def __init__(self):
self.cache_path = os.path.join(config.DATA_DIR, "MSCOCO", "cache",
"cache.pkl")
self.caption_file_names = {
"eval": "captions_val2014.json",
"train": "captions_train2014.json"
}
self.image_feature_names = {
"eval": "eval_img_feature.pkl",
"test": "test_img_feature.pkl",
"train": "train_img_feature.pkl"
}
DataLoader.__init__(self, MSCOCODataLoader.MSCOCO_text_processor,
self.cache_path)
def test_get_paths_training_set(self):
loader_class = DataLoader(datafolder, True, resample_fs)
file_paths = loader_class.__get_paths__()
expected = [
datafolder + '01/2_raw_data_13-13_22.03.16.txt',
datafolder + '02/2_raw_data_14-21_22.03.16.txt',
datafolder + '03/2_raw_data_09-34_11.04.16.txt'
]
self.assertEqual(file_paths, expected,
'Path of text files with prefix 2_ should be loaded')
self.assertEqual(len(file_paths), 3,
'It should return a total of 3 file paths')
self.assertNotIn('mock_data/EMG_dataset/README.txt', file_paths,
'It should not load any extraneous path')
def main():
args = get_args()
m_config = process_config(args.config)
config = tf.ConfigProto(log_device_placement=True)
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
# create_dirs([config.summary_dir, config.checkpoint_dir])
data = DataLoader(config=m_config)
model = VGG16(data_loader=data, config=m_config)
logger = Logger(sess=sess, config=m_config)
trainer = Trainer(
sess=sess,
model=model,
data=data,
config=m_config,
logger=logger)
trainer.train()
def main(model_config, train_config, track_config):
# GPU Config
gpu_list = train_config['train_data_config'].get('gpu_ids', '0')
num_gpus = len(gpu_list.split(','))
if num_gpus > 1:
os.environ['CUDA_VISIBLE_DEVICES'] = gpu_list
else:
os.environ['CUDA_VISIBLE_DEVICES'] = auto_select_gpu()
# Create training directory which will be used to save: configurations, model files, TensorBoard logs
train_dir = train_config['train_dir']
if not osp.isdir(train_dir):
logging.info('Creating training directory: %s', train_dir)
mkdir_p(train_dir)
g = tf.Graph()
with g.as_default():
# Set fixed seed for reproducible experiments
random.seed(train_config['seed'])
np.random.seed(train_config['seed'])
tf.set_random_seed(train_config['seed'])
#Build global step
with tf.name_scope('train/'):
global_step = tf.Variable(initial_value=0,
name='global_step',
trainable=False,
collections=[
tf.GraphKeys.GLOBAL_STEP,
tf.GraphKeys.GLOBAL_VARIABLES
])
Model = get_model(model_config['Model'])
# build training dataloader and validation dataloader
#---train
train_dataloader = DataLoader(train_config['train_data_config'],
is_training=True)
train_dataloader.build()
train_inputs = train_dataloader.get_one_batch()
#---validation
val_dataloader = DataLoader(train_config['validation_data_config'],
is_training=False)
val_dataloader.build()
val_inputs = val_dataloader.get_one_batch()
# Save configurations for future reference
save_cfgs(train_dir, model_config, train_config, track_config)
if train_config['lr_config'].get('lr_warmup', False):
warmup_epoch_num = 10
init_lr_ratio = 0.8
warmup_steps = warmup_epoch_num * int(
train_config['train_data_config']['num_examples_per_epoch']
) // train_config['train_data_config']['batch_size']
inc_per_step = (
1 - init_lr_ratio
) * train_config['lr_config']['initial_lr'] / warmup_steps
warmup_lr = train_config['lr_config'][
'initial_lr'] * init_lr_ratio + inc_per_step * tf.to_float(
global_step)
learning_rate = tf.cond(
tf.less(global_step,
warmup_steps), lambda: tf.identity(warmup_lr),
lambda: _configure_learning_rate(train_config, global_step -
warmup_steps))
else:
learning_rate = _configure_learning_rate(train_config, global_step)
optimizer = _configure_optimizer(train_config, learning_rate)
tf.summary.scalar('learning_rate', learning_rate)
# Set up the training ops
examplars, instances, gt_examplar_boxes, gt_instance_boxes = tf.split(train_inputs[0],num_gpus), \
tf.split(train_inputs[1],num_gpus), \
tf.split(train_inputs[2],num_gpus), \
tf.split(train_inputs[3],num_gpus)
if train_config['train_data_config'].get('time_decay', False):
time_intervals = tf.split(train_inputs[4], num_gpus)
tower_grads = []
with tf.variable_scope(tf.get_variable_scope()):
for i in range(num_gpus):
with tf.device('/gpu:%d' % i):
if train_config['train_data_config'].get(
'time_decay', False):
inputs = [
examplars[i], instances[i], gt_examplar_boxes[i],
gt_instance_boxes[i], time_intervals[i]
]
else:
inputs = [
examplars[i], instances[i], gt_examplar_boxes[i],
gt_instance_boxes[i]
]
model = tower_model(Model,
inputs,
model_config,
train_config,
mode='train')
# Reuse variables for the next tower.
tf.get_variable_scope().reuse_variables()
grads = optimizer.compute_gradients(model.total_loss)
tower_grads.append(grads)
grads = average_gradients(tower_grads)
#Clip gradient
gradients, tvars = zip(*grads)
clip_gradients, _ = tf.clip_by_global_norm(
gradients, train_config['clip_gradients'])
train_op = optimizer.apply_gradients(zip(clip_gradients, tvars),
global_step=global_step)
#Build validation model
with tf.device('/gpu:0'):
model_va = Model(model_config,
train_config,
mode='validation',
inputs=val_inputs)
model_va.build(reuse=True)
#Save Model setup
saver = tf.train.Saver(
tf.global_variables(),
max_to_keep=train_config['max_checkpoints_to_keep'])
summary_writer = tf.summary.FileWriter(train_dir, g)
summary_op = tf.summary.merge_all()
global_variables_init_op = tf.global_variables_initializer()
local_variables_init_op = tf.local_variables_initializer()
# Dynamically allocate GPU memory
gpu_options = tf.GPUOptions(allow_growth=True)
sess_config = tf.ConfigProto(gpu_options=gpu_options,
allow_soft_placement=True)
#inter_op_parallelism_threads = 16, intra_op_parallelism_threads = 16, log_device_placement=True)
######Debug timeline
if Debug:
from tensorflow.python.client import timeline
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
######Debug timeline
sess = tf.Session(config=sess_config)
model_path = tf.train.latest_checkpoint(train_config['train_dir'])
if not model_path:
sess.run(global_variables_init_op)
sess.run(local_variables_init_op)
start_step = 0
if model_config['embed_config']['embedding_checkpoint_file']:
model.init_fn(sess)
elif model_config['finetuned_checkpoint_file']:
finetuned_checkpoint_file = tf.train.latest_checkpoint(
model_config['finetuned_checkpoint_file'])
logging.info('Restore from last checkpoint: {}'.format(
finetuned_checkpoint_file))
sess.run(local_variables_init_op)
sess.run(global_variables_init_op)
restore_op = tf.contrib.slim.assign_from_checkpoint_fn(
finetuned_checkpoint_file,
tf.global_variables(),
ignore_missing_vars=True)
restore_op(sess)
#reset global step saved in checkpoint
global_step_reset_op = global_step.assign(0)
sess.run(global_step_reset_op)
else:
logging.info('Restore from last checkpoint: {}'.format(model_path))
sess.run(local_variables_init_op)
sess.run(global_variables_init_op)
#saver.restore(sess, model_path)
restore_op = tf.contrib.slim.assign_from_checkpoint_fn(
model_path, tf.global_variables(), ignore_missing_vars=True)
restore_op(sess)
start_step = tf.train.global_step(sess, global_step.name) + 1
print_trainable(sess) #help function, can be disenable
g.finalize() # Finalize graph to avoid adding ops by mistake
# Training loop
data_config = train_config['train_data_config']
total_steps = int(data_config['epoch'] *
data_config['num_examples_per_epoch'] /
data_config['batch_size'])
logging.info('Train for {} steps'.format(total_steps))
for step in range(start_step, total_steps):
try:
start_time = time.time()
if Debug:
_, loss, batch_loss, current_lr = sess.run(
[
train_op, model.total_loss, model.batch_loss,
learning_rate
],
run_metadata=run_metadata,
options=run_options)
t1 = timeline.Timeline(run_metadata.step_stats)
ctf = t1.generate_chrome_trace_format()
with open('timeline.json', 'w') as f:
f.write(ctf)
else:
_, loss, batch_loss, current_lr = sess.run([
train_op, model.total_loss, model.batch_loss,
learning_rate
])
duration = time.time() - start_time
if step % 10 == 0:
examples_per_sec = data_config['batch_size'] / float(
duration)
time_remain = data_config['batch_size'] * (
total_steps - step) / examples_per_sec
current_epoch = (
step * data_config['batch_size']
) // data_config['num_examples_per_epoch'] + 1
m, s = divmod(time_remain, 60)
h, m = divmod(m, 60)
format_str = (
'%s: epoch %d-step %d,lr = %f, total loss = %.3f, batch loss = %.3f (%.1f examples/sec; %.3f '
'sec/batch; %dh:%02dm:%02ds remains)')
logging.info(
format_str %
(datetime.now(), current_epoch, step, current_lr, loss,
batch_loss, examples_per_sec, duration, h, m, s))
if step % 200 == 0:
summary_str = sess.run(summary_op)
summary_writer.add_summary(summary_str, step)
if step % train_config['save_model_every_n_step'] == 0 or (
step + 1) == total_steps:
checkpoint_path = osp.join(train_config['train_dir'],
'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
except KeyboardInterrupt:
checkpoint_path = osp.join(train_config['train_dir'],
'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
print("save model.ckpt-%d" % (step))
break
except:
print(traceback.format_exc())
print("Error found in current step, continue")
def __init__(self):
self.cache_path = os.path.join(Filckr8kImageProcessor.data_dir,
"cache", "cache.pkl")
DataLoader.__init__(self, Flickr8KDataLoader.Flickr8k_text_processor,
self.cache_path)