = dh.batch_iter(list(zip(x_test, y_test, y_test_tuple, y_test_labels)),
args.batch_size, 1, shuffle=False)
test_counter, test_loss = 0, 0.0
# Collect the predictions here
true_labels = []
predicted_labels = []
predicted_scores = []
# Collect for calculating metrics
true_onehot_labels = []
predicted_onehot_scores = []
predicted_onehot_labels_ts = []
predicted_onehot_labels_tk = [[] for _ in range(args.topK)]
true_onehot_first_labels = []
true_onehot_second_labels = []
true_onehot_third_labels = []
true_onehot_fourth_labels = []
predicted_onehot_scores_first = []
predicted_onehot_scores_second = []
predicted_onehot_scores_third = []
predicted_onehot_scores_fourth = []
predicted_onehot_labels_first = []
predicted_onehot_labels_second = []
predicted_onehot_labels_third = []
predicted_onehot_labels_fourth = []
for batch_test in batches:
x_batch_test, y_batch_test, y_batch_test_tuple, y_batch_test_labels = zip(*batch_test)
y_batch_test_first = [i[0] for i in y_batch_test_tuple]
y_batch_test_second = [j[1] for j in y_batch_test_tuple]
y_batch_test_third = [k[2] for k in y_batch_test_tuple]
y_batch_test_fourth = [t[3] for t in y_batch_test_tuple]
feed_dict = {
input_x: x_batch_test,
input_y_first: y_batch_test_first,
input_y_second: y_batch_test_second,
input_y_third: y_batch_test_third,
input_y_fourth: y_batch_test_fourth,
input_y: y_batch_test,
dropout_keep_prob: 1.0,
beta: args.beta,
is_training: False
}
batch_first_scores, batch_second_scores, batch_third_scores, batch_fourth_scores, batch_scores, cur_loss = \
sess.run([first_scores, second_scores, third_scores, fourth_scores, scores, loss], feed_dict)
# Prepare for calculating metrics
for onehot_labels in y_batch_test:
true_onehot_labels.append(onehot_labels)
for onehot_labels in y_batch_test_first:
true_onehot_first_labels.append(onehot_labels)
for onehot_labels in y_batch_test_second:
true_onehot_second_labels.append(onehot_labels)
for onehot_labels in y_batch_test_third:
true_onehot_third_labels.append(onehot_labels)
for onehot_labels in y_batch_test_fourth:
true_onehot_fourth_labels.append(onehot_labels)
for onehot_scores in batch_scores:
predicted_onehot_scores.append(onehot_scores)
for onehot_scores in batch_first_scores:
predicted_onehot_scores_first.append(onehot_scores)
for onehot_scores in batch_second_scores:
predicted_onehot_scores_second.append(onehot_scores)
for onehot_scores in batch_third_scores:
predicted_onehot_scores_third.append(onehot_scores)
for onehot_scores in batch_fourth_scores:
predicted_onehot_scores_fourth.append(onehot_scores)
# Get the predicted labels by threshold
batch_predicted_labels_ts, batch_predicted_scores_ts = \
dh.get_label_threshold(scores=batch_scores, threshold=args.threshold)
# Add results to collection
for labels in y_batch_test_labels:
true_labels.append(labels)
for labels in batch_predicted_labels_ts:
predicted_labels.append(labels)
for values in batch_predicted_scores_ts:
predicted_scores.append(values)
# Get one-hot prediction by threshold
batch_predicted_onehot_labels_ts = \
dh.get_onehot_label_threshold(scores=batch_scores, threshold=args.threshold)
batch_predicted_onehot_labels_first = \
dh.get_onehot_label_threshold(scores=batch_first_scores, threshold=args.threshold)
batch_predicted_onehot_labels_second = \
dh.get_onehot_label_threshold(scores=batch_second_scores, threshold=args.threshold)
batch_predicted_onehot_labels_third = \
dh.get_onehot_label_threshold(scores=batch_third_scores, threshold=args.threshold)
batch_predicted_onehot_labels_fourth = \
dh.get_onehot_label_threshold(scores=batch_fourth_scores, threshold=args.threshold)
for onehot_labels in batch_predicted_onehot_labels_ts:
predicted_onehot_labels_ts.append(onehot_labels)
for onehot_labels in batch_predicted_onehot_labels_first:
predicted_onehot_labels_first.append(onehot_labels)
for onehot_labels in batch_predicted_onehot_labels_second:
predicted_onehot_labels_second.append(onehot_labels)
for onehot_labels in batch_predicted_onehot_labels_third:
predicted_onehot_labels_third.append(onehot_labels)
for onehot_labels in batch_predicted_onehot_labels_fourth:
predicted_onehot_labels_fourth.append(onehot_labels)
# Get one-hot prediction by topK
for i in range(args.topK):
batch_predicted_onehot_labels_tk = dh.get_onehot_label_topk(scores=batch_scores, top_num=i + 1)
for onehot_labels in batch_predicted_onehot_labels_tk:
predicted_onehot_labels_tk[i].append(onehot_labels)
test_loss = test_loss + cur_loss
test_counter = test_counter + 1
# Calculate Precision & Recall & F1
test_pre_ts = precision_score(y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_ts), average='micro')
test_pre_first = precision_score(y_true=np.array(true_onehot_first_labels),
y_pred=np.array(predicted_onehot_labels_first), average='micro')
test_pre_second = precision_score(y_true=np.array(true_onehot_second_labels),
y_pred=np.array(predicted_onehot_labels_second), average='micro')
test_pre_third = precision_score(y_true=np.array(true_onehot_third_labels),
y_pred=np.array(predicted_onehot_labels_third), average='micro')
test_pre_fourth = precision_score(y_true=np.array(true_onehot_fourth_labels),
y_pred=np.array(predicted_onehot_labels_fourth), average='micro')
test_rec_ts = recall_score(y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_ts), average='micro')
test_rec_first = recall_score(y_true=np.array(true_onehot_first_labels),
y_pred=np.array(predicted_onehot_labels_first), average='micro')
test_rec_second = recall_score(y_true=np.array(true_onehot_second_labels),
y_pred=np.array(predicted_onehot_labels_second), average='micro')
test_rec_third = recall_score(y_true=np.array(true_onehot_third_labels),
y_pred=np.array(predicted_onehot_labels_third), average='micro')
test_rec_fourth = recall_score(y_true=np.array(true_onehot_fourth_labels),
y_pred=np.array(predicted_onehot_labels_fourth), average='micro')
test_F1_ts = f1_score(y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_ts), average='micro')
test_F1_first = f1_score(y_true=np.array(true_onehot_first_labels),
y_pred=np.array(predicted_onehot_labels_first), average='micro')
test_F1_second = f1_score(y_true=np.array(true_onehot_second_labels),
y_pred=np.array(predicted_onehot_labels_second), average='micro')
test_F1_third = f1_score(y_true=np.array(true_onehot_third_labels),
y_pred=np.array(predicted_onehot_labels_third), average='micro')
test_F1_fourth = f1_score(y_true=np.array(true_onehot_fourth_labels),
y_pred=np.array(predicted_onehot_labels_fourth), average='micro')
# Calculate the average AUC
test_auc = roc_auc_score(y_true=np.array(true_onehot_labels),
y_score=np.array(predicted_onehot_scores), average='micro')
# Calculate the average PR
test_prc = average_precision_score(y_true=np.array(true_onehot_labels),
y_score=np.array(predicted_onehot_scores), average="micro")
test_prc_first = average_precision_score(y_true=np.array(true_onehot_first_labels),
y_score=np.array(predicted_onehot_scores_first), average="micro")
test_prc_second = average_precision_score(y_true=np.array(true_onehot_second_labels),
y_score=np.array(predicted_onehot_scores_second), average="micro")
test_prc_third = average_precision_score(y_true=np.array(true_onehot_third_labels),
y_score=np.array(predicted_onehot_scores_third), average="micro")
test_prc_fourth = average_precision_score(y_true=np.array(true_onehot_fourth_labels),
y_score=np.array(predicted_onehot_scores_fourth), average="micro")
test_loss = float(test_loss / test_counter)
logger.info("All Test Dataset: Loss {0:g} | AUC {1:g} | AUPRC {2:g}"
.format(test_loss, test_auc, test_prc))
# Predict by threshold
logger.info("Predict by threshold: Precision {0:g}, Recall {1:g}, F1 {2:g}"
.format(test_pre_ts, test_rec_ts, test_F1_ts))
logger.info("Predict by threshold in Level-1: Precision {0:g}, Recall {1:g}, F1 {2:g}, AUPRC {3:g}"
.format(test_pre_first, test_rec_first, test_F1_first, test_prc_first))
logger.info("Predict by threshold in Level-2: Precision {0:g}, Recall {1:g}, F1 {2:g}, AUPRC {3:g}"
.format(test_pre_second, test_rec_second, test_F1_second, test_prc_second))
logger.info("Predict by threshold in Level-3: Precision {0:g}, Recall {1:g}, F1 {2:g}, AUPRC {3:g}"
.format(test_pre_third, test_rec_third, test_F1_third, test_prc_third))
logger.info("Predict by threshold in Level-4: Precision {0:g}, Recall {1:g}, F1 {2:g}, AUPRC {3:g}"
.format(test_pre_fourth, test_rec_fourth, test_F1_fourth, test_prc_fourth))
# Save the prediction result
if not os.path.exists(SAVE_DIR):
os.makedirs(SAVE_DIR)
dh.create_prediction_file(output_file=SAVE_DIR + "/predictions.json", data_id=test_data.patent_id,
all_labels=true_labels, all_predict_labels=predicted_labels,
all_predict_scores=predicted_scores)
logger.info("All Done.")
def train(ckpt_path, log_path, class_path, decay_steps=2000, decay_rate=0.8):
""" Function to train the model.
ckpt_path: string, path for saving/restoring the model
log_path: string, path for saving the training/validation logs
class_path: string, path for the classes of the dataset
decay_steps: int, steps after which the learning rate is to be decayed
decay_rate: float, rate to carrying out exponential decay
"""
# Getting the anchors
anchors = read_anchors(config.anchors_path)
if not os.path.exists(config.data_dir):
os.mkdir(config.data_dir)
classes = get_classes(class_path)
# Building the training pipeline
graph = tf.get_default_graph()
with graph.as_default():
# Getting the training data
with tf.name_scope('data_parser/'):
train_reader = Parser('train',
config.data_dir,
config.anchors_path,
config.output_dir,
config.num_classes,
input_shape=config.input_shape,
max_boxes=config.max_boxes)
train_data = train_reader.build_dataset(config.train_batch_size //
config.subdivisions)
train_iterator = train_data.make_one_shot_iterator()
val_reader = Parser('val',
config.data_dir,
config.anchors_path,
config.output_dir,
config.num_classes,
input_shape=config.input_shape,
max_boxes=config.max_boxes)
val_data = val_reader.build_dataset(config.val_batch_size //
config.subdivisions)
val_iterator = val_data.make_one_shot_iterator()
is_training = tf.placeholder(
dtype=tf.bool, shape=[], name='train_flag'
) # Used for different behaviour of batch normalization
mode = tf.placeholder(dtype=tf.int16, shape=[], name='mode_flag')
def train():
# images, bbox, bbox_true_13, bbox_true_26, bbox_true_52 = train_iterator.get_next()
return train_iterator.get_next()
def valid():
# images, bbox, bbox_true_13, bbox_true_26, bbox_true_52 = val_iterator.get_next()
return val_iterator.get_next()
images, bbox, bbox_true_13, bbox_true_26, bbox_true_52 = tf.cond(
pred=tf.equal(mode, 1),
true_fn=train,
false_fn=valid,
name='train_val_cond')
images.set_shape([None, config.input_shape, config.input_shape, 3])
bbox.set_shape([None, config.max_boxes, 5])
grid_shapes = [
config.input_shape // 32, config.input_shape // 16,
config.input_shape // 8
]
draw_box(images, bbox)
# Extracting the pre-defined yolo graph from the darknet cfg file
if not os.path.exists(ckpt_path):
os.mkdir(ckpt_path)
output = yolo(images, is_training, config.yolov3_cfg_path,
config.num_classes)
# Declaring the parameters for GT
with tf.name_scope('Targets'):
bbox_true_13.set_shape([
None, grid_shapes[0], grid_shapes[0], 3, 5 + config.num_classes
])
bbox_true_26.set_shape([
None, grid_shapes[1], grid_shapes[1], 3, 5 + config.num_classes
])
bbox_true_52.set_shape([
None, grid_shapes[2], grid_shapes[2], 3, 5 + config.num_classes
])
y_true = [bbox_true_13, bbox_true_26, bbox_true_52]
# Compute Loss
with tf.name_scope('Loss_and_Detect'):
yolo_loss = compute_loss(output,
y_true,
anchors,
config.num_classes,
print_loss=False)
l2_loss = tf.losses.get_regularization_loss()
loss = yolo_loss + l2_loss
yolo_loss_summary = tf.summary.scalar('yolo_loss', yolo_loss)
l2_loss_summary = tf.summary.scalar('l2_loss', l2_loss)
total_loss_summary = tf.summary.scalar('Total_loss', loss)
# Declaring the parameters for training the model
with tf.name_scope('train_parameters'):
epoch_loss = []
global_step = tf.Variable(0, trainable=False, name='global_step')
learning_rate = tf.train.exponential_decay(config.learning_rate,
global_step,
decay_steps, decay_rate)
tf.summary.scalar('learning rate', learning_rate)
# Define optimizer for minimizing the computed loss
with tf.name_scope('Optimizer'):
#optimizer = tf.train.MomentumOptimizer(learning_rate=learning_rate, momentum=config.momentum)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
if config.pre_train:
train_vars = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope='yolo')
grads = optimizer.compute_gradients(loss=loss,
var_list=train_vars)
gradients = [(tf.placeholder(dtype=tf.float32,
shape=grad[1].get_shape()),
grad[1]) for grad in grads]
gradients = gradients * config.subdivisions
train_step = optimizer.apply_gradients(
grads_and_vars=gradients, global_step=global_step)
else:
grads = optimizer.compute_gradients(loss=loss)
gradients = [(tf.placeholder(dtype=tf.float32,
shape=grad[1].get_shape()),
grad[1]) for grad in grads]
gradients = gradients * config.subdivisions
train_step = optimizer.apply_gradients(
grads_and_vars=gradients, global_step=global_step)
#################################### Training loop ############################################################
# A saver object for saving the model
best_ckpt_saver = checkmate.BestCheckpointSaver(save_dir=ckpt_path,
num_to_keep=5)
summary_op = tf.summary.merge_all()
summary_op_valid = tf.summary.merge(
[yolo_loss_summary, l2_loss_summary, total_loss_summary])
init_op = tf.global_variables_initializer()
# Defining some train loop dependencies
gpu_config = tf.ConfigProto(log_device_placement=False)
gpu_config.gpu_options.allow_growth = True
sess = tf.Session(config=gpu_config)
tf.logging.set_verbosity(tf.logging.ERROR)
train_summary_writer = tf.summary.FileWriter(
os.path.join(log_path, 'train'), sess.graph)
val_summary_writer = tf.summary.FileWriter(
os.path.join(log_path, 'val'), sess.graph)
# Restoring the model
ckpt = tf.train.get_checkpoint_state(ckpt_path)
if ckpt and tf.train.checkpoint_exists(ckpt.model_checkpoint_path):
print('Restoring model ', checkmate.get_best_checkpoint(ckpt_path))
tf.train.Saver().restore(sess,
checkmate.get_best_checkpoint(ckpt_path))
print('Model Loaded!')
elif config.pre_train is True:
load_ops = load_weights(tf.global_variables(scope='darknet53'),
config.darknet53_weights_path)
sess.run(load_ops)
else:
sess.run(init_op)
print('Uninitialized variables: ',
sess.run(tf.report_uninitialized_variables()))
epochbar = tqdm(range(config.Epoch))
for epoch in epochbar:
epochbar.set_description('Epoch %s of %s' % (epoch, config.Epoch))
mean_loss_train = []
mean_loss_valid = []
trainbar = tqdm(range(config.train_num // config.train_batch_size))
for k in trainbar:
total_grad = []
for minibach in range(config.subdivisions):
train_summary, loss_train, grads_and_vars = sess.run(
[summary_op, loss, grads],
feed_dict={
is_training: True,
mode: 1
})
total_grad += grads_and_vars
feed_dict = {is_training: True, mode: 1}
for i in range(len(gradients)):
feed_dict[gradients[i][0]] = total_grad[i][0]
# print(np.shape(feed_dict))
_ = sess.run(train_step, feed_dict=feed_dict)
train_summary_writer.add_summary(train_summary, epoch)
train_summary_writer.flush()
mean_loss_train.append(loss_train)
trainbar.set_description('Train loss: %s' % str(loss_train))
print('Validating.....')
valbar = tqdm(range(config.val_num // config.val_batch_size))
for k in valbar:
val_summary, loss_valid = sess.run([summary_op_valid, loss],
feed_dict={
is_training: False,
mode: 0
})
val_summary_writer.add_summary(val_summary, epoch)
val_summary_writer.flush()
mean_loss_valid.append(loss_valid)
valbar.set_description('Validation loss: %s' % str(loss_valid))
mean_loss_train = np.mean(mean_loss_train)
mean_loss_valid = np.mean(mean_loss_valid)
print('\n')
print('Train loss after %d epochs is: %f' %
(epoch + 1, mean_loss_train))
print('Validation loss after %d epochs is: %f' %
(epoch + 1, mean_loss_valid))
print('\n\n')
if ((epoch + 1) % 3) == 0:
best_ckpt_saver.handle(mean_loss_valid, sess,
tf.constant(epoch))
print('Tuning Completed!!')
train_summary_writer.close()
val_summary_writer.close()
sess.close()
def test_cnn():
"""Test ABCNN model."""
# Load data
logger.info("✔︎ Loading data...")
logger.info("Recommended padding Sequence length is: {0}".format(
FLAGS.pad_seq_len))
logger.info("✔︎ Test data processing...")
test_data = dh.load_data_and_labels(FLAGS.test_data_file,
FLAGS.embedding_dim)
logger.info("✔︎ Test data padding...")
x_test_front, x_test_behind, y_test = dh.pad_data(test_data,
FLAGS.pad_seq_len)
y_test_labels = test_data.labels
# Load abcnn model
BEST_OR_LATEST = input("☛ Load Best or Latest Model?(B/L): ")
while not (BEST_OR_LATEST.isalpha()
and BEST_OR_LATEST.upper() in ['B', 'L']):
BEST_OR_LATEST = input(
"✘ The format of your input is illegal, please re-input: ")
if BEST_OR_LATEST == 'B':
logger.info("✔︎ Loading best model...")
checkpoint_file = cm.get_best_checkpoint(FLAGS.best_checkpoint_dir,
select_maximum_value=True)
else:
logger.info("✔︎ Loading latest model...")
checkpoint_file = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
logger.info(checkpoint_file)
graph = tf.Graph()
with graph.as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=FLAGS.allow_soft_placement,
log_device_placement=FLAGS.log_device_placement)
session_conf.gpu_options.allow_growth = FLAGS.gpu_options_allow_growth
sess = tf.Session(config=session_conf)
with sess.as_default():
# Load the saved meta graph and restore variables
saver = tf.train.import_meta_graph(
"{0}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
# Get the placeholders from the graph by name
input_x_front = graph.get_operation_by_name(
"input_x_front").outputs[0]
input_x_behind = graph.get_operation_by_name(
"input_x_behind").outputs[0]
input_y = graph.get_operation_by_name("input_y").outputs[0]
dropout_keep_prob = graph.get_operation_by_name(
"dropout_keep_prob").outputs[0]
is_training = graph.get_operation_by_name("is_training").outputs[0]
# Tensors we want to evaluate
predictions = graph.get_operation_by_name(
"output/predictions").outputs[0]
topKPreds = graph.get_operation_by_name(
"output/topKPreds").outputs[0]
accuracy = graph.get_operation_by_name(
"accuracy/accuracy").outputs[0]
loss = graph.get_operation_by_name("loss/loss").outputs[0]
# Split the output nodes name by '|' if you have several output nodes
output_node_names = "output/logits|output/predictions|output/softmax_scores|output/topKPreds"
# Save the .pb model file
output_graph_def = tf.graph_util.convert_variables_to_constants(
sess, sess.graph_def, output_node_names.split("|"))
tf.train.write_graph(output_graph_def,
"graph",
"graph-abcnn-{0}.pb".format(MODEL),
as_text=False)
# Generate batches for one epoch
batches = dh.batch_iter(list(
zip(x_test_front, x_test_behind, y_test, y_test_labels)),
FLAGS.batch_size,
1,
shuffle=False)
# Collect the predictions here
all_labels = []
all_predicted_labels = []
all_predicted_values = []
for index, x_test_batch in enumerate(batches):
x_batch_front, x_batch_behind, y_batch, y_batch_labels = zip(
*x_test_batch)
feed_dict = {
input_x_front: x_batch_front,
input_x_behind: x_batch_behind,
input_y: y_batch,
dropout_keep_prob: 1.0,
is_training: False
}
all_labels = np.append(all_labels, y_batch_labels)
batch_predicted_labels = sess.run(predictions, feed_dict)
all_predicted_labels = np.concatenate(
[all_predicted_labels, batch_predicted_labels])
batch_predicted_values = sess.run(topKPreds, feed_dict)
all_predicted_values = np.append(all_predicted_values,
batch_predicted_values)
batch_loss = sess.run(loss, feed_dict)
batch_acc = sess.run(accuracy, feed_dict)
logger.info(
"✔︎ Test batch {0}: loss {1:g}, accuracy {2:g}.".format(
(index + 1), batch_loss, batch_acc))
# Save the prediction result
if not os.path.exists(SAVE_DIR):
os.makedirs(SAVE_DIR)
dh.create_prediction_file(output_file=SAVE_DIR +
"/predictions.json",
front_data_id=test_data.front_testid,
behind_data_id=test_data.behind_testid,
all_labels=all_labels,
all_predict_labels=all_predicted_labels,
all_predict_values=all_predicted_values)
logger.info("✔︎ Done.")
def train(self,
Data,
n_epochs,
l_bs,
u_bs,
lr,
eval_samps=None,
binarize=False,
verbose=1,
decay_ratio=0.75,
decay_period=200,
h_opt=False,
keep_ckpt=True,
restore=False):
""" Method for training the models """
self.data_init(Data, eval_samps, l_bs, u_bs)
self.global_step = tf.Variable(0, trainable=False, name='global_step')
# self.global_epoch = tf.Variable(0, trainable=False, name='global_epoch')
self.epoch = 0
#self.lr = self.set_learning_rate([lr[0], 1600, lr[0] / 10.0])
self.lr = self.set_learning_rate(
[lr[0], lr[0] / 10.0, decay_period, decay_ratio], 'exp')
# define optimizer
optimizer = tf.train.AdamOptimizer(learning_rate=self.lr)
gvs = optimizer.compute_gradients(self.loss)
# clip gradients
capped_gvs = [(tf.clip_by_value(grad, -1., 1.), var)
for grad, var in gvs]
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
self.optimizer = optimizer.apply_gradients(
capped_gvs, global_step=self.global_step)
self.y_pred = self.predict(self.x)
self.curve_array = np.zeros((n_epochs + 1, 14))
if self.learning_paradigm == 'unsupervised':
self.elbo_l_curve = tf.reduce_mean(self.unlabelled_loss(self.x))
self.qy_ll_curve = tf.reduce_mean(self.qy_loss(self.x))
self.elbo_u_curve = tf.reduce_mean(self.unlabelled_loss(self.x))
else:
if self.model_name == 'adgm' or self.model_name == 'adg_dgm':
self.elbo_l_curve = tf.reduce_mean(
self.labelled_loss(self.x, self.y)[0])
self.qy_ll_curve = tf.reduce_mean(
self.labelled_loss(self.x, self.y)[1])
else:
self.elbo_l_curve = tf.reduce_mean(
self.labelled_loss(self.x, self.y))
self.qy_ll_curve = tf.reduce_mean(self.qy_loss(self.x, self.y))
self.elbo_u_curve = tf.reduce_mean(self.unlabelled_loss(self.x))
self.compute_accuracies()
# initialize session and train
with self.session as sess:
sess.run(tf.global_variables_initializer())
if restore == True:
saver_for_restore = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(self.ckpt_dir)
best_ckpt = get_best_checkpoint(self.ckpt_dir)
best_epoch = int(re.match('.*?([0-9]+)$', best_ckpt).group(1))
best_ckpt_usable = re.sub('-([0-9]+)$', "", best_ckpt)
saver_for_restore.restore(sess, best_ckpt_usable)
self.epoch = best_epoch
self.curve_array[self.epoch] = self.calc_curve_vals(sess, Data)
if verbose == 3:
self.print_verbose3(self.epoch)
if keep_ckpt == True:
saver = BestCheckpointSaver(save_dir=self.ckpt_dir,
num_to_keep=2,
maximize=True)
while self.epoch < n_epochs:
x_labelled, labels, x_unlabelled, _ = \
Data.next_batch(l_bs, u_bs)
if binarize is True:
x_labelled = self.binarize(x_labelled)
x_unlabelled = self.binarize(x_unlabelled)
fd = self.training_fd(x_labelled, labels, x_unlabelled)
_, loss_batch = sess.run([self.optimizer, self.loss], fd)
if Data._epochs_unlabelled > self.epoch:
self.epoch += 1
# sess.run(self.global_epoch.assign(self.epoch)
self.curve_array[self.epoch] = \
self.calc_curve_vals(sess, Data)
if h_opt == True and self.epoch > 20:
if self.curve_array[self.epoch, 12] < 0.07:
raise Exception('results too bad')
if h_opt == True and self.epoch > 40:
if self.curve_array[self.epoch, 12] < 0.1:
raise Exception('results too bad')
if keep_ckpt == True:
saver.handle(self.curve_array[self.epoch, 6], sess,
self.global_step, self.epoch)
if verbose == 1:
fd = self._printing_feed_dict(Data, x_labelled,
x_unlabelled, labels,
eval_samps, binarize)
self.print_verbose1(self.epoch, fd, sess)
elif verbose == 2:
fd = self._printing_feed_dict(Data, x_labelled,
x_unlabelled, labels,
eval_samps, binarize)
self.print_verbose2(self.epoch, fd, sess)
elif verbose == 3:
self.print_verbose3(self.epoch)
if self.epoch % 10 == 0:
y_pred_test = sess.run([self.y_pred], {
self.x: Data.data['x_test'],
K.learning_phase(): 0
})[0]
conf_mat = confusion_matrix(
Data.data['y_test'].argmax(1),
y_pred_test.argmax(1))
np.save(
os.path.join(
self.output_dir, 'conf_mat_' + self.name +
'_' + str(self.epoch) + '.npy'), conf_mat)
np.save(
os.path.join(
self.output_dir, 'y_pred_' + self.name +
'_' + str(self.epoch) + '.npy'),
y_pred_test)
np.save(
os.path.join(
self.output_dir, 'y_true_' + self.name +
'_' + str(self.epoch) + '.npy'),
Data.data['y_test'])
if np.sum(np.isnan(self.curve_array)) > 0:
print(
'loss is nan, going back to previous best checkpoint'
)
best_ckpt = get_best_checkpoint(self.ckpt_dir)
best_epoch = int(
re.match('.*?([0-9]+)$', best_ckpt).group(1))
best_ckpt_usable = re.sub('-([0-9]+)$', "", best_ckpt)
self.epoch = best_epoch
saver._saver.restore(sess, best_ckpt_usable)
return self.curve_array
def test_parnn():
"""Test PARNN model."""
# Load data
logger.info("✔︎ Loading data...")
logger.info("✔︎ Test data processing...")
test_data = dh.load_data_and_labels(FLAGS.test_data_file,
FLAGS.embedding_dim,
data_aug_flag=False)
logger.info("✔︎ Test data padding...")
x_test_content, x_test_question, x_test_option, y_test = dh.pad_data(
test_data, FLAGS.pad_seq_len)
# Load parnn model
BEST_OR_LATEST = input("☛ Load Best or Latest Model?(B/L): ")
while not (BEST_OR_LATEST.isalpha()
and BEST_OR_LATEST.upper() in ['B', 'L']):
BEST_OR_LATEST = input(
"✘ The format of your input is illegal, please re-input: ")
if BEST_OR_LATEST.upper() == 'B':
logger.info("✔︎ Loading best model...")
checkpoint_file = cm.get_best_checkpoint(FLAGS.best_checkpoint_dir,
select_maximum_value=True)
else:
logger.info("✔︎ Loading latest model...")
checkpoint_file = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
logger.info(checkpoint_file)
graph = tf.Graph()
with graph.as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=FLAGS.allow_soft_placement,
log_device_placement=FLAGS.log_device_placement)
session_conf.gpu_options.allow_growth = FLAGS.gpu_options_allow_growth
sess = tf.Session(config=session_conf)
with sess.as_default():
# Load the saved meta graph and restore variables
saver = tf.train.import_meta_graph(
"{0}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
# Get the placeholders from the graph by name
input_x_content = graph.get_operation_by_name(
"input_x_content").outputs[0]
input_x_question = graph.get_operation_by_name(
"input_x_question").outputs[0]
input_x_option = graph.get_operation_by_name(
"input_x_option").outputs[0]
input_y = graph.get_operation_by_name("input_y").outputs[0]
dropout_keep_prob = graph.get_operation_by_name(
"dropout_keep_prob").outputs[0]
is_training = graph.get_operation_by_name("is_training").outputs[0]
# Tensors we want to evaluate
scores = graph.get_operation_by_name("output/scores").outputs[0]
loss = graph.get_operation_by_name("loss/loss").outputs[0]
# Split the output nodes name by '|' if you have several output nodes
output_node_names = "output/scores"
# Save the .pb model file
output_graph_def = tf.graph_util.convert_variables_to_constants(
sess, sess.graph_def, output_node_names.split("|"))
tf.train.write_graph(output_graph_def,
"graph",
"graph-parnn-{0}.pb".format(MODEL),
as_text=False)
# Generate batches for one epoch
batches = dh.batch_iter(
(x_test_content, x_test_question, x_test_option, y_test),
FLAGS.batch_size,
1,
shuffle=False)
test_counter, test_loss = 0, 0.0
# Collect the predictions here
true_labels = []
predicted_scores = []
for batch_test in batches:
x_content_batch_front, x_content_batch_behind, \
x_question_batch_front, x_question_batch_behind, \
x_option_batch_front, x_option_batch_behind, \
y_batch_front, y_batch_behind = zip(*batch_test)
feed_dict = {
input_x_content:
[x_content_batch_front, x_content_batch_behind],
input_x_question:
[x_question_batch_front, x_question_batch_behind],
input_x_option:
[x_option_batch_front, x_option_batch_behind],
input_y: [y_batch_front, y_batch_behind],
dropout_keep_prob:
1.0,
is_training:
False
}
batch_scores, cur_loss = sess.run([scores, loss], feed_dict)
# Prepare for calculating metrics
for i in y_batch_front:
true_labels.append(i)
for j in batch_scores[0]:
predicted_scores.append(j)
print(predicted_scores)
test_loss = test_loss + cur_loss
test_counter = test_counter + 1
# Calculate RMSE
rmse = mean_squared_error(true_labels, predicted_scores)**0.5
test_loss = float(test_loss / test_counter)
logger.info("☛ All Test Dataset: Loss {0:g} | RMSE {1:g}".format(
test_loss, rmse))
# Save the prediction result
if not os.path.exists(SAVE_DIR):
os.makedirs(SAVE_DIR)
dh.create_prediction_file(output_file=SAVE_DIR +
"/predictions.json",
all_id=test_data.id,
all_labels=true_labels,
all_predict_scores=predicted_scores)
logger.info("✔︎ Done.")
def run_inference(img_path, output_dir, args):
""" A function making inference using the pre-trained darknet weights in the tensorflow
framework
Input:
img_path: string, path to the image on which inference is to be run, path to the image directory containing images in the case of multiple images.
output_dir: string, directory for saving the output
args: argparse object
"""
# Reading the images
if not os.path.exists(output_dir):
os.mkdir(output_dir)
if not os.path.exists(os.path.join(output_dir, 'images')):
os.mkdir(os.path.join(output_dir, 'images'))
if not os.path.exists(os.path.join(output_dir, 'labels')):
os.mkdir(os.path.join(output_dir, 'labels'))
output_dir_images = os.path.join(output_dir, 'images')
output_dir_labels = os.path.join(output_dir, 'labels')
file_names = sorted(os.listdir(img_path))
images_batch = read_image(img_path)
# Getting anchors and labels for the prediction
class_names = get_classes(config.classes_path)
num_classes = config.num_classes
num_anchors = config.num_anchors
# Retriving the input shape of the model i.e. (608x608), (416x416), (320x320)
input_shape = (config.input_shape, config.input_shape)
# Defining placeholder for passing the image data onto the model
image_tensor = tf.placeholder(dtype=tf.float32,
shape=[None, None, None, 3])
image_shape = tf.placeholder(dtype=tf.int32, shape=[2])
model = model(image_tensor,
is_training=False,
num_classes=config.num_classes)
output_nodes, model_layers = model.forward()
print('Summary of the model created.......\n')
for layer in model_layers:
print(layer)
# Creating a session for running the model
gpu_config = tf.ConfigProto(log_device_placement=False)
gpu_config.gpu_options.allow_growth = True
sess = tf.Session(config=gpu_config)
output_values = predict(output_nodes, num_classes, input_shape,
image_shape)
ckpt_path = config.model_dir + 'valid/'
exponential_moving_average_obj = tf.train.ExponentialMovingAverage(
config.weight_decay)
saver = tf.train.Saver(
exponential_moving_average_obj.variables_to_restore())
ckpt = tf.train.get_checkpoint_state(ckpt_path)
# chkp.print_tensors_in_checkpoint_file(checkmate.get_best_checkpoint(ckpt_path), tensor_name='', all_tensors=True)
# exit()
if ckpt and tf.train.checkpoint_exists(ckpt.model_checkpoint_path):
print('Restoring model ', checkmate.get_best_checkpoint(ckpt_path))
saver.restore(sess, checkmate.get_best_checkpoint(ckpt_path))
print('Model Loaded!')
total_time_pred = []
for x in range(len(images_batch)):
image = images_batch[x]
new_image_size = (config.input_shape, config.input_shape)
image_data = np.array(resize_image(image, new_image_size))
print('Image height: {}\tImage width: {}'.format(
image.shape[0], image.shape[1]))
img = image_data / 255.
img = np.expand_dims(img, 0) # Adding the batch dimension
tick = time()
# Actually run the graph in a tensorflow session to get the outputs
out_values = sess.run([output_values],
feed_dict={
image_tensor: img,
image_shape:
[image.shape[0], image.shape[1]]
})
tock = time()
total_time_pred.append(tock - tick)
print('Found {} boxes for {} in {}sec'.format(len(out_boxes), 'img',
tock - tick))
######################## Visualization ######################
font = ImageFont.truetype(font='./font/FiraMono-Medium.otf',
size=np.floor(1e-2 * image.shape[1] +
0.5).astype(np.int32))
thickness = (image.shape[0] + image.shape[1]) // 500 # do day cua BB
image = Image.fromarray((image).astype('uint8'), mode='RGB')
output_labels = open(
os.path.join(output_dir_labels,
file_names[x].split('.')[0] + '.txt'), 'w')
### DO ALL THE PLOTTING THING IF REQUIRED ###
### SAVE THE IMAGE ###
output_labels.close() # Saving labels
sess.close()
total_time_pred = sum(total_time_pred[1:])
print('FPS of model with post processing over {} images is {}'.format(
len(images_batch) - 1, (len(images_batch) - 1) / total_time_pred))
def test_han():
"""Test HAN model."""
# Load data
logger.info("✔︎ Loading data...")
logger.info("Recommended padding Sequence length is: {0}".format(
FLAGS.pad_seq_len))
logger.info("✔︎ Test data processing...")
test_data = dh.load_data_and_labels(FLAGS.test_data_file,
FLAGS.num_classes,
FLAGS.embedding_dim,
data_aug_flag=False)
logger.info("✔︎ Test data padding...")
x_test, y_test = dh.pad_data(test_data, FLAGS.pad_seq_len)
y_test_labels = test_data.labels
# Load han model
BEST_OR_LATEST = input("☛ Load Best or Latest Model?(B/L): ")
while not (BEST_OR_LATEST.isalpha()
and BEST_OR_LATEST.upper() in ['B', 'L']):
BEST_OR_LATEST = input(
"✘ The format of your input is illegal, please re-input: ")
if BEST_OR_LATEST == 'B':
logger.info("✔︎ Loading best model...")
checkpoint_file = cm.get_best_checkpoint(FLAGS.best_checkpoint_dir,
select_maximum_value=True)
else:
logger.info("✔︎ Loading latest model...")
checkpoint_file = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
logger.info(checkpoint_file)
graph = tf.Graph()
with graph.as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=FLAGS.allow_soft_placement,
log_device_placement=FLAGS.log_device_placement)
session_conf.gpu_options.allow_growth = FLAGS.gpu_options_allow_growth
sess = tf.Session(config=session_conf)
with sess.as_default():
# Load the saved meta graph and restore variables
saver = tf.train.import_meta_graph(
"{0}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
# Get the placeholders from the graph by name
input_x = graph.get_operation_by_name("input_x").outputs[0]
input_y = graph.get_operation_by_name("input_y").outputs[0]
dropout_keep_prob = graph.get_operation_by_name(
"dropout_keep_prob").outputs[0]
is_training = graph.get_operation_by_name("is_training").outputs[0]
# Tensors we want to evaluate
scores = graph.get_operation_by_name("output/scores").outputs[0]
loss = graph.get_operation_by_name("loss/loss").outputs[0]
# Split the output nodes name by '|' if you have several output nodes
output_node_names = "output/logits|output/scores"
# Save the .pb model file
output_graph_def = tf.graph_util.convert_variables_to_constants(
sess, sess.graph_def, output_node_names.split("|"))
tf.train.write_graph(output_graph_def,
"graph",
"graph-han-{0}.pb".format(MODEL),
as_text=False)
# Generate batches for one epoch
batches = dh.batch_iter(list(zip(x_test, y_test, y_test_labels)),
FLAGS.batch_size,
1,
shuffle=False)
# Collect the predictions here
all_labels = []
all_predicted_labels = []
all_predicted_values = []
# Calculate the metric
test_counter, test_loss, test_rec, test_pre, test_F = 0, 0.0, 0.0, 0.0, 0.0
for batch_test in batches:
x_batch_test, y_batch_test, y_batch_test_labels = zip(
*batch_test)
feed_dict = {
input_x: x_batch_test,
input_y: y_batch_test,
dropout_keep_prob: 1.0,
is_training: False
}
batch_scores, cur_loss = sess.run([scores, loss], feed_dict)
# Predict by threshold
predicted_labels_threshold, predicted_values_threshold = \
dh.get_label_using_scores_by_threshold(scores=batch_scores, threshold=FLAGS.threshold)
cur_rec, cur_pre, cur_F = 0.0, 0.0, 0.0
for index, predicted_label_threshold in enumerate(
predicted_labels_threshold):
rec_inc, pre_inc = dh.cal_metric(predicted_label_threshold,
y_batch_test[index])
cur_rec, cur_pre = cur_rec + rec_inc, cur_pre + pre_inc
cur_rec = cur_rec / len(y_batch_test)
cur_pre = cur_pre / len(y_batch_test)
test_rec, test_pre = test_rec + cur_rec, test_pre + cur_pre
# Add results to collection
for item in y_batch_test_labels:
all_labels.append(item)
for item in predicted_labels_threshold:
all_predicted_labels.append(item)
for item in predicted_values_threshold:
all_predicted_values.append(item)
test_loss = test_loss + cur_loss
test_counter = test_counter + 1
test_loss = float(test_loss / test_counter)
test_rec = float(test_rec / test_counter)
test_pre = float(test_pre / test_counter)
test_F = dh.cal_F(test_rec, test_pre)
logger.info("☛ All Test Dataset: Loss {0:g}".format(test_loss))
# Predict by threshold
logger.info(
"☛ Predict by threshold: Recall {0:g}, Precision {1:g}, F {2:g}"
.format(test_rec, test_pre, test_F))
# Save the prediction result
if not os.path.exists(SAVE_DIR):
os.makedirs(SAVE_DIR)
dh.create_prediction_file(output_file=SAVE_DIR +
"/predictions.json",
data_id=test_data.testid,
all_labels=all_labels,
all_predict_labels=all_predicted_labels,
all_predict_values=all_predicted_values)
logger.info("✔︎ Done.")
def visualize():
"""Visualize HARNN model."""
# Load word2vec model
word2idx, embedding_matrix = dh.load_word2vec_matrix(args.word2vec_file)
# Load data
logger.info("Loading data...")
logger.info("Data processing...")
test_data = dh.load_data_and_labels(args, args.test_file, word2idx)
# Load harnn model
OPTION = dh._option(pattern=1)
if OPTION == 'B':
logger.info("Loading best model...")
checkpoint_file = cm.get_best_checkpoint(BEST_CPT_DIR,
select_maximum_value=True)
else:
logger.info("Loading latest model...")
checkpoint_file = tf.train.latest_checkpoint(CPT_DIR)
logger.info(checkpoint_file)
graph = tf.Graph()
with graph.as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=args.allow_soft_placement,
log_device_placement=args.log_device_placement)
session_conf.gpu_options.allow_growth = args.gpu_options_allow_growth
sess = tf.Session(config=session_conf)
with sess.as_default():
# Load the saved meta graph and restore variables
saver = tf.train.import_meta_graph(
"{0}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
# Get the placeholders from the graph by name
input_x = graph.get_operation_by_name("input_x").outputs[0]
input_y_first = graph.get_operation_by_name(
"input_y_first").outputs[0]
input_y_second = graph.get_operation_by_name(
"input_y_second").outputs[0]
input_y_third = graph.get_operation_by_name(
"input_y_third").outputs[0]
input_y_fourth = graph.get_operation_by_name(
"input_y_fourth").outputs[0]
input_y = graph.get_operation_by_name("input_y").outputs[0]
dropout_keep_prob = graph.get_operation_by_name(
"dropout_keep_prob").outputs[0]
alpha = graph.get_operation_by_name("alpha").outputs[0]
is_training = graph.get_operation_by_name("is_training").outputs[0]
# Tensors we want to evaluate
first_visual = graph.get_operation_by_name(
"first-output/visual").outputs[0]
second_visual = graph.get_operation_by_name(
"second-output/visual").outputs[0]
third_visual = graph.get_operation_by_name(
"third-output/visual").outputs[0]
fourth_visual = graph.get_operation_by_name(
"fourth-output/visual").outputs[0]
# Split the output nodes name by '|' if you have several output nodes
output_node_names = "first-output/visual|second-output/visual|third-output/visual|fourth-output/visual|output/scores"
# Save the .pb model file
output_graph_def = tf.graph_util.convert_variables_to_constants(
sess, sess.graph_def, output_node_names.split("|"))
tf.train.write_graph(output_graph_def,
"graph",
"graph-harnn-{0}.pb".format(MODEL),
as_text=False)
# Generate batches for one epoch
batches = dh.batch_iter(list(create_input_data(test_data)),
args.batch_size,
1,
shuffle=False)
for batch_id, batch_test in enumerate(batches):
x, x_content, sec, subsec, group, subgroup, y_onehot = zip(
*batch_test)
feed_dict = {
input_x: x,
input_y_first: sec,
input_y_second: subsec,
input_y_third: group,
input_y_fourth: subgroup,
input_y: y_onehot,
dropout_keep_prob: 1.0,
alpha: args.alpha,
is_training: False
}
batch_first_visual, batch_second_visual, batch_third_visual, batch_fourth_visual = \
sess.run([first_visual, second_visual, third_visual, fourth_visual], feed_dict)
batch_visual = [
batch_first_visual, batch_second_visual,
batch_third_visual, batch_fourth_visual
]
seq_len = len(x_content[0])
pad_len = len(batch_first_visual[0])
length = (pad_len if seq_len >= pad_len else seq_len)
visual_list = []
for visual in batch_visual:
visual_list.append(
normalization(visual[0].tolist(), length))
create_visual_file(batch_id, x_content, visual_list, seq_len)
logger.info("Done.")
def run_inference(img_path, output_dir, args):
""" A function making inference using the pre-trained darknet weights in the tensorflow
framework
Input:
img_path: string, path to the image on which inference is to be run, path to the image directory containing images in the case of multiple images.
output_dir: string, directory for saving the output
args: argparse object
"""
# Reading the images
if not os.path.exists(output_dir):
os.mkdir(output_dir)
if not os.path.exists(os.path.join(output_dir, 'images')):
os.mkdir(os.path.join(output_dir, 'images'))
if not os.path.exists(os.path.join(output_dir, 'labels')):
os.mkdir(os.path.join(output_dir, 'labels'))
output_dir_images = os.path.join(output_dir, 'images')
output_dir_labels = os.path.join(output_dir, 'labels')
file_names = sorted(os.listdir(img_path))
images_batch = read_image(img_path)
# Getting anchors and labels for the prediction
class_names = get_classes(config.classes_path)
anchors = read_anchors(config.anchors_path)
num_classes = config.num_classes
num_anchors = config.num_anchors
# Retriving the input shape of the model i.e. (608x608), (416x416), (320x320)
input_shape = (config.input_shape, config.input_shape)
# Generate colors for drawing bounding boxes.
hsv_tuples = [(x / len(class_names), 1., 1.)
for x in range(len(class_names))]
colors = list(map(lambda x: colorsys.hsv_to_rgb(*x), hsv_tuples))
colors = list(
map(lambda x: (int(x[0] * 255), int(x[1] * 255), int(x[2] * 255)),
colors))
random.seed(10101) # Fixed seed for consistent colors across runs.
random.shuffle(colors) # Shuffle colors to decorrelate adjacent classes.
random.seed(None) # Reset seed to default.
# Defining placeholder for passing the image data onto the model
image_tensor = tf.placeholder(dtype=tf.float32,
shape=[None, None, None, 3])
image_shape = tf.placeholder(dtype=tf.int32, shape=[2])
# Building the model for running inference
output_nodes = yolo(input_images=image_tensor,
is_training=False,
config_path=config.yolov3_cfg_path,
num_classes=num_classes)
# Creating a session for running the model
gpu_config = tf.ConfigProto(log_device_placement=False)
gpu_config.gpu_options.allow_growth = True
sess = tf.Session(config=gpu_config)
boxes, scores, classes = predict(output_nodes, anchors, num_classes,
input_shape, image_shape)
total_time_pred = []
total_time_yolo = []
for x in range(len(images_batch)):
image = images_batch[x]
new_image_size = (config.input_shape, config.input_shape)
image_data = np.array(resize_image(image, new_image_size))
print('Image height: {}\tImage width: {}'.format(
image.shape[0], image.shape[1]))
img = image_data / 255.
img = np.expand_dims(img, 0) # Adding the batch dimension
# Loading the model/weights for running the model
if x < 1:
if args.darknet_model is not None:
print('Loading pre-trained weights.....')
if not os.path.exists(config.yolov3_weights_path):
print('yolov3 weights not found.....\n')
if not os.path.exists('./yolov3.weights'):
os.system(
'wget https://pjreddie.com/media/files/yolov3.weights'
)
os.system(
'mv yolov3.weights ./darknet_data/yolov3.weights')
load_op = load_weights(tf.global_variables(),
weights_file=config.yolov3_weights_path)
sess.run(load_op)
else:
ckpt_path = config.model_dir
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(ckpt_path)
if ckpt and tf.train.checkpoint_exists(
ckpt.model_checkpoint_path):
print('Restoring model ',
checkmate.get_best_checkpoint(ckpt_path))
saver.restore(sess,
checkmate.get_best_checkpoint(ckpt_path))
print('Model Loaded!')
# tick = time()
# sess.run(output_nodes, feed_dict={image_tensor: img, image_shape: [image.shape[0], image.shape[1]]})
# tock = time()
# print("Prediction time: ", tock-tick)
# total_time_yolo.append(tock-tick)
tick = time()
# Actually run the graph in a tensorflow session to get the outputs
out_boxes, out_scores, out_classes = sess.run(
[boxes, scores, classes],
feed_dict={
image_tensor: img,
image_shape: [image.shape[0], image.shape[1]]
})
tock = time()
total_time_pred.append(tock - tick)
print('Found {} boxes for {} in {}sec'.format(len(out_boxes), 'img',
tock - tick))
######################## Visualization ######################
font = ImageFont.truetype(font='./font/FiraMono-Medium.otf',
size=np.floor(1e-2 * image.shape[1] +
0.5).astype(np.int32))
thickness = (image.shape[0] + image.shape[1]) // 1000 # do day cua BB
image = Image.fromarray((image).astype('uint8'), mode='RGB')
output_labels = open(
os.path.join(output_dir_labels,
file_names[x].split(',')[0] + '.txt'), 'w')
for i, c in reversed(list(enumerate(out_classes))):
predicted_class = class_names[c]
box = out_boxes[i]
score = out_scores[i]
label = '{} {:.2f}'.format(predicted_class, score)
draw = ImageDraw.Draw(image)
label_size = draw.textsize(label, font)
# print(label_size)
top, left, bottom, right = box # y_min, x_min, y_max, x_max
top = max(0, np.floor(top + 0.5).astype(np.int32))
left = max(0, np.floor(left + 0.5).astype(np.int32))
bottom = min(image.size[1],
np.floor(bottom + 0.5).astype(np.int32))
right = min(image.size[0], np.floor(right + 0.5).astype(np.int32))
print(label, (left, top),
(right, bottom)) # (x_min, y_min), (x_max, y_max)
output_labels.write(
str(left) + ',' + str(top) + ',' + str(right) + ',' +
str(bottom) + ',' + str(c) + ',' + str(score) + '\n')
if top - label_size[1] >= 0:
text_origin = np.array([left, top - label_size[1]])
else:
text_origin = np.array([left, top + 1])
# My kingdom for a good redistributable image drawing library.
for j in range(thickness):
draw.rectangle([left + j, top + j, right - j, bottom - j],
outline=colors[c])
draw.rectangle(
[tuple(text_origin),
tuple(text_origin + label_size)],
fill=colors[c])
draw.text(text_origin, label, fill=(0, 0, 0), font=font)
del draw
# image.show()
image.save(os.path.join(output_dir_images, file_names[x]),
compress_level=1)
output_labels.close()
sess.close()
total_time_pred = sum(total_time_pred[1:])
# total_time_yolo = sum(total_time_yolo[1:])
print('FPS of model with post processing over {} images is {}'.format(
len(images_batch) - 1, (len(images_batch) - 1) / total_time_pred))
def train(ckpt_path, log_path, class_path):
""" Function to train the model.
ckpt_path: string, path for saving/restoring the model
log_path: string, path for saving the training/validation logs
class_path: string, path for the classes of the dataset
decay_steps: int, steps after which the learning rate is to be decayed
decay_rate: float, rate to carrying out exponential decay
"""
# Getting the anchors
anchors = read_anchors(config.anchors_path)
classes = get_classes(class_path)
if anchors.shape[0] // 3 == 2:
yolo_tiny = True
else:
yolo_tiny = False
# Building the training pipeline
graph = tf.get_default_graph()
with graph.as_default():
# Getting the training data
with tf.name_scope('data_parser/'):
train_reader = Parser('train',
config.anchors_path,
config.output_dir,
config.num_classes,
input_shape=config.input_shape,
max_boxes=config.max_boxes)
train_data = train_reader.build_dataset(config.train_batch_size //
config.subdivisions)
train_iterator = train_data.make_one_shot_iterator()
val_reader = Parser('val',
config.anchors_path,
config.output_dir,
config.num_classes,
input_shape=config.input_shape,
max_boxes=config.max_boxes)
val_data = val_reader.build_dataset(config.val_batch_size //
config.subdivisions)
val_iterator = val_data.make_one_shot_iterator()
is_training = tf.placeholder(
dtype=tf.bool, shape=[], name='train_flag'
) # Used for different behaviour of batch normalization
mode = tf.placeholder(dtype=tf.int16, shape=[], name='mode_flag')
def train():
# images, bbox, bbox_true_13, bbox_true_26, bbox_true_52 = train_iterator.get_next()
return train_iterator.get_next()
def valid():
# images, bbox, bbox_true_13, bbox_true_26, bbox_true_52 = val_iterator.get_next()
return val_iterator.get_next()
if yolo_tiny:
images, bbox, bbox_true_13, bbox_true_26 = tf.cond(
pred=tf.equal(mode, 1),
true_fn=train,
false_fn=valid,
name='train_val__data')
grid_shapes = [
config.input_shape // 32, config.input_shape // 16
]
else:
images, bbox, bbox_true_13, bbox_true_26, bbox_true_52 = tf.cond(
pred=tf.equal(mode, 1),
true_fn=train,
false_fn=valid,
name='train_val_data')
grid_shapes = [
config.input_shape // 32, config.input_shape // 16,
config.input_shape // 8
]
images.set_shape([None, config.input_shape, config.input_shape, 3])
bbox.set_shape([None, config.max_boxes, 5])
# image_summary = draw_box(images, bbox)
# Extracting the pre-defined yolo graph from the darknet cfg file
if not os.path.exists(ckpt_path):
os.mkdir(ckpt_path)
output = yolo(images, is_training, config.yolov3_cfg_path,
config.num_classes)
# Declaring the parameters for GT
with tf.name_scope('Targets'):
if yolo_tiny:
bbox_true_13.set_shape([
None, grid_shapes[0], grid_shapes[0],
config.num_anchors_per_scale, 5 + config.num_classes
])
bbox_true_26.set_shape([
None, grid_shapes[1], grid_shapes[1],
config.num_anchors_per_scale, 5 + config.num_classes
])
y_true = [bbox_true_13, bbox_true_26]
else:
bbox_true_13.set_shape([
None, grid_shapes[0], grid_shapes[0],
config.num_anchors_per_scale, 5 + config.num_classes
])
bbox_true_26.set_shape([
None, grid_shapes[1], grid_shapes[1],
config.num_anchors_per_scale, 5 + config.num_classes
])
bbox_true_52.set_shape([
None, grid_shapes[2], grid_shapes[2],
config.num_anchors_per_scale, 5 + config.num_classes
])
y_true = [bbox_true_13, bbox_true_26, bbox_true_52]
# Compute Loss
with tf.name_scope('Loss_and_Detect'):
loss_scale, yolo_loss, xy_loss, wh_loss, obj_loss, noobj_loss, conf_loss, class_loss = compute_loss(
output,
y_true,
anchors,
config.num_classes,
config.input_shape,
ignore_threshold=config.ignore_thresh)
loss = yolo_loss
exponential_moving_average_op = tf.train.ExponentialMovingAverage(
config.weight_decay).apply(
var_list=tf.trainable_variables()) # For regularisation
scale1_loss_summary = tf.summary.scalar('scale_loss_1',
loss_scale[0],
family='Loss')
scale2_loss_summary = tf.summary.scalar('scale_loss_2',
loss_scale[1],
family='Loss')
yolo_loss_summary = tf.summary.scalar('yolo_loss',
yolo_loss,
family='Loss')
# total_loss_summary = tf.summary.scalar('Total_loss', loss, family='Loss')
xy_loss_summary = tf.summary.scalar('xy_loss',
xy_loss,
family='Loss')
wh_loss_summary = tf.summary.scalar('wh_loss',
wh_loss,
family='Loss')
obj_loss_summary = tf.summary.scalar('obj_loss',
obj_loss,
family='Loss')
noobj_loss_summary = tf.summary.scalar('noobj_loss',
noobj_loss,
family='Loss')
conf_loss_summary = tf.summary.scalar('confidence_loss',
conf_loss,
family='Loss')
class_loss_summary = tf.summary.scalar('class_loss',
class_loss,
family='Loss')
# Declaring the parameters for training the model
with tf.name_scope('train_parameters'):
global_step = tf.Variable(0, trainable=False, name='global_step')
def learning_rate_scheduler(learning_rate,
scheduler_name,
global_step,
decay_steps=100):
if scheduler_name == 'exponential':
lr = tf.train.exponential_decay(
learning_rate,
global_step,
decay_steps,
decay_rate,
staircase=True,
name='exponential_learning_rate')
return tf.maximum(lr, config.learning_rate_lower_bound)
elif scheduler_name == 'polynomial':
lr = tf.train.polynomial_decay(
learning_rate,
global_step,
decay_steps,
config.learning_rate_lower_bound,
power=0.8,
cycle=True,
name='polynomial_learning_rate')
return tf.maximum(lr, config.learning_rate_lower_bound)
elif scheduler_name == 'cosine':
lr = tf.train.cosine_decay(learning_rate,
global_step,
decay_steps,
alpha=0.5,
name='cosine_learning_rate')
return tf.maximum(lr, config.learning_rate_lower_bound)
elif scheduler_name == 'linear':
return tf.convert_to_tensor(learning_rate,
name='linear_learning_rate')
else:
raise ValueError(
'Unsupported learning rate scheduler\n[supported types: exponential, polynomial, linear]'
)
if config.use_warm_up:
learning_rate = tf.cond(
pred=tf.less(
global_step,
config.burn_in_epochs *
(config.train_num // config.train_batch_size)),
true_fn=lambda: learning_rate_scheduler(
config.init_learning_rate, config.warm_up_lr_scheduler,
global_step),
false_fn=lambda: learning_rate_scheduler(
config.learning_rate,
config.lr_scheduler,
global_step,
decay_steps=500))
else:
learning_rate = learning_rate_scheduler(config.learning_rate,
config.lr_scheduler,
global_step,
decay_steps=2000)
tf.summary.scalar('learning rate', learning_rate)
# Define optimizer for minimizing the computed loss
with tf.name_scope('Optimizer'):
# optimizer = tf.train.MomentumOptimizer(learning_rate=learning_rate, momentum=config.momentum)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
# optimizer = tf.train.RMSPropOptimizer(learning_rate=learning_rate, momentum=config.momentum)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
if config.pre_train:
train_vars = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope='yolo')
else:
train_vars = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES)
grads = optimizer.compute_gradients(loss=loss,
var_list=train_vars)
gradients = [(tf.placeholder(dtype=tf.float32,
shape=grad[1].get_shape()),
grad[1]) for grad in grads]
optimizing_op = optimizer.apply_gradients(
grads_and_vars=gradients, global_step=global_step)
# optimizing_op = optimizer.minimize(loss=loss, global_step=global_step)
with tf.control_dependencies([optimizing_op]):
with tf.control_dependencies([exponential_moving_average_op]):
train_op_with_mve = tf.no_op()
train_op = train_op_with_mve
#################################### Training loop ############################################################
# A saver object for saving the model
best_ckpt_saver_train = checkmate.BestCheckpointSaver(
save_dir=ckpt_path + 'train/', num_to_keep=5)
best_ckpt_saver_valid = checkmate.BestCheckpointSaver(
save_dir=ckpt_path + 'valid/', num_to_keep=5)
summary_op = tf.summary.merge_all()
summary_op_valid = tf.summary.merge([
yolo_loss_summary, xy_loss_summary, wh_loss_summary,
obj_loss_summary, noobj_loss_summary, conf_loss_summary,
class_loss_summary, scale1_loss_summary, scale2_loss_summary
])
init_op = tf.global_variables_initializer()
# Defining some train loop dependencies
gpu_config = tf.ConfigProto(log_device_placement=False)
gpu_config.gpu_options.allow_growth = True
sess = tf.Session(config=gpu_config)
tf.logging.set_verbosity(tf.logging.ERROR)
train_summary_writer = tf.summary.FileWriter(
os.path.join(log_path, 'train'), sess.graph)
val_summary_writer = tf.summary.FileWriter(
os.path.join(log_path, 'val'), sess.graph)
# Restoring the model
ckpt = tf.train.get_checkpoint_state(ckpt_path + 'valid/')
if ckpt and tf.train.checkpoint_exists(ckpt.model_checkpoint_path):
print('Restoring model ',
checkmate.get_best_checkpoint(ckpt_path + 'valid/'))
tf.train.Saver().restore(
sess, checkmate.get_best_checkpoint(ckpt_path + 'valid/'))
print('Model Loaded!')
elif config.pre_train is True:
sess.run(init_op)
load_ops = load_weights(tf.global_variables(scope='darknet53'),
config.darknet53_weights_path)
sess.run(load_ops)
else:
sess.run(init_op)
print('Uninitialized variables: ',
sess.run(tf.report_uninitialized_variables()))
epochbar = tqdm(range(config.Epoch))
for epoch in epochbar:
epochbar.set_description('Epoch %s of %s' % (epoch, config.Epoch))
mean_loss_train = []
mean_loss_valid = []
trainbar = tqdm(range(config.train_num // config.train_batch_size))
for k in trainbar:
all_grads_and_vars = []
for minibatch in range(config.train_batch_size //
config.subdivisions):
num_steps, train_summary, loss_train, grads_and_vars = sess.run(
[global_step, summary_op, loss, grads],
feed_dict={
is_training: True,
mode: 1
})
all_grads_and_vars += grads_and_vars
train_summary_writer.add_summary(train_summary, epoch)
train_summary_writer.flush()
mean_loss_train.append(loss_train)
trainbar.set_description('Train loss: %s' %
str(loss_train))
feed_dict = {is_training: True, mode: 1}
for i in range(len(gradients), len(all_grads_and_vars)):
all_grads_and_vars[
i % len(gradients)] += all_grads_and_vars[i][0]
all_grads_and_vars = all_grads_and_vars[:len(gradients)]
for i in range(len(gradients)):
feed_dict[gradients[i][0]] = all_grads_and_vars[i][0]
# print(np.shape(feed_dict))
_ = sess.run(train_op, feed_dict=feed_dict)
print('Validating.....')
valbar = tqdm(range(config.val_num // config.val_batch_size))
for k in valbar:
for minibatch in range(config.train_batch_size //
config.subdivisions):
val_summary, loss_valid = sess.run(
[summary_op_valid, loss],
feed_dict={
is_training: False,
mode: 0
})
val_summary_writer.add_summary(val_summary, epoch)
val_summary_writer.flush()
mean_loss_valid.append(loss_valid)
valbar.set_description('Validation loss: %s' %
str(loss_valid))
mean_loss_train = np.mean(mean_loss_train)
mean_loss_valid = np.mean(mean_loss_valid)
print('\n')
print('Train loss after %d epochs is: %f' %
(epoch + 1, mean_loss_train))
print('Validation loss after %d epochs is: %f' %
(epoch + 1, mean_loss_valid))
print('\n\n')
if (config.use_warm_up):
if (num_steps > config.burn_in_epochs *
(config.train_num // config.train_batch_size)):
best_ckpt_saver_train.handle(mean_loss_train, sess,
global_step)
best_ckpt_saver_valid.handle(mean_loss_valid, sess,
global_step)
else:
continue
else:
best_ckpt_saver_train.handle(mean_loss_train, sess,
global_step)
best_ckpt_saver_valid.handle(mean_loss_valid, sess,
global_step)
print('Tuning Completed!!')
train_summary_writer.close()
val_summary_writer.close()
sess.close()
def train(ckpt_path, log_path, class_path):
""" Function to train the model.
ckpt_path: string, path for saving/restoring the model
log_path: string, path for saving the training/validation logs
class_path: string, path for the classes of the dataset
decay_steps: int, steps after which the learning rate is to be decayed
decay_rate: float, rate to carrying out exponential decay
"""
# Getting the anchors
anchors = read_anchors(config.anchors_path)
if not os.path.exists(config.data_dir):
os.mkdir(config.data_dir)
classes = get_classes(class_path)
# Building the training pipeline
graph = tf.get_default_graph()
with graph.as_default():
# Getting the training data
with tf.name_scope('data_parser/'):
train_reader = Parser('train', config.data_dir, config.anchors_path, config.output_dir,
config.num_classes, input_shape=config.input_shape, max_boxes=config.max_boxes)
train_data = train_reader.build_dataset(config.train_batch_size//config.subdivisions)
train_iterator = train_data.make_one_shot_iterator()
val_reader = Parser('val', config.data_dir, config.anchors_path, config.output_dir,
config.num_classes, input_shape=config.input_shape, max_boxes=config.max_boxes)
val_data = val_reader.build_dataset(config.val_batch_size)
val_iterator = val_data.make_one_shot_iterator()
is_training = tf.placeholder(dtype=tf.bool, shape=[], name='train_flag') # Used for different behaviour of batch normalization
mode = tf.placeholder(dtype=tf.int16, shape=[], name='mode_flag')
def train():
return train_iterator.get_next()
def valid():
return val_iterator.get_next()
images, labels = tf.cond(pred=tf.equal(mode, 1), true_fn=train, false_fn=valid, name='train_val_data')
grid_shapes = [config.input_shape // 32, config.input_shape // 16, config.input_shape // 8]
images.set_shape([None, config.input_shape, config.input_shape, 3])
labels.set_shape([None, required_shape, 5])
# image_summary = draw_box(images, bbox, file_name)
if not os.path.exists(ckpt_path):
os.mkdir(ckpt_path)
model = model(images, is_training, config.num_classes, config.num_anchors_per_scale, config.weight_decay, config.norm_decay)
output, model_layers = model.forward()
print('Summary of the created model.......\n')
for layer in model_layers:
print(layer)
# Declaring the parameters for GT
with tf.name_scope('Targets'):
### GT PROCESSING ###
# Compute Loss
with tf.name_scope('Loss_and_Detect'):
loss_scale,summaries = compute_loss(output, y_true, config.num_classes, ignore_threshold=config.ignore_thresh)
exponential_moving_average_op = tf.train.ExponentialMovingAverage(config.weight_decay).apply(var_list=tf.trainable_variables())
loss = model_loss
model_loss_summary = tf.summary.scalar('model_loss', summaries, family='Losses')
# Declaring the parameters for training the model
with tf.name_scope('train_parameters'):
global_step = tf.Variable(0, trainable=False, name='global_step')
# Declaring the parameters for training the model
with tf.name_scope('train_parameters'):
global_step = tf.Variable(0, trainable=False, name='global_step')
def learning_rate_scheduler(learning_rate, scheduler_name, global_step, decay_steps=100):
if scheduler_name == 'exponential':
lr = tf.train.exponential_decay(learning_rate, global_step,
decay_steps, decay_rate, staircase=True, name='exponential_learning_rate')
return tf.maximum(lr, config.learning_rate_lower_bound)
elif scheduler_name == 'polynomial':
lr = tf.train.polynomial_decay(learning_rate, global_step,
decay_steps, config.learning_rate_lower_bound, power=0.8, cycle=True, name='polynomial_learning_rate')
return tf.maximum(lr, config.learning_rate_lower_bound)
elif scheduler_name == 'cosine':
lr = tf.train.cosine_decay(learning_rate, global_step,
decay_steps, alpha=0.5, name='cosine_learning_rate')
return tf.maximum(lr, config.learning_rate_lower_bound)
elif scheduler_name == 'linear':
return tf.convert_to_tensor(learning_rate, name='linear_learning_rate')
else:
raise ValueError('Unsupported learning rate scheduler\n[supported types: exponential, polynomial, linear]')
if config.use_warm_up:
learning_rate = tf.cond(pred=tf.less(global_step, config.burn_in_epochs * (config.train_num // config.train_batch_size)),
true_fn=lambda: learning_rate_scheduler(config.init_learning_rate, config.warm_up_lr_scheduler, global_step),
false_fn=lambda: learning_rate_scheduler(config.learning_rate, config.lr_scheduler, global_step, decay_steps=2000))
else:
learning_rate = learning_rate_scheduler(config.learning_rate, config.lr_scheduler, global_step=global_step, decay_steps=2000)
tf.summary.scalar('learning rate', learning_rate, family='Train_Parameters')
# Define optimizer for minimizing the computed loss
with tf.name_scope('Optimizer'):
optimizer = tf.train.MomentumOptimizer(learning_rate=learning_rate, momentum=config.momentum)
# optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
# optimizer = tf.train.RMSPropOptimizer(learning_rate=learning_rate, momentum=config.momentum)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
# grads = optimizer.compute_gradients(loss=loss)
# gradients = [(tf.placeholder(dtype=tf.float32, shape=grad[1].get_shape()), grad[1]) for grad in grads]
# train_step = optimizer.apply_gradients(grads_and_vars=gradients, global_step=global_step)
optimizing_op = optimizer.minimize(loss=loss, global_step=global_step)
with tf.control_dependencies([optimizing_op]):
with tf.control_dependencies([exponential_moving_average_op]):
train_op_with_mve = tf.no_op()
train_op = train_op_with_mve
#################################### Training loop ############################################################
# A saver object for saving the model
best_ckpt_saver_train = checkmate.BestCheckpointSaver(save_dir=ckpt_path+'train/', num_to_keep=5)
best_ckpt_saver_valid = checkmate.BestCheckpointSaver(save_dir=ckpt_path+'valid/', num_to_keep=5)
summary_op = tf.summary.merge_all()
summary_op_valid = tf.summary.merge([model_loss_summary_without_learning_rate])
init_op = tf.global_variables_initializer()
# Defining some train loop dependencies
gpu_config = tf.ConfigProto(log_device_placement=False)
gpu_config.gpu_options.allow_growth = True
sess = tf.Session(config=gpu_config)
tf.logging.set_verbosity(tf.logging.ERROR)
train_summary_writer = tf.summary.FileWriter(os.path.join(log_path, 'train'), sess.graph)
val_summary_writer = tf.summary.FileWriter(os.path.join(log_path, 'val'), sess.graph)
print(sess.run(receptive_field))
# Restoring the model
ckpt = tf.train.get_checkpoint_state(ckpt_path+'train/')
if ckpt and tf.train.checkpoint_exists(ckpt.model_checkpoint_path):
print('Restoring model ', checkmate.get_best_checkpoint(ckpt_path+'train/'))
tf.train.Saver().restore(sess, checkmate.get_best_checkpoint(ckpt_path+'train/'))
print('Model Loaded!')
else:
sess.run(init_op)
print('Uninitialized variables: ', sess.run(tf.report_uninitialized_variables()))
epochbar = tqdm(range(config.Epoch))
for epoch in epochbar:
epochbar.set_description('Epoch %s of %s' % (epoch, config.Epoch))
mean_loss_train = []
mean_loss_valid = []
trainbar = tqdm(range(config.train_num//config.train_batch_size))
for k in trainbar:
num_steps, train_summary, loss_train, _ = sess.run([global_step, summary_op, loss,
train_op], feed_dict={is_training: True, mode: 1})
train_summary_writer.add_summary(train_summary, epoch)
train_summary_writer.flush()
mean_loss_train.append(loss_train)
trainbar.set_description('Train loss: %s' %str(loss_train))
print('Validating.....')
valbar = tqdm(range(config.val_num//config.val_batch_size))
for k in valbar:
val_summary, loss_valid = sess.run([summary_op_valid, loss], feed_dict={is_training: False, mode: 0})
val_summary_writer.add_summary(val_summary, epoch)
val_summary_writer.flush()
mean_loss_valid.append(loss_valid)
valbar.set_description('Validation loss: %s' %str(loss_valid))
mean_loss_train = np.mean(mean_loss_train)
mean_loss_valid = np.mean(mean_loss_valid)
print('\n')
print('Train loss after %d epochs is: %f' %(epoch+1, mean_loss_train))
print('Validation loss after %d epochs is: %f' %(epoch+1, mean_loss_valid))
print('\n\n')
if (config.use_warm_up):
if (num_steps > config.burn_in_epochs * (config.train_num // config.train_batch_size)):
best_ckpt_saver_train.handle(mean_loss_train, sess, global_step)
best_ckpt_saver_valid.handle(mean_loss_valid, sess, global_step)
else:
continue
else:
best_ckpt_saver_train.handle(mean_loss_train, sess, global_step)
best_ckpt_saver_valid.handle(mean_loss_valid, sess, global_step)
print('Tuning Completed!!')
train_summary_writer.close()
val_summary_writer.close()
sess.close()
def main():
""" main function which calls all the other required functions for training """
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = str(config.gpu_num)
train(config.model_dir, config.logs_dir, config.classes_path)
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
if __name__ == '__main__':
main()
def test_harnn():
"""Test HARNN model."""
# Load data
logger.info("✔︎ Loading data...")
logger.info("Recommended padding Sequence length is: {0}".format(
FLAGS.pad_seq_len))
logger.info("✔︎ Test data processing...")
test_data = dh.load_data_and_labels(FLAGS.test_data_file,
FLAGS.num_classes_list,
FLAGS.total_classes,
FLAGS.embedding_dim,
data_aug_flag=False)
logger.info("✔︎ Test data padding...")
x_test, y_test, y_test_tuple = dh.pad_data(test_data, FLAGS.pad_seq_len)
y_test_labels = test_data.labels
# Load harnn model
BEST_OR_LATEST = input("☛ Load Best or Latest Model?(B/L): ")
while not (BEST_OR_LATEST.isalpha()
and BEST_OR_LATEST.upper() in ['B', 'L']):
BEST_OR_LATEST = input(
"✘ The format of your input is illegal, please re-input: ")
if BEST_OR_LATEST.upper() == 'B':
logger.info("✔︎ Loading best model...")
checkpoint_file = cm.get_best_checkpoint(FLAGS.best_checkpoint_dir,
select_maximum_value=True)
else:
logger.info("✔︎ Loading latest model...")
checkpoint_file = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
logger.info(checkpoint_file)
graph = tf.Graph()
with graph.as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=FLAGS.allow_soft_placement,
log_device_placement=FLAGS.log_device_placement)
session_conf.gpu_options.allow_growth = FLAGS.gpu_options_allow_growth
sess = tf.Session(config=session_conf)
with sess.as_default():
# Load the saved meta graph and restore variables
saver = tf.train.import_meta_graph(
"{0}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
# Get the placeholders from the graph by name
input_x = graph.get_operation_by_name("input_x").outputs[0]
input_y_first = graph.get_operation_by_name(
"input_y_first").outputs[0]
input_y_second = graph.get_operation_by_name(
"input_y_second").outputs[0]
input_y_third = graph.get_operation_by_name(
"input_y_third").outputs[0]
input_y = graph.get_operation_by_name("input_y").outputs[0]
dropout_keep_prob = graph.get_operation_by_name(
"dropout_keep_prob").outputs[0]
beta = graph.get_operation_by_name("beta").outputs[0]
is_training = graph.get_operation_by_name("is_training").outputs[0]
# Tensors we want to evaluate
first_attention = graph.get_operation_by_name(
"first-attention/attention").outputs[0]
first_visual = graph.get_operation_by_name(
"first-output/visual").outputs[0]
second_visual = graph.get_operation_by_name(
"second-output/visual").outputs[0]
third_visual = graph.get_operation_by_name(
"third-output/visual").outputs[0]
scores = graph.get_operation_by_name("output/scores").outputs[0]
loss = graph.get_operation_by_name("loss/loss").outputs[0]
# Split the output nodes name by '|' if you have several output nodes
output_node_names = "first-output/scores|second-output/scores|third-output/scores|output/scores"
# Save the .pb model file
output_graph_def = tf.graph_util.convert_variables_to_constants(
sess, sess.graph_def, output_node_names.split("|"))
tf.train.write_graph(output_graph_def,
"graph",
"graph-harnn-{0}.pb".format(MODEL),
as_text=False)
# Generate batches for one epoch
batches = dh.batch_iter(list(
zip(x_test, y_test, y_test_tuple, y_test_labels)),
FLAGS.batch_size,
1,
shuffle=False)
test_counter, test_loss = 0, 0.0
# Collection
true_labels = []
predicted_labels = []
predicted_scores = []
# Collect for calculating metrics
true_onehot_labels = []
predicted_onehot_scores = []
predicted_onehot_labels_ts = []
predicted_onehot_labels_tk = [[] for _ in range(FLAGS.top_num)]
for batch_test in batches:
x_batch_test, y_batch_test, y_batch_test_tuple, y_batch_test_labels = zip(
*batch_test)
y_batch_test_first = [i[0] for i in y_batch_test_tuple]
y_batch_test_second = [j[1] for j in y_batch_test_tuple]
y_batch_test_third = [k[2] for k in y_batch_test_tuple]
feed_dict = {
input_x: x_batch_test,
input_y_first: y_batch_test_first,
input_y_second: y_batch_test_second,
input_y_third: y_batch_test_third,
input_y: y_batch_test,
dropout_keep_prob: 1.0,
beta: FLAGS.beta,
is_training: False
}
batch_first_attention, batch_first_visual, batch_second_visual, batch_third_visual, batch_scores, cur_loss = \
sess.run([first_attention, first_visual, second_visual, third_visual, scores, loss], feed_dict)
print(batch_first_visual)
print(batch_first_visual[0])
print(batch_first_visual[1])
print(batch_first_visual[0, :15])
f = open('attention.html', 'w')
f.write(
'<html style="margin:0;padding:0;"><body style="margin:0;padding:0;">\n'
)
f.write('<div style="margin:25px;">\n')
for k in range(len(batch_first_attention[0])):
f.write('<p style="margin:10px;">\n')
for i in range(len(batch_first_attention[0][0])):
alpha = "{:.2f}".format(batch_first_attention[0][k][i])
word = x_batch_test[i]
f.write(
f'\t<span style="margin-left:3px;background-color:rgba(255,0,0,{alpha})">{word}</span>\n'
)
f.write('</p>\n')
f.write('</div>\n')
f.write('</body></html>')
f.close()
# Prepare for calculating metrics
for onehot_labels in y_batch_test:
true_onehot_labels.append(onehot_labels)
for onehot_scores in batch_scores:
predicted_onehot_scores.append(onehot_scores)
# Get the predicted labels by threshold
batch_predicted_labels_ts, batch_predicted_scores_ts = \
dh.get_label_threshold(scores=batch_scores, threshold=FLAGS.threshold)
# Add results to collection
for labels in y_batch_test_labels:
true_labels.append(labels)
for labels in batch_predicted_labels_ts:
predicted_labels.append(labels)
for values in batch_predicted_scores_ts:
predicted_scores.append(values)
# Get one-hot prediction by threshold
batch_predicted_onehot_labels_ts = \
dh.get_onehot_label_threshold(scores=batch_scores, threshold=FLAGS.threshold)
for onehot_labels in batch_predicted_onehot_labels_ts:
predicted_onehot_labels_ts.append(onehot_labels)
# Get one-hot prediction by topK
for i in range(FLAGS.top_num):
batch_predicted_onehot_labels_tk = dh.get_onehot_label_topk(
scores=batch_scores, top_num=i + 1)
for onehot_labels in batch_predicted_onehot_labels_tk:
predicted_onehot_labels_tk[i].append(onehot_labels)
test_loss = test_loss + cur_loss
test_counter = test_counter + 1
# Calculate Precision & Recall & F1
test_pre_ts = precision_score(
y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_ts),
average='micro')
test_rec_ts = recall_score(
y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_ts),
average='micro')
test_F_ts = f1_score(y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_ts),
average='micro')
# Calculate the average AUC
test_auc = roc_auc_score(y_true=np.array(true_onehot_labels),
y_score=np.array(predicted_onehot_scores),
average='micro')
# Calculate the average PR
test_prc = average_precision_score(
y_true=np.array(true_onehot_labels),
y_score=np.array(predicted_onehot_scores),
average="micro")
test_loss = float(test_loss / test_counter)
logger.info(
"☛ All Test Dataset: Loss {0:g} | AUC {1:g} | AUPRC {2:g}".
format(test_loss, test_auc, test_prc))
# Predict by threshold
logger.info(
"☛ Predict by threshold: Precision {0:g}, Recall {1:g}, F1 {2:g}"
.format(test_pre_ts, test_rec_ts, test_F_ts))
# Save the prediction result
if not os.path.exists(SAVE_DIR):
os.makedirs(SAVE_DIR)
dh.create_prediction_file(output_file=SAVE_DIR +
"/predictions.json",
data_id=test_data.patent_id,
all_labels=true_labels,
all_predict_labels=predicted_labels,
all_predict_scores=predicted_scores)
logger.info("✔︎ Done.")
def test_fasttext():
"""Test FASTTEXT model."""
# Print parameters used for the model
dh.tab_printer(args, logger)
# Load data
logger.info("Loading data...")
logger.info("Data processing...")
test_data = dh.load_data_and_labels(args.test_file,
args.num_classes,
args.word2vec_file,
data_aug_flag=False)
logger.info("Data padding...")
x_test, y_test = dh.pad_data(test_data, args.pad_seq_len)
y_test_labels = test_data.labels
# Load fasttext model
OPTION = dh._option(pattern=1)
if OPTION == 'B':
logger.info("Loading best model...")
checkpoint_file = cm.get_best_checkpoint(BEST_CPT_DIR,
select_maximum_value=True)
else:
logger.info("Loading latest model...")
checkpoint_file = tf.train.latest_checkpoint(CPT_DIR)
logger.info(checkpoint_file)
graph = tf.Graph()
with graph.as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=args.allow_soft_placement,
log_device_placement=args.log_device_placement)
session_conf.gpu_options.allow_growth = args.gpu_options_allow_growth
sess = tf.Session(config=session_conf)
with sess.as_default():
# Load the saved meta graph and restore variables
saver = tf.train.import_meta_graph(
"{0}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
# Get the placeholders from the graph by name
input_x = graph.get_operation_by_name("input_x").outputs[0]
input_y = graph.get_operation_by_name("input_y").outputs[0]
dropout_keep_prob = graph.get_operation_by_name(
"dropout_keep_prob").outputs[0]
is_training = graph.get_operation_by_name("is_training").outputs[0]
# Tensors we want to evaluate
scores = graph.get_operation_by_name("output/scores").outputs[0]
loss = graph.get_operation_by_name("loss/loss").outputs[0]
# Split the output nodes name by '|' if you have several output nodes
output_node_names = "output/scores"
# Save the .pb model file
output_graph_def = tf.graph_util.convert_variables_to_constants(
sess, sess.graph_def, output_node_names.split("|"))
tf.train.write_graph(output_graph_def,
"graph",
"graph-fasttext-{0}.pb".format(MODEL),
as_text=False)
# Generate batches for one epoch
batches = dh.batch_iter(list(zip(x_test, y_test, y_test_labels)),
args.batch_size,
1,
shuffle=False)
test_counter, test_loss = 0, 0.0
test_pre_tk = [0.0] * args.topK
test_rec_tk = [0.0] * args.topK
test_F1_tk = [0.0] * args.topK
# Collect the predictions here
true_labels = []
predicted_labels = []
predicted_scores = []
# Collect for calculating metrics
true_onehot_labels = []
predicted_onehot_scores = []
predicted_onehot_labels_ts = []
predicted_onehot_labels_tk = [[] for _ in range(args.topK)]
for batch_test in batches:
x_batch_test, y_batch_test, y_batch_test_labels = zip(
*batch_test)
feed_dict = {
input_x: x_batch_test,
input_y: y_batch_test,
dropout_keep_prob: 1.0,
is_training: False
}
batch_scores, cur_loss = sess.run([scores, loss], feed_dict)
# Prepare for calculating metrics
for i in y_batch_test:
true_onehot_labels.append(i)
for j in batch_scores:
predicted_onehot_scores.append(j)
# Get the predicted labels by threshold
batch_predicted_labels_ts, batch_predicted_scores_ts = \
dh.get_label_threshold(scores=batch_scores, threshold=args.threshold)
# Add results to collection
for i in y_batch_test_labels:
true_labels.append(i)
for j in batch_predicted_labels_ts:
predicted_labels.append(j)
for k in batch_predicted_scores_ts:
predicted_scores.append(k)
# Get onehot predictions by threshold
batch_predicted_onehot_labels_ts = \
dh.get_onehot_label_threshold(scores=batch_scores, threshold=args.threshold)
for i in batch_predicted_onehot_labels_ts:
predicted_onehot_labels_ts.append(i)
# Get onehot predictions by topK
for top_num in range(args.topK):
batch_predicted_onehot_labels_tk = dh.get_onehot_label_topk(
scores=batch_scores, top_num=top_num + 1)
for i in batch_predicted_onehot_labels_tk:
predicted_onehot_labels_tk[top_num].append(i)
test_loss = test_loss + cur_loss
test_counter = test_counter + 1
# Calculate Precision & Recall & F1
test_pre_ts = precision_score(
y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_ts),
average='micro')
test_rec_ts = recall_score(
y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_ts),
average='micro')
test_F1_ts = f1_score(y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_ts),
average='micro')
for top_num in range(args.topK):
test_pre_tk[top_num] = precision_score(
y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_tk[top_num]),
average='micro')
test_rec_tk[top_num] = recall_score(
y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_tk[top_num]),
average='micro')
test_F1_tk[top_num] = f1_score(
y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_tk[top_num]),
average='micro')
# Calculate the average AUC
test_auc = roc_auc_score(y_true=np.array(true_onehot_labels),
y_score=np.array(predicted_onehot_scores),
average='micro')
# Calculate the average PR
test_prc = average_precision_score(
y_true=np.array(true_onehot_labels),
y_score=np.array(predicted_onehot_scores),
average="micro")
test_loss = float(test_loss / test_counter)
logger.info(
"All Test Dataset: Loss {0:g} | AUC {1:g} | AUPRC {2:g}".
format(test_loss, test_auc, test_prc))
# Predict by threshold
logger.info(
"Predict by threshold: Precision {0:g}, Recall {1:g}, F1 {2:g}"
.format(test_pre_ts, test_rec_ts, test_F1_ts))
# Predict by topK
logger.info("Predict by topK:")
for top_num in range(args.topK):
logger.info(
"Top{0}: Precision {1:g}, Recall {2:g}, F1 {3:g}".format(
top_num + 1, test_pre_tk[top_num],
test_rec_tk[top_num], test_F1_tk[top_num]))
# Save the prediction result
if not os.path.exists(SAVE_DIR):
os.makedirs(SAVE_DIR)
dh.create_prediction_file(output_file=SAVE_DIR +
"/predictions.json",
data_id=test_data.testid,
all_labels=true_labels,
all_predict_labels=predicted_labels,
all_predict_scores=predicted_scores)
logger.info("All Done.")
def visualize():
"""visualize HARNN model."""
# Load data
logger.info("✔︎ Loading data...")
logger.info("Recommended padding Sequence length is: {0}".format(FLAGS.pad_seq_len))
logger.info("✔︎ Test data processing...")
test_data = dh.load_data_and_labels(FLAGS.test_data_file, FLAGS.num_classes_list, FLAGS.total_classes,
FLAGS.embedding_dim, data_aug_flag=False)
logger.info("✔︎ Test data padding...")
x_test, y_test, y_test_tuple = dh.pad_data(test_data, FLAGS.pad_seq_len)
x_test_content, y_test_labels = test_data.abstract_content, test_data.labels
# Load harnn model
BEST_OR_LATEST = input("☛ Load Best or Latest Model?(B/L): ")
while not (BEST_OR_LATEST.isalpha() and BEST_OR_LATEST.upper() in ['B', 'L']):
BEST_OR_LATEST = input("✘ The format of your input is illegal, please re-input: ")
if BEST_OR_LATEST.upper() == 'B':
logger.info("✔︎ Loading best model...")
checkpoint_file = cm.get_best_checkpoint(FLAGS.best_checkpoint_dir, select_maximum_value=True)
else:
logger.info("✔︎ Loading latest model...")
checkpoint_file = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
logger.info(checkpoint_file)
graph = tf.Graph()
with graph.as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=FLAGS.allow_soft_placement,
log_device_placement=FLAGS.log_device_placement)
session_conf.gpu_options.allow_growth = FLAGS.gpu_options_allow_growth
sess = tf.Session(config=session_conf)
with sess.as_default():
# Load the saved meta graph and restore variables
saver = tf.train.import_meta_graph("{0}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
# Get the placeholders from the graph by name
input_x = graph.get_operation_by_name("input_x").outputs[0]
input_y_first = graph.get_operation_by_name("input_y_first").outputs[0]
input_y_second = graph.get_operation_by_name("input_y_second").outputs[0]
input_y_third = graph.get_operation_by_name("input_y_third").outputs[0]
input_y = graph.get_operation_by_name("input_y").outputs[0]
dropout_keep_prob = graph.get_operation_by_name("dropout_keep_prob").outputs[0]
beta = graph.get_operation_by_name("beta").outputs[0]
is_training = graph.get_operation_by_name("is_training").outputs[0]
# Tensors we want to evaluate
first_visual = graph.get_operation_by_name("first-output/visual").outputs[0]
second_visual = graph.get_operation_by_name("second-output/visual").outputs[0]
third_visual = graph.get_operation_by_name("third-output/visual").outputs[0]
scores = graph.get_operation_by_name("output/scores").outputs[0]
# Split the output nodes name by '|' if you have several output nodes
output_node_names = "first-output/visual|second-output/visual|third-output/visual|output/scores"
# Save the .pb model file
output_graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def,
output_node_names.split("|"))
tf.train.write_graph(output_graph_def, "graph", "graph-harnn-{0}.pb".format(MODEL), as_text=False)
# Generate batches for one epoch
batches = dh.batch_iter(list(zip(x_test, y_test, y_test_tuple, x_test_content, y_test_labels)),
FLAGS.batch_size, 1, shuffle=False)
for batch_test in batches:
x_batch_test, y_batch_test, y_batch_test_tuple, \
x_batch_test_content, y_batch_test_labels = zip(*batch_test)
y_batch_test_first = [i[0] for i in y_batch_test_tuple]
y_batch_test_second = [j[1] for j in y_batch_test_tuple]
y_batch_test_third = [k[2] for k in y_batch_test_tuple]
feed_dict = {
input_x: x_batch_test,
input_y_first: y_batch_test_first,
input_y_second: y_batch_test_second,
input_y_third: y_batch_test_third,
input_y: y_batch_test,
dropout_keep_prob: 1.0,
beta: FLAGS.beta,
is_training: False
}
batch_first_visual, batch_second_visual, batch_third_visual, batch_scores = \
sess.run([first_visual, second_visual, third_visual, scores], feed_dict)
seq_len = len(x_batch_test_content[0])
pad_len = len(batch_first_visual[0])
if seq_len >= pad_len:
length = pad_len
else:
length = seq_len
# print(seq_len, pad_len, length)
final_first_visual = normalization(batch_first_visual[0].tolist(), length)
final_second_visual = normalization(batch_second_visual[0].tolist(), length)
final_third_visual = normalization(batch_third_visual[0].tolist(), length)
visual_list = [final_first_visual, final_second_visual, final_third_visual]
print(visual_list)
f = open('attention.html', 'w')
f.write('<html style="margin:0;padding:0;"><body style="margin:0;padding:0;">\n')
f.write('<div style="margin:25px;">\n')
for k in range(len(visual_list)):
f.write('<p style="margin:10px;">\n')
for i in range(seq_len):
alpha = "{:.2f}".format(visual_list[k][i])
word = x_batch_test_content[0][i]
f.write('\t<span style="margin-left:3px;background-color:rgba(255,0,0,{0})">{1}</span>\n'
.format(alpha, word))
f.write('</p>\n')
f.write('</div>\n')
f.write('</body></html>')
f.close()
logger.info("✔︎ Done.")
def visualize():
"""Visualize HARNN model."""
# Load data
logger.info("Loading data...")
logger.info("Data processing...")
test_data = dh.load_data_and_labels(args.test_file,
args.num_classes_list,
args.total_classes,
args.word2vec_file,
data_aug_flag=False)
logger.info("Data padding...")
x_test, y_test, y_test_tuple = dh.pad_data(test_data, args.pad_seq_len)
x_test_content, y_test_labels = test_data.abstract_content, test_data.labels
# Load harnn model
OPTION = dh._option(pattern=1)
if OPTION == 'B':
logger.info("Loading best model...")
checkpoint_file = cm.get_best_checkpoint(BEST_CPT_DIR,
select_maximum_value=True)
else:
logger.info("Loading latest model...")
checkpoint_file = tf.train.latest_checkpoint(CPT_DIR)
logger.info(checkpoint_file)
graph = tf.Graph()
with graph.as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=args.allow_soft_placement,
log_device_placement=args.log_device_placement)
session_conf.gpu_options.allow_growth = args.gpu_options_allow_growth
sess = tf.Session(config=session_conf)
with sess.as_default():
# Load the saved meta graph and restore variables
saver = tf.train.import_meta_graph(
"{0}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
# Get the placeholders from the graph by name
input_x = graph.get_operation_by_name("input_x").outputs[0]
input_y_first = graph.get_operation_by_name(
"input_y_first").outputs[0]
input_y_second = graph.get_operation_by_name(
"input_y_second").outputs[0]
input_y_third = graph.get_operation_by_name(
"input_y_third").outputs[0]
input_y_fourth = graph.get_operation_by_name(
"input_y_fourth").outputs[0]
input_y = graph.get_operation_by_name("input_y").outputs[0]
dropout_keep_prob = graph.get_operation_by_name(
"dropout_keep_prob").outputs[0]
alpha = graph.get_operation_by_name("alpha").outputs[0]
is_training = graph.get_operation_by_name("is_training").outputs[0]
# Tensors we want to evaluate
first_visual = graph.get_operation_by_name(
"first-output/visual").outputs[0]
second_visual = graph.get_operation_by_name(
"second-output/visual").outputs[0]
third_visual = graph.get_operation_by_name(
"third-output/visual").outputs[0]
fourth_visual = graph.get_operation_by_name(
"fourth-output/visual").outputs[0]
scores = graph.get_operation_by_name("output/scores").outputs[0]
# Split the output nodes name by '|' if you have several output nodes
output_node_names = "first-output/visual|second-output/visual|third-output/visual|fourth-output/visual|output/scores"
# Save the .pb model file
output_graph_def = tf.graph_util.convert_variables_to_constants(
sess, sess.graph_def, output_node_names.split("|"))
tf.train.write_graph(output_graph_def,
"graph",
"graph-harnn-{0}.pb".format(MODEL),
as_text=False)
# Generate batches for one epoch
batches = dh.batch_iter(list(
zip(x_test, y_test, y_test_tuple, x_test_content,
y_test_labels)),
args.batch_size,
1,
shuffle=False)
for batch_test in batches:
x_batch_test, y_batch_test, y_batch_test_tuple, \
x_batch_test_content, y_batch_test_labels = zip(*batch_test)
y_batch_test_first = [i[0] for i in y_batch_test_tuple]
y_batch_test_second = [j[1] for j in y_batch_test_tuple]
y_batch_test_third = [k[2] for k in y_batch_test_tuple]
y_batch_test_fourth = [t[3] for t in y_batch_test_tuple]
feed_dict = {
input_x: x_batch_test,
input_y_first: y_batch_test_first,
input_y_second: y_batch_test_second,
input_y_third: y_batch_test_third,
input_y_fourth: y_batch_test_fourth,
input_y: y_batch_test,
dropout_keep_prob: 1.0,
alpha: args.alpha,
is_training: False
}
batch_first_visual, batch_second_visual, batch_third_visual, batch_fourth_visual = \
sess.run([first_visual, second_visual, third_visual, fourth_visual], feed_dict)
seq_len = len(x_batch_test_content[0])
pad_len = len(batch_first_visual[0])
length = (pad_len if seq_len >= pad_len else seq_len)
# print(seq_len, pad_len, length)
final_first_visual = normalization(
batch_first_visual[0].tolist(), length)
final_second_visual = normalization(
batch_second_visual[0].tolist(), length)
final_third_visual = normalization(
batch_third_visual[0].tolist(), length)
final_fourth_visual = normalization(
batch_fourth_visual[0].tolist(), length)
visual_list = [
final_first_visual, final_second_visual,
final_third_visual, final_fourth_visual
]
print(visual_list)
f = open('attention.html', 'w')
f.write(
'<html style="margin:0;padding:0;"><body style="margin:0;padding:0;">\n'
)
f.write('<div style="margin:25px;">\n')
for k in range(len(visual_list)):
f.write('<p style="margin:10px;">\n')
for i in range(seq_len):
alpha = "{:.2f}".format(visual_list[k][i])
word = x_batch_test_content[0][i]
f.write(
'\t<span style="margin-left:3px;background-color:rgba(255,0,0,{0})">{1}</span>\n'
.format(alpha, word))
f.write('</p>\n')
f.write('</div>\n')
f.write('</body></html>')
f.close()
logger.info("Done.")
def test_rmidp():
"""Test RMIDP model."""
# Print parameters used for the model
dh.tab_printer(args, logger)
# Load data
logger.info("Loading data...")
logger.info("Data processing...")
test_data = dh.load_data_and_labels(args.test_file, args.word2vec_file, data_aug_flag=False)
logger.info("Data padding...")
x_test_content, x_test_question, x_test_option, y_test = dh.pad_data(test_data, args.pad_seq_len)
# Load rmidp model
OPTION = dh.option(pattern=1)
if OPTION == 'B':
logger.info("Loading best model...")
checkpoint_file = cm.get_best_checkpoint(BEST_CPT_DIR, select_maximum_value=True)
else:
logger.info("Loading latest model...")
checkpoint_file = tf.train.latest_checkpoint(CPT_DIR)
logger.info(checkpoint_file)
graph = tf.Graph()
with graph.as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=args.allow_soft_placement,
log_device_placement=args.log_device_placement)
session_conf.gpu_options.allow_growth = args.gpu_options_allow_growth
sess = tf.Session(config=session_conf)
with sess.as_default():
# Load the saved meta graph and restore variables
saver = tf.train.import_meta_graph("{0}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
# Get the placeholders from the graph by name
input_x_content = graph.get_operation_by_name("input_x_content").outputs[0]
input_x_question = graph.get_operation_by_name("input_x_question").outputs[0]
input_x_option = graph.get_operation_by_name("input_x_option").outputs[0]
input_y = graph.get_operation_by_name("input_y").outputs[0]
dropout_keep_prob = graph.get_operation_by_name("dropout_keep_prob").outputs[0]
is_training = graph.get_operation_by_name("is_training").outputs[0]
# Tensors we want to evaluate
scores = graph.get_operation_by_name("output/scores").outputs[0]
loss = graph.get_operation_by_name("loss/loss").outputs[0]
# Split the output nodes name by '|' if you have several output nodes
output_node_names = "output/scores"
# Save the .pb model file
output_graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def,
output_node_names.split("|"))
tf.train.write_graph(output_graph_def, "graph", "graph-rmidp-{0}.pb".format(MODEL), as_text=False)
# Generate batches for one epoch
batches = dh.batch_iter(list(zip(x_test_content, x_test_question, x_test_option, y_test)),
args.batch_size, 1, shuffle=False)
test_counter, test_loss = 0, 0.0
# Collect the predictions here
true_labels = []
predicted_scores = []
for batch_test in batches:
x_batch_content, x_batch_question, x_batch_option, y_batch = zip(*batch_test)
feed_dict = {
input_x_content: x_batch_content,
input_x_question: x_batch_question,
input_x_option: x_batch_option,
input_y: y_batch,
dropout_keep_prob: 1.0,
is_training: False
}
batch_scores, cur_loss = sess.run([scores, loss], feed_dict)
# Prepare for calculating metrics
for i in y_batch:
true_labels.append(i)
for j in batch_scores:
predicted_scores.append(j)
test_loss = test_loss + cur_loss
test_counter = test_counter + 1
# Calculate PCC & DOA
pcc, doa = dh.evaluation(true_labels, predicted_scores)
# Calculate RMSE
rmse = mean_squared_error(true_labels, predicted_scores) ** 0.5
r2 = r2_score(true_labels, predicted_scores)
test_loss = float(test_loss / test_counter)
logger.info("All Test Dataset: Loss {0:g} | PCC {1:g} | DOA {2:g} | RMSE {3:g} | R2 {4:g}"
.format(test_loss, pcc, doa, rmse, r2))
# Save the prediction result
if not os.path.exists(SAVE_DIR):
os.makedirs(SAVE_DIR)
dh.create_prediction_file(output_file=SAVE_DIR + "/predictions.json", all_id=test_data.id,
all_labels=true_labels, all_predict_scores=predicted_scores)
logger.info("All Done.")
def test():
# Load data
logger.info("Loading data...")
logger.info("Training data processing...")
test_students, test_max_num_problems, test_max_skill_num = dh.read_test_data_from_csv_file(
FLAGS.test_data_file)
max_num_steps = test_max_num_problems
max_num_skills = test_max_skill_num
# Load rnn model
BEST_OR_LATEST = input("Load Best or Latest Model?(B/L): ")
while not (BEST_OR_LATEST.isalpha()
and BEST_OR_LATEST.upper() in ['B', 'L']):
BEST_OR_LATEST = input(
"he format of your input is illegal, please re-input: ")
if BEST_OR_LATEST == 'B':
logger.info("Loading best model...")
checkpoint_file = cm.get_best_checkpoint(FLAGS.best_checkpoint_dir,
select_maximum_value=True)
if BEST_OR_LATEST == 'L':
logger.info("latest")
checkpoint_file = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
logger.info(checkpoint_file)
graph = tf.Graph()
with graph.as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=FLAGS.allow_soft_placement,
log_device_placement=FLAGS.log_device_placement)
session_conf.gpu_options.allow_growth = FLAGS.gpu_options_allow_growth
sess = tf.Session(config=session_conf)
with sess.as_default():
# Load the saved meta graph and restore variables
saver = tf.train.import_meta_graph(
"{0}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
# Get the placeholders from the graph by name
input_data = graph.get_operation_by_name("input_data").outputs[0]
input_skill = graph.get_operation_by_name("input_skill").outputs[0]
l = graph.get_operation_by_name("l").outputs[0]
next_id = graph.get_operation_by_name("next_id").outputs[0]
target_id = graph.get_operation_by_name("target_id").outputs[0]
target_correctness = graph.get_operation_by_name(
"target_correctness").outputs[0]
target_id2 = graph.get_operation_by_name("target_id2").outputs[0]
target_correctness2 = graph.get_operation_by_name(
"target_correctness2").outputs[0]
dropout_keep_prob = graph.get_operation_by_name(
"dropout_keep_prob").outputs[0]
is_training = graph.get_operation_by_name("is_training").outputs[0]
skill = graph.get_operation_by_name("skill_w").outputs[0]
states = graph.get_operation_by_name("states").outputs[0]
pred = graph.get_operation_by_name("pred").outputs[0]
data_size = len(test_students)
index = 0
actual_labels = []
pred_labels = []
while (index + FLAGS.batch_size < data_size):
x = np.zeros((FLAGS.batch_size, max_num_steps))
xx = np.zeros((FLAGS.batch_size, max_num_steps))
next_id_b = np.zeros((FLAGS.batch_size, max_num_steps))
l_b = np.ones(
(FLAGS.batch_size, max_num_steps, max_num_skills))
target_id_b = []
target_correctness_b = []
target_id2_b = []
target_correctness2_b = []
for i in range(FLAGS.batch_size):
student = test_students[index + i]
problem_ids = student[1]
correctness = student[2]
leng = len(problem_ids)
correct_num = np.zeros(max_num_skills)
answer_count = np.ones(max_num_skills)
for j in range(len(problem_ids) - 1):
problem_id = int(problem_ids[j])
if (int(correctness[j]) == 0):
x[i, j] = problem_id + max_num_skills
else:
x[i, j] = problem_id
correct_num[problem_id] += 1
l_b[i, j] = correct_num / answer_count
answer_count[problem_id] += 1
xx[i, j] = problem_id
next_id_b[i, j] = int(problem_ids[j + 1])
target_id_b.append(i * max_num_steps + j)
target_correctness_b.append(int(correctness[j + 1]))
actual_labels.append(int(correctness[j + 1]))
target_id2_b.append(i * max_num_steps + j)
target_correctness2_b.append(int(correctness[j + 1]))
index += FLAGS.batch_size
feed_dict = {
input_data: x,
input_skill: xx,
l: l_b,
next_id: next_id_b,
target_id: target_id_b,
target_correctness: target_correctness_b,
target_id2: target_id2_b,
target_correctness2: target_correctness2_b,
dropout_keep_prob: 1.0,
is_training: False
}
'''
skill_b = sess.run([skill], feed_dict)
print(np.shape(skill_b))
item = skill_b[0]
with open('skill_2009.txt', 'a')as fi:
for temp in item:
for iii in temp:
fi.write(str(iii) + ',')
fi.write('\n')
break
'''
pred_b, state = sess.run([pred, states], feed_dict)
print(np.shape(state))
print(np.shape(pred_b))
state = np.squeeze(state, axis=0)
state = state[:leng]
if leng > 50 and leng < 100:
writer = csv.writer(open('state.csv', 'a', newline=''))
writer.writerow([len(problem_ids)])
writer.writerow(student[1])
writer.writerow(student[2])
writer.writerow(state)
writer.writerow('\n')
for p in pred_b:
pred_labels.append(p)
rmse = sqrt(mean_squared_error(actual_labels, pred_labels))
fpr, tpr, thresholds = metrics.roc_curve(actual_labels,
pred_labels,
pos_label=1)
auc = metrics.auc(fpr, tpr)
#calculate r^2
r2 = r2_score(actual_labels, pred_labels)
print("epochs {0}: rmse {1:g} auc {2:g} r2 {3:g} ".format(
1, rmse, auc, r2))
logger.info("epochs {0}: rmse {1:g} auc {2:g} r2 {3:g} ".format(
1, rmse, auc, r2))
logger.info("Done.")
def test_cnn():
"""Test CNN model."""
# Print parameters used for the model
dh.tab_printer(args, logger)
# Load word2vec model
word2idx, embedding_matrix = dh.load_word2vec_matrix(args.word2vec_file)
# Load data
logger.info("Loading data...")
logger.info("Data processing...")
test_data = dh.load_data_and_labels(args, args.test_file, word2idx)
# Load cnn model
OPTION = dh._option(pattern=1)
if OPTION == 'B':
logger.info("Loading best model...")
checkpoint_file = cm.get_best_checkpoint(BEST_CPT_DIR, select_maximum_value=True)
else:
logger.info("Loading latest model...")
checkpoint_file = tf.train.latest_checkpoint(CPT_DIR)
logger.info(checkpoint_file)
graph = tf.Graph()
with graph.as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=args.allow_soft_placement,
log_device_placement=args.log_device_placement)
session_conf.gpu_options.allow_growth = args.gpu_options_allow_growth
sess = tf.Session(config=session_conf)
with sess.as_default():
# Load the saved meta graph and restore variables
saver = tf.train.import_meta_graph("{0}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
# Get the placeholders from the graph by name
input_x_front = graph.get_operation_by_name("input_x_front").outputs[0]
input_x_behind = graph.get_operation_by_name("input_x_behind").outputs[0]
input_y = graph.get_operation_by_name("input_y").outputs[0]
dropout_keep_prob = graph.get_operation_by_name("dropout_keep_prob").outputs[0]
is_training = graph.get_operation_by_name("is_training").outputs[0]
# Tensors we want to evaluate
scores = graph.get_operation_by_name("output/topKPreds").outputs[0]
predictions = graph.get_operation_by_name("output/topKPreds").outputs[1]
loss = graph.get_operation_by_name("loss/loss").outputs[0]
# Split the output nodes name by '|' if you have several output nodes
output_node_names = "output/topKPreds"
# Save the .pb model file
output_graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def,
output_node_names.split("|"))
tf.train.write_graph(output_graph_def, "graph", "graph-cnn-{0}.pb".format(MODEL), as_text=False)
# Generate batches for one epoch
batches_test = dh.batch_iter(list(create_input_data(test_data)), args.batch_size, 1, shuffle=False)
# Collect the predictions here
test_counter, test_loss = 0, 0.0
true_labels = []
predicted_labels = []
predicted_scores = []
for batch_test in batches_test:
x_f, x_b, y_onehot = zip(*batch_test)
feed_dict = {
input_x_front: x_f,
input_x_behind: x_b,
input_y: y_onehot,
dropout_keep_prob: 1.0,
is_training: False
}
batch_predicted_scores, batch_predicted_labels, batch_loss \
= sess.run([scores, predictions, loss], feed_dict)
for i in y_onehot:
true_labels.append(np.argmax(i))
for j in batch_predicted_scores:
predicted_scores.append(j[0])
for k in batch_predicted_labels:
predicted_labels.append(k[0])
test_loss = test_loss + batch_loss
test_counter = test_counter + 1
test_loss = float(test_loss / test_counter)
# Calculate Precision & Recall & F1
test_acc = accuracy_score(y_true=np.array(true_labels), y_pred=np.array(predicted_labels))
test_pre = precision_score(y_true=np.array(true_labels),
y_pred=np.array(predicted_labels), average='micro')
test_rec = recall_score(y_true=np.array(true_labels),
y_pred=np.array(predicted_labels), average='micro')
test_F1 = f1_score(y_true=np.array(true_labels),
y_pred=np.array(predicted_labels), average='micro')
# Calculate the average AUC
test_auc = roc_auc_score(y_true=np.array(true_labels),
y_score=np.array(predicted_scores), average='micro')
logger.info("All Test Dataset: Loss {0:g} | Acc {1:g} | Precision {2:g} | "
"Recall {3:g} | F1 {4:g} | AUC {5:g}"
.format(test_loss, test_acc, test_pre, test_rec, test_F1, test_auc))
# Save the prediction result
if not os.path.exists(SAVE_DIR):
os.makedirs(SAVE_DIR)
dh.create_prediction_file(output_file=SAVE_DIR + "/predictions.json", front_data_id=test_data['f_id'],
behind_data_id=test_data['b_id'], true_labels=true_labels,
predict_labels=predicted_labels, predict_scores=predicted_scores)
logger.info("All Done.")
def test_harnn():
"""Test HARNN model."""
# Print parameters used for the model
dh.tab_printer(args, logger)
# Load word2vec model
word2idx, embedding_matrix = dh.load_word2vec_matrix(args.word2vec_file)
# Load data
logger.info("Loading data...")
logger.info("Data processing...")
test_data = dh.load_data_and_labels(args, args.test_file, word2idx)
# Load harnn model
OPTION = dh._option(pattern=1)
if OPTION == 'B':
logger.info("Loading best model...")
checkpoint_file = cm.get_best_checkpoint(BEST_CPT_DIR,
select_maximum_value=True)
else:
logger.info("Loading latest model...")
checkpoint_file = tf.train.latest_checkpoint(CPT_DIR)
logger.info(checkpoint_file)
graph = tf.Graph()
with graph.as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=args.allow_soft_placement,
log_device_placement=args.log_device_placement)
session_conf.gpu_options.allow_growth = args.gpu_options_allow_growth
sess = tf.Session(config=session_conf)
with sess.as_default():
# Load the saved meta graph and restore variables
saver = tf.train.import_meta_graph(
"{0}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
# Get the placeholders from the graph by name
input_x = graph.get_operation_by_name("input_x").outputs[0]
input_y_first = graph.get_operation_by_name(
"input_y_first").outputs[0]
input_y_second = graph.get_operation_by_name(
"input_y_second").outputs[0]
input_y_third = graph.get_operation_by_name(
"input_y_third").outputs[0]
input_y_fourth = graph.get_operation_by_name(
"input_y_fourth").outputs[0]
input_y = graph.get_operation_by_name("input_y").outputs[0]
dropout_keep_prob = graph.get_operation_by_name(
"dropout_keep_prob").outputs[0]
alpha = graph.get_operation_by_name("alpha").outputs[0]
is_training = graph.get_operation_by_name("is_training").outputs[0]
# Tensors we want to evaluate
first_scores = graph.get_operation_by_name(
"first-output/scores").outputs[0]
second_scores = graph.get_operation_by_name(
"second-output/scores").outputs[0]
third_scores = graph.get_operation_by_name(
"third-output/scores").outputs[0]
fourth_scores = graph.get_operation_by_name(
"fourth-output/scores").outputs[0]
scores = graph.get_operation_by_name("output/scores").outputs[0]
# Split the output nodes name by '|' if you have several output nodes
output_node_names = "first-output/scores|second-output/scores|third-output/scores|fourth-output/scores|output/scores"
# Save the .pb model file
output_graph_def = tf.graph_util.convert_variables_to_constants(
sess, sess.graph_def, output_node_names.split("|"))
tf.train.write_graph(output_graph_def,
"graph",
"graph-harnn-{0}.pb".format(MODEL),
as_text=False)
# Generate batches for one epoch
batches = dh.batch_iter(list(create_input_data(test_data)),
args.batch_size,
1,
shuffle=False)
# Collect the predictions here
true_labels = []
predicted_labels = []
predicted_scores = []
# Collect for calculating metrics
true_onehot_labels = [[], [], [], [], []]
predicted_onehot_scores = [[], [], [], [], []]
predicted_onehot_labels = [[], [], [], [], []]
for batch_test in batches:
x, sec, subsec, group, subgroup, y_onehot, y = zip(*batch_test)
y_batch_test_list = [y_onehot, sec, subsec, group, subgroup]
feed_dict = {
input_x: x,
input_y_first: sec,
input_y_second: subsec,
input_y_third: group,
input_y_fourth: subgroup,
input_y: y_onehot,
dropout_keep_prob: 1.0,
alpha: args.alpha,
is_training: False
}
batch_global_scores, batch_first_scores, batch_second_scores, batch_third_scores, batch_fourth_scores = \
sess.run([scores, first_scores, second_scores, third_scores, fourth_scores], feed_dict)
batch_scores = [
batch_global_scores, batch_first_scores,
batch_second_scores, batch_third_scores,
batch_fourth_scores
]
# Get the predicted labels by threshold
batch_predicted_labels_ts, batch_predicted_scores_ts = \
dh.get_label_threshold(scores=batch_scores[0], threshold=args.threshold)
# Add results to collection
for labels in y:
true_labels.append(labels)
for labels in batch_predicted_labels_ts:
predicted_labels.append(labels)
for values in batch_predicted_scores_ts:
predicted_scores.append(values)
for index in range(len(predicted_onehot_scores)):
for onehot_labels in y_batch_test_list[index]:
true_onehot_labels[index].append(onehot_labels)
for onehot_scores in batch_scores[index]:
predicted_onehot_scores[index].append(onehot_scores)
# Get one-hot prediction by threshold
predicted_onehot_labels_ts = \
dh.get_onehot_label_threshold(scores=batch_scores[index], threshold=args.threshold)
for onehot_labels in predicted_onehot_labels_ts:
predicted_onehot_labels[index].append(onehot_labels)
# Calculate Precision & Recall & F1
for index in range(len(predicted_onehot_scores)):
test_pre = precision_score(
y_true=np.array(true_onehot_labels[index]),
y_pred=np.array(predicted_onehot_labels[index]),
average='micro')
test_rec = recall_score(
y_true=np.array(true_onehot_labels[index]),
y_pred=np.array(predicted_onehot_labels[index]),
average='micro')
test_F1 = f1_score(y_true=np.array(true_onehot_labels[index]),
y_pred=np.array(
predicted_onehot_labels[index]),
average='micro')
test_auc = roc_auc_score(
y_true=np.array(true_onehot_labels[index]),
y_score=np.array(predicted_onehot_scores[index]),
average='micro')
test_prc = average_precision_score(
y_true=np.array(true_onehot_labels[index]),
y_score=np.array(predicted_onehot_scores[index]),
average="micro")
if index == 0:
logger.info(
"[Global] Predict by threshold: Precision {0:g}, Recall {1:g}, "
"F1 {2:g}, AUC {3:g}, AUPRC {4:g}".format(
test_pre, test_rec, test_F1, test_auc, test_prc))
else:
logger.info(
"[Local] Predict by threshold in Level-{0}: Precision {1:g}, Recall {2:g}, "
"F1 {3:g}, AUPRC {4:g}".format(index, test_pre,
test_rec, test_F1,
test_prc))
# Save the prediction result
if not os.path.exists(SAVE_DIR):
os.makedirs(SAVE_DIR)
dh.create_prediction_file(output_file=SAVE_DIR +
"/predictions.json",
data_id=test_data['uniq_id'],
true_labels=true_labels,
predict_labels=predicted_labels,
predict_scores=predicted_scores)
logger.info("All Done.")
def test_cnn():
"""Test CNN model."""
# Load data
logger.info("✔︎ Loading data...")
logger.info("Recommended padding Sequence length is: {0}".format(FLAGS.pad_seq_len))
logger.info("✔︎ Test data processing...")
test_data = dh.load_data_and_labels(FLAGS.test_data_file, FLAGS.num_classes,
FLAGS.embedding_dim, data_aug_flag=False)
logger.info("✔︎ Test data padding...")
x_test, y_test = dh.pad_data(test_data, FLAGS.pad_seq_len)
y_test_labels = test_data.labels
# Load cnn model
BEST_OR_LATEST = input("☛ Load Best or Latest Model?(B/L): ")
while not (BEST_OR_LATEST.isalpha() and BEST_OR_LATEST.upper() in ['B', 'L']):
BEST_OR_LATEST = input("✘ The format of your input is illegal, please re-input: ")
if BEST_OR_LATEST.upper() == 'B':
logger.info("✔︎ Loading best model...")
checkpoint_file = cm.get_best_checkpoint(FLAGS.best_checkpoint_dir, select_maximum_value=True)
else:
logger.info("✔︎ Loading latest model...")
checkpoint_file = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
logger.info(checkpoint_file)
graph = tf.Graph()
with graph.as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=FLAGS.allow_soft_placement,
log_device_placement=FLAGS.log_device_placement)
session_conf.gpu_options.allow_growth = FLAGS.gpu_options_allow_growth
sess = tf.Session(config=session_conf)
with sess.as_default():
# Load the saved meta graph and restore variables
saver = tf.train.import_meta_graph("{0}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
# Get the placeholders from the graph by name
input_x = graph.get_operation_by_name("input_x").outputs[0]
input_y = graph.get_operation_by_name("input_y").outputs[0]
dropout_keep_prob = graph.get_operation_by_name("dropout_keep_prob").outputs[0]
is_training = graph.get_operation_by_name("is_training").outputs[0]
# Tensors we want to evaluate
scores = graph.get_operation_by_name("output/scores").outputs[0]
loss = graph.get_operation_by_name("loss/loss").outputs[0]
# Split the output nodes name by '|' if you have several output nodes
output_node_names = "output/scores"
# Save the .pb model file
output_graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def,
output_node_names.split("|"))
tf.train.write_graph(output_graph_def, "graph", "graph-cnn-{0}.pb".format(MODEL), as_text=False)
# Generate batches for one epoch
batches = dh.batch_iter(list(zip(x_test, y_test, y_test_labels)), FLAGS.batch_size, 1, shuffle=False)
test_counter, test_loss = 0, 0.0
test_pre_tk = [0.0] * FLAGS.top_num
test_rec_tk = [0.0] * FLAGS.top_num
test_F_tk = [0.0] * FLAGS.top_num
# Collect the predictions here
true_labels = []
predicted_labels = []
predicted_scores = []
# Collect for calculating metrics
true_onehot_labels = []
predicted_onehot_scores = []
predicted_onehot_labels_ts = []
predicted_onehot_labels_tk = [[] for _ in range(FLAGS.top_num)]
for batch_test in batches:
x_batch_test, y_batch_test, y_batch_test_labels = zip(*batch_test)
feed_dict = {
input_x: x_batch_test,
input_y: y_batch_test,
dropout_keep_prob: 1.0,
is_training: False
}
batch_scores, cur_loss = sess.run([scores, loss], feed_dict)
# Prepare for calculating metrics
for i in y_batch_test:
true_onehot_labels.append(i)
for j in batch_scores:
predicted_onehot_scores.append(j)
# Get the predicted labels by threshold
batch_predicted_labels_ts, batch_predicted_scores_ts = \
dh.get_label_threshold(scores=batch_scores, threshold=FLAGS.threshold)
# Add results to collection
for i in y_batch_test_labels:
true_labels.append(i)
for j in batch_predicted_labels_ts:
predicted_labels.append(j)
for k in batch_predicted_scores_ts:
predicted_scores.append(k)
# Get onehot predictions by threshold
batch_predicted_onehot_labels_ts = \
dh.get_onehot_label_threshold(scores=batch_scores, threshold=FLAGS.threshold)
for i in batch_predicted_onehot_labels_ts:
predicted_onehot_labels_ts.append(i)
# Get onehot predictions by topK
for top_num in range(FLAGS.top_num):
batch_predicted_onehot_labels_tk = dh.get_onehot_label_topk(scores=batch_scores, top_num=top_num+1)
for i in batch_predicted_onehot_labels_tk:
predicted_onehot_labels_tk[top_num].append(i)
test_loss = test_loss + cur_loss
test_counter = test_counter + 1
# Calculate Precision & Recall & F1 (threshold & topK)
test_pre_ts = precision_score(y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_ts), average='micro')
test_rec_ts = recall_score(y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_ts), average='micro')
test_F_ts = f1_score(y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_ts), average='micro')
for top_num in range(FLAGS.top_num):
test_pre_tk[top_num] = precision_score(y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_tk[top_num]),
average='micro')
test_rec_tk[top_num] = recall_score(y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_tk[top_num]),
average='micro')
test_F_tk[top_num] = f1_score(y_true=np.array(true_onehot_labels),
y_pred=np.array(predicted_onehot_labels_tk[top_num]),
average='micro')
# Calculate the average AUC
test_auc = roc_auc_score(y_true=np.array(true_onehot_labels),
y_score=np.array(predicted_onehot_scores), average='micro')
# Calculate the average PR
test_prc = average_precision_score(y_true=np.array(true_onehot_labels),
y_score=np.array(predicted_onehot_scores), average="micro")
test_loss = float(test_loss / test_counter)
logger.info("☛ All Test Dataset: Loss {0:g} | AUC {1:g} | AUPRC {2:g}"
.format(test_loss, test_auc, test_prc))
# Predict by threshold
logger.info("☛ Predict by threshold: Precision {0:g}, Recall {1:g}, F1 {2:g}"
.format(test_pre_ts, test_rec_ts, test_F_ts))
# Predict by topK
logger.info("☛ Predict by topK:")
for top_num in range(FLAGS.top_num):
logger.info("Top{0}: Precision {1:g}, Recall {2:g}, F {3:g}"
.format(top_num + 1, test_pre_tk[top_num], test_rec_tk[top_num], test_F_tk[top_num]))
# Save the prediction result
if not os.path.exists(SAVE_DIR):
os.makedirs(SAVE_DIR)
dh.create_prediction_file(output_file=SAVE_DIR + "/predictions.json", data_id=test_data.testid,
all_labels=true_labels, all_predict_labels=predicted_labels,
all_predict_scores=predicted_scores)
logger.info("✔︎ Done.")
def test():
test_students, test_max_num_problems, test_max_skill_num = dh.read_test_data_from_csv_file(
FLAGS.test_data_file)
max_num_steps = test_max_num_problems
max_num_skills = test_max_skill_num
fileName = "clustered_skill_name.txt"
same_b, differ_b = eb.embedding(fileName)
BEST_OR_LATEST = input("Load Best or Latest Model?(B/L): ")
while not (BEST_OR_LATEST.isalpha()
and BEST_OR_LATEST.upper() in ['B', 'L']):
BEST_OR_LATEST = input(
"he format of your input is illegal, please re-input: ")
if BEST_OR_LATEST == 'B':
logger.info("Loading best model...")
checkpoint_file = cm.get_best_checkpoint(FLAGS.best_checkpoint_dir,
select_maximum_value=True)
if BEST_OR_LATEST == 'L':
logger.info("latest")
checkpoint_file = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
logger.info(checkpoint_file)
graph = tf.Graph()
with graph.as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=FLAGS.allow_soft_placement,
log_device_placement=FLAGS.log_device_placement)
session_conf.gpu_options.allow_growth = FLAGS.gpu_options_allow_growth
sess = tf.Session(config=session_conf)
with sess.as_default():
saver = tf.train.import_meta_graph(
"{0}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
input_data = graph.get_operation_by_name("input_data").outputs[0]
next_id = graph.get_operation_by_name("next_id").outputs[0]
target_id = graph.get_operation_by_name("target_id").outputs[0]
target_correctness = graph.get_operation_by_name(
"target_correctness").outputs[0]
dropout_keep_prob = graph.get_operation_by_name(
"dropout_keep_prob").outputs[0]
is_training = graph.get_operation_by_name("is_training").outputs[0]
same = graph.get_operation_by_name("same").outputs[0]
differ = graph.get_operation_by_name("differ").outputs[0]
skill = graph.get_operation_by_name("skill_w").outputs[0]
states = graph.get_operation_by_name("states").outputs[0]
pred = graph.get_operation_by_name("pred").outputs[0]
data_size = len(test_students)
index = 0
actual_labels = []
pred_labels = []
while (index < data_size):
x = np.zeros((FLAGS.batch_size, max_num_steps))
xx = np.zeros((FLAGS.batch_size, max_num_steps))
next_id_b = np.zeros((FLAGS.batch_size, max_num_steps))
l_b = np.ones(
(FLAGS.batch_size, max_num_steps, max_num_skills))
target_id_b = []
target_correctness_b = []
for i in range(FLAGS.batch_size):
student = test_students[index + i]
problem_ids = student[1]
correctness = student[2]
correct_num = np.zeros(max_num_skills)
answer_count = np.ones(max_num_skills)
for j in range(len(problem_ids) - 1):
problem_id = int(problem_ids[j])
if (int(correctness[j]) == 0):
x[i, j] = problem_id + max_num_skills
else:
x[i, j] = problem_id
correct_num[problem_id] += 1
l_b[i, j] = correct_num / answer_count
answer_count[problem_id] += 1
xx[i, j] = problem_id
next_id_b[i, j] = int(problem_ids[j + 1])
target_id_b.append(i * max_num_steps + j)
target_correctness_b.append(int(correctness[j + 1]))
actual_labels.append(int(correctness[j + 1]))
index += FLAGS.batch_size
feed_dict = {
input_data: x,
next_id: next_id_b,
target_id: target_id_b,
target_correctness: target_correctness_b,
dropout_keep_prob: 1.0,
is_training: False,
same: same_b,
differ: differ_b
}
pred_b, state, skill_b = sess.run([pred, states, skill],
feed_dict)
for p in pred_b:
pred_labels.append(p)
rmse = sqrt(mean_squared_error(actual_labels, pred_labels))
fpr, tpr, thresholds = metrics.roc_curve(actual_labels,
pred_labels,
pos_label=1)
auc = metrics.auc(fpr, tpr)
r2 = r2_score(actual_labels, pred_labels)
pred_score = np.greater_equal(pred_labels, 0.5)
pred_score = pred_score.astype(int)
pred_score = np.equal(actual_labels, pred_score)
acc = np.mean(pred_score.astype(int))
logger.info(
"epochs {0}: rmse {1:g} auc {2:g} r2 {3:g} acc {4:g}".
format(1, rmse, auc, r2, acc))