def __init__(self, num_frames, model_path):
super(InferenceModel, self).__init__()
self.num_frames = num_frames
self.model_path = model_path
with tf.Graph().as_default():
self.inputs_pl = tf.placeholder(
tf.float32, shape=(
1, 240, 320, cfg.num_frames))
self.is_training_pl = tf.placeholder(tf.bool, shape=())
pred = build_graph(self.inputs_pl, self.is_training_pl,
weight_decay=0.0, bn_decay=None)
self.pred=tf.nn.softmax(pred)
# config_ss = tf.ConfigProto()
self.sess = tf.Session()
saver = tf.train.Saver()
saver.restore(self.sess, model_path)
print("\nLoaded model... ", model_path)
def train():
log_string('***** Config *****')
log_string('***** Building Point {}...'.format(MODEL_NAME))
log_string('** num_frames: {}'.format(cfg.num_frames))
log_string('** num_classes: {}'.format(cfg.num_classes))
log_string('** batch_size: {}'.format(cfg.batch_size))
log_string('** epoch: {}'.format(cfg.epoch))
log_string('** init_learning_rate: {}'.format(cfg.init_learning_rate))
log_string('** decay_step: {}'.format(cfg.decay_step))
log_string('** decay_rate: {}'.format(cfg.decay_rate))
log_string('** weight_decay: {}'.format(cfg.weight_decay))
with tf.Graph().as_default():
inputs, labels = placeholder_inputs(cfg.batch_size, cfg.num_frames)
is_training_pl = tf.placeholder(tf.bool, shape=())
keep_prob_pl = tf.placeholder(tf.float32)
global_step = tf.Variable(0, dtype=tf.int64)
bn_decay = get_bn_decay(global_step)
tf.summary.scalar('bn_decay', bn_decay)
pred = build_graph(inputs,
is_training_pl,
weight_decay=cfg.weight_decay,
keep_prob=keep_prob_pl,
bn_decay=bn_decay)
loss = get_loss(pred, labels)
# raise
tf.summary.scalar('total_loss', loss)
correct = tf.equal(tf.argmax(pred, 1), tf.to_int64(labels))
correct = tf.reduce_sum(tf.cast(correct, tf.float32))
accuracy = correct / float(cfg.batch_size)
tf.summary.scalar('accuracy', accuracy)
# Get training operator
learning_rate = get_learning_rate(global_step)
tf.summary.scalar('learning_rate', learning_rate)
optimizer = tf.train.AdamOptimizer(learning_rate)
# optimizer = tf.train.MomentumOptimizer(learning_rate, momentum=0.9)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(loss, global_step=global_step)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
# config.log_device_placement = True
# config.gpu_options.allocator_type = "BFC"
sess = tf.Session(config=config)
# # restore model #################
load_model_path = LOGDIR + '/model_epoch_{}'.format(
cfg.load_model_epoch)
try:
saver = tf.train.Saver()
saver.restore(sess, load_model_path)
print("\nPrevious model restored... ", load_model_path)
except Exception as e:
print("\nCannot find the requested model... {}".format(e))
sess.run(tf.global_variables_initializer())
# %% create a saver object
saver = tf.train.Saver()
print("\nCreating new model...", load_model_path)
if cfg.debug:
sess = tf_debug.LocalCLIDebugWrapperSession(sess)
# init = tf.global_variables_initializer()
# sess.run(init, {is_training_pl: True})
# saver = tf.train.Saver()
# Plot Variable Histogram
t_vars = tf.trainable_variables()
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(LOGDIR + '/train')
train_writer.add_graph(tf.get_default_graph())
test_writer = tf.summary.FileWriter(LOGDIR + '/test')
test_writer.add_graph(tf.get_default_graph())
# running_vars = tf.get_collection('metric_vars')
# running_vars = tf.get_collection(tf.GraphKeys.METRIC_VARIABLES)
# running_vars = [ var for var in running_vars if isinstance(var, tf.Variable)]
# print(running_vars)
# running_vars_initializer = tf.variables_initializer(var_list=running_vars)
# Count number of trainable parameters
num_params = np.sum([np.prod(v.get_shape().as_list()) for v in t_vars])
log_string(
'************ The Number of Trainable Parameters: {} ************'.
format(num_params))
num_g_params = np.sum(
[np.prod(v.get_shape().as_list()) for v in tf.global_variables()])
log_string(
'************ The Number of Global Parameters: {} ************'.
format(num_g_params))
ops = {
'inputs_pl': inputs,
'labels_pl': labels,
'is_training_pl': is_training_pl,
'keep_prob_pl': keep_prob_pl,
'pred': pred,
'loss': loss,
'train_op': train_op,
'merged': merged,
'step': global_step
}
# %%
# training_dataset = np.load(
# '/media/tjosh/vault/MSRAction3D/all_npy_5_t1_training.npy')
# validation_dataset = np.load(
# '/media/tjosh/vault/MSRAction3D/all_npy_5_t1_validation.npy')
# for utkinect
training_dataset = np.load(
'D:/datasets/UTKinectAction3D_npy_5/training.npy')
validation_dataset = np.load(
'D:/datasets/UTKinectAction3D_npy_5/validation.npy')
# '/media/tjosh/vault/MSRAction3D/all_npy_5_t3_training.npy')
# '/media/tjosh/vault/MSRAction3D/all_npy_5_t3_validation.npy')
train_data_gen = DataGenerator(training_dataset,
batch_size=cfg.batch_size,
augment=False)
validation_data_gen = DataGenerator(validation_dataset,
batch_size=cfg.batch_size,
augment=False)
# %%
# dataset = glob.glob(
# '/media/tjosh/vault/UTKinectAction3D_train_npy_6/**/*.npy')
# dataset_size = len(dataset)
for epoch in range(1, cfg.epoch + 1):
# dataset = shuffle(dataset)
# train_cut = int(dataset_size*0.50)
# training_dataset = dataset[:train_cut]
# validation_dataset = dataset[train_cut:]
# train_data_gen = DataGenerator(training_dataset, batch_size=cfg.batch_size)
# validation_data_gen = DataGenerator(validation_dataset, batch_size=cfg.batch_size, augment=False)
log_string('\n******** Training:---Epoch_{}/{} *********'.format(
epoch, cfg.epoch))
log_string('Training ...')
train_one_epoch(sess, train_data_gen, ops, train_writer)
log_string('Validating ...')
val_one_epoch(sess, validation_data_gen, ops, test_writer)
if epoch % cfg.save_model_freq == 0:
saver.save(sess, LOGDIR + '/model_epoch_{}'.format(epoch))
log_string('Model saved at epoch {}'.format(epoch))
def test_model():
with tf.Graph().as_default():
validation_dataset = np.load('path/to/dataset')
validation_data_gen = DataGenerator(validation_dataset,
batch_size=cfg.batch_size)
inputs, labels = placeholder_inputs(cfg.batch_size, cfg.num_frames)
is_training_pl = tf.placeholder(tf.bool, shape=())
keep_prob_pl = tf.placeholder(tf.float32)
global_step = tf.Variable(0, dtype=tf.int64)
bn_decay = get_bn_decay(global_step)
tf.summary.scalar('bn_decay', bn_decay)
# Get model and loss
pred = build_graph(inputs,
is_training_pl,
weight_decay=cfg.weight_decay,
keep_prob=keep_prob_pl,
bn_decay=bn_decay)
loss = get_loss(pred, labels)
tf.summary.scalar('total_loss', loss)
# raise
correct = tf.equal(tf.argmax(pred, 1), tf.to_int64(labels))
correct = tf.reduce_sum(tf.cast(correct, tf.float32))
accuracy = correct / float(cfg.batch_size)
tf.summary.scalar('accuracy', accuracy)
# Get training operator
learning_rate = get_learning_rate(global_step)
tf.summary.scalar('learning_rate', learning_rate)
optimizer = tf.train.AdamOptimizer(learning_rate)
# optimizer = tf.train.MomentumOptimizer(learning_rate, momentum=0.9)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(loss, global_step=global_step)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
# config.log_device_placement = True
# config.gpu_options.allocator_type = "BFC"
sess = tf.Session(config=config)
# # restore model #################
# load_model_path = LOGDIR+'/model_epoch_{}'.format(cfg.load_model_epoch)
load_model_path = LOGDIR + '/model_epoch_{}'.format(
cfg.load_model_epoch)
try:
saver = tf.train.Saver()
saver.restore(sess, load_model_path)
print("\nLoaded previous model... ", load_model_path)
except Exception as e:
raise
if cfg.debug:
sess = tf_debug.LocalCLIDebugWrapperSession(sess)
# Plot Variable Histogram
t_vars = tf.trainable_variables()
# for var in t_vars:
# tf.summary.histogram(var.op.name, var)
# saver = tf.train.Saver()
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(LOGDIR + '/train')
train_writer.add_graph(tf.get_default_graph())
test_writer = tf.summary.FileWriter(LOGDIR + '/test')
test_writer.add_graph(tf.get_default_graph())
# Count number of trainable parameters
num_params = np.sum([np.prod(v.get_shape().as_list()) for v in t_vars])
print(
'************ The Number of Trainable Parameters: {} ************'.
format(num_params))
num_g_params = np.sum(
[np.prod(v.get_shape().as_list()) for v in tf.global_variables()])
print('************ The Number of Global Parameters: {} ************'.
format(num_g_params))
ops = {
'inputs_pl': inputs,
'labels_pl': labels,
'keep_prob_pl': keep_prob_pl,
'is_training_pl': is_training_pl,
'pred': pred,
'loss': loss,
'train_op': train_op,
'merged': merged,
'step': global_step
}
# validation_data_gen = DataGenerator(validation_dataset, batch_size=cfg.batch_size)
print('Validating ...')
val_one_epoch(sess,
validation_data_gen,
ops,
test_writer,
logging=False)