def after_run(self, run_context, run_values):
predictions_val = run_values.results['predictions']
labels_val = run_values.results['labels']
step = run_values.results['step']
hit_at_one = eval_util.calculate_hit_at_one(predictions_val,
labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(
predictions_val, labels_val)
gap = eval_util.calculate_gap(predictions_val, labels_val)
self.writer.add_summary(
utils.MakeSummary("model/Training_Hit@1", hit_at_one), step)
self.writer.add_summary(utils.MakeSummary("model/Training_Perr", perr),
step)
self.writer.add_summary(utils.MakeSummary("model/Training_GAP", gap),
step)
def run(self, start_new_model=False):
"""Performs training on the currently defined Tensorflow graph.
Returns:
A tuple of the training Hit@1 and the training PERR.
"""
if self.is_master and start_new_model:
self.remove_training_directory(self.train_dir)
target, device_fn = self.start_server_if_distributed()
meta_filename = self.get_meta_filename(start_new_model, self.train_dir)
with tf.Graph().as_default() as graph:
if meta_filename:
saver = self.recover_model(meta_filename)
with tf.device(device_fn):
if not meta_filename:
saver = self.build_model(self.model, self.reader)
global_step = tf.get_collection("global_step")[0]
restart_learning_rate = tf.get_collection(
"restart_learning_rate")[0]
layers_keep_probs = tf.get_collection("layers_keep_probs")[0]
loss = tf.get_collection("loss")[0]
predictions = tf.get_collection("predictions")[0]
labels = tf.get_collection("labels")[0]
train_op = tf.get_collection("train_op")[0]
if FLAGS.use_ema == True:
ema_op = tf.get_collection("ema_op")[0]
e_loss = tf.get_collection("e_loss")[0]
e_labels = tf.get_collection("e_labels")[0]
e_predictions = tf.get_collection("e_predictions")[0]
init_op = tf.global_variables_initializer()
restart_op = tf.assign(restart_learning_rate,
FLAGS.restart_learning_rate)
# getting a proper number of keep_prob parameters for dropout
# max is 10 and we have to pad the vector with 1s
# not the nicest solution, but works
tmp_layers = []
if FLAGS.layers_keep_probs is not None:
tmp_layers = [
float(x) for x in FLAGS.layers_keep_probs.replace(
' ', '').split(',')
]
tmp_layers_padded = tmp_layers + [
1.0 for x in range(10 - len(tmp_layers))
]
with tf.variable_scope("tower", reuse=True) as scope:
keep_op = tf.assign(layers_keep_probs, tmp_layers_padded)
sv = tf.train.Supervisor(graph,
logdir=self.train_dir,
init_op=init_op,
is_chief=self.is_master,
global_step=global_step,
save_model_secs=FLAGS.save_model_minutes * 60,
save_summaries_secs=120,
saver=saver)
logging.info("%s: Starting managed session.",
task_as_string(self.task))
with sv.managed_session(target, config=self.config) as sess:
try:
if FLAGS.restart_learning_rate > 0.0:
sess.run(restart_op)
logging.info("restart learning rate: %f\n" %
FLAGS.restart_learning_rate)
if FLAGS.layers_keep_probs != "1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0":
logging.info("============")
sess.run(keep_op)
logging.info("layers keep probabilites: %s" %
FLAGS.layers_keep_probs)
logging.info("%s: Entering training loop.",
task_as_string(self.task))
while (not sv.should_stop()) and (not self.max_steps_reached):
batch_start_time = time.time()
_, global_step_val, loss_val, predictions_val, labels_val = sess.run(
[train_op, global_step, loss, predictions, labels])
seconds_per_batch = time.time() - batch_start_time
examples_per_second = labels_val.shape[
0] / seconds_per_batch
if self.max_steps and self.max_steps <= global_step_val:
self.max_steps_reached = True
if self.is_master and global_step_val % 50 == 0 and self.train_dir:
eval_start_time = time.time()
hit_at_one = eval_util.calculate_hit_at_one(
predictions_val, labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(
predictions_val, labels_val)
gap = eval_util.calculate_gap(predictions_val,
labels_val)
eval_end_time = time.time()
eval_time = eval_end_time - eval_start_time
logging.info("training step " + str(global_step_val) +
" | Loss: " + ("%.2f" % loss_val) +
" | Hit@1: " + ("%.2f" % hit_at_one) +
" PERR: " + ("%.2f" % perr) + " GAP: " +
("%.4f" % gap))
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Hit@1",
hit_at_one), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Perr", perr),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_GAP", gap),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("global_step/Examples/Second",
examples_per_second),
global_step_val)
#also do eval
e_loss_val, e_predictions_val, e_labels_val = sess.run(
[e_loss, e_predictions, e_labels])
e_hit_at_one = eval_util.calculate_hit_at_one(
e_predictions_val, e_labels_val)
e_perr = eval_util.calculate_precision_at_equal_recall_rate(
e_predictions_val, e_labels_val)
e_gap = eval_util.calculate_gap(
e_predictions_val, e_labels_val)
logging.info("training step " + str(global_step_val) +
" | eLoss: " + ("%.2f" % e_loss_val) +
" |eHit@1: " + ("%.2f" % e_hit_at_one) +
" ePERR: " + ("%.2f" % e_perr) +
" eGAP: " + ("%.4f" % e_gap))
sv.summary_writer.add_summary(
utils.MakeSummary("model/Eval_Hit@1",
e_hit_at_one), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Eval_Perr", e_perr),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Eval_GAP", e_gap),
global_step_val)
sv.summary_writer.flush()
# Exporting the model every x steps
time_to_export = (
(self.last_model_export_step == 0)
or (global_step_val - self.last_model_export_step
>= self.export_model_steps))
if self.is_master and time_to_export:
self.export_model(global_step_val, sv.saver,
sv.save_path, sess)
self.last_model_export_step = global_step_val
else:
logging.info("training step " + str(global_step_val) +
" | Loss: " + ("%.2f" % loss_val) +
" Examples/sec: " +
("%.2f" % examples_per_second))
except tf.errors.OutOfRangeError:
logging.info("%s: Done training -- epoch limit reached.",
task_as_string(self.task))
logging.info("%s: Exited training loop.", task_as_string(self.task))
sv.Stop()
def model_fn(features, labels, mode, params):
is_training = mode == learn.ModeKeys.TRAIN
optimizer_class = find_class_by_name(params.optimizer, [tf.train])
label_loss_fn = find_class_by_name(params.label_loss, [losses])()
model = find_class_by_name(params.model,
[frame_level_models, video_level_models])()
global_step = tf.train.get_or_create_global_step()
learning_rate = tf.train.exponential_decay(
params.base_learning_rate,
global_step * params.batch_size * params.num_towers,
params.learning_rate_decay_examples,
params.learning_rate_decay,
staircase=True,
)
tf.summary.scalar('learning_rate', learning_rate)
optimizer = optimizer_class(learning_rate)
tf.summary.histogram("model/input_raw", features['model_input'])
feature_dim = len(features['model_input'].get_shape()) - 1
model_input = tf.nn.l2_normalize(features['model_input'], feature_dim)
tower_inputs = tf.split(model_input, params.num_towers)
if mode == learn.ModeKeys.INFER:
# ***
# this is a quick hack so that the existing model_fn code,
# taken from train.py, doesn't break in inference (or serving) mode.
# Normally, we would write model_fn such that the 'labels' input arg
# can be None in inference mode, but this existing model code was not written this
# way. See the serving_input_fn() defined below, to see where 'labels_batch'
# is added to the features dict, just to make this code work properly
labels = features['labels_batch']
tower_labels = tf.split(labels, params.num_towers)
tower_num_frames = tf.split(features['num_frames'], params.num_towers)
tower_gradients = []
tower_predictions = []
tower_label_losses = []
tower_reg_losses = []
for i in range(params.num_towers):
# For some reason these 'with' statements can't be combined onto the same
# line. They have to be nested.
with tf.device(params.device_string % i):
with (tf.variable_scope(("tower"), reuse=True if i > 0 else None)):
with (slim.arg_scope([slim.model_variable, slim.variable],
device="/cpu:0"
if params.num_gpus != 1 else "/gpu:0")):
result = model.create_model(
tower_inputs[i],
num_frames=tower_num_frames[i],
vocab_size=params.reader.num_classes,
labels=tower_labels[i],
is_training=is_training)
for variable in slim.get_model_variables():
tf.summary.histogram(variable.op.name, variable)
predictions = result["predictions"]
tower_predictions.append(predictions)
if "loss" in result.keys():
label_loss = result["loss"]
else:
label_loss = label_loss_fn.calculate_loss(
predictions, tower_labels[i])
if "regularization_loss" in result.keys():
reg_loss = result["regularization_loss"]
else:
reg_loss = tf.constant(0.0)
reg_losses = tf.losses.get_regularization_losses()
if reg_losses:
reg_loss += tf.add_n(reg_losses)
tower_reg_losses.append(reg_loss)
# Adds update_ops (e.g., moving average updates in batch normalization) as
# a dependency to the train_op.
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
if "update_ops" in result.keys():
update_ops += result["update_ops"]
if update_ops:
with tf.control_dependencies(update_ops):
barrier = tf.no_op(name="gradient_barrier")
with tf.control_dependencies([barrier]):
label_loss = tf.identity(label_loss)
tower_label_losses.append(label_loss)
final_loss = params.regularization_penalty * reg_loss + label_loss
gradients = optimizer.compute_gradients(
final_loss, colocate_gradients_with_ops=False)
tower_gradients.append(gradients)
pred_dict = {}
label_loss = tf.reduce_mean(tf.stack(tower_label_losses))
predictions = tf.concat(tower_predictions, 0)
pred_dict['predictions'] = predictions
tf.summary.scalar("label_loss", label_loss)
if params.regularization_penalty != 0:
reg_loss = tf.reduce_mean(tf.stack(tower_reg_losses))
tf.summary.scalar("reg_loss", reg_loss)
if is_training:
# Incorporate the L2 weight penalties, etc.
merged_gradients = utils.combine_gradients(tower_gradients)
if params.clip_gradient_norm > 0:
with tf.name_scope('clip_grads'):
merged_gradients = utils.clip_gradient_norms(
merged_gradients, params.clip_gradient_norm)
train_op = optimizer.apply_gradients(merged_gradients,
global_step=global_step)
else:
train_op = None
eval_metric_ops = {}
if mode == learn.ModeKeys.EVAL or is_training:
eval_metric_ops['hit_at_one'] = metrics.streaming_mean(
tf.py_func(
lambda x, y: np.float32(eval_util.calculate_hit_at_one(x, y)),
[predictions, labels],
tf.float32,
stateful=False,
))
eval_metric_ops['perr'] = metrics.streaming_mean(
tf.py_func(
lambda x, y: np.float32(
eval_util.calculate_precision_at_equal_recall_rate(x, y)),
[predictions, labels],
tf.float32,
stateful=False,
))
eval_metric_ops['gap'] = metrics.streaming_mean(
tf.py_func(
lambda x, y: np.float32(eval_util.calculate_gap(x, y)),
[predictions, labels],
tf.float32,
stateful=False,
))
else:
pass
top_predictions, top_indices = tf.nn.top_k(predictions,
_TOP_PREDICTIONS_IN_OUTPUT)
pred_dict['top_predictions'] = top_predictions
pred_dict['top_indices'] = top_indices
#add eval summaries and update ops for training
for key, val in eval_metric_ops.items():
tf.summary.scalar(key, val[0]) #create summary for each eval op
tf.add_to_collection(
tf.GraphKeys.UPDATE_OPS, val[1]
) # add the update op for each eval up to update ops collection, so that it will be run every train_op call
# tf.add_to_collection("global_step", global_step)
# tf.add_to_collection("loss", label_loss)
tf.add_to_collection("predictions", tf.concat(tower_predictions, 0))
# tf.add_to_collection("input_batch_raw", model_input_raw)
# tf.add_to_collection("input_batch", model_input)
# tf.add_to_collection("num_frames", num_frames)
tf.add_to_collection("labels", tf.cast(labels, tf.float32))
# tf.add_to_collection("train_op", train_op)
tf.summary.scalar("loss", label_loss)
export_outputs = {
tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
tf.estimator.export.PredictOutput(pred_dict)
}
return tf.estimator.EstimatorSpec(mode=mode,
predictions=pred_dict,
loss=label_loss,
train_op=train_op,
export_outputs=export_outputs,
eval_metric_ops=eval_metric_ops)
def run(self, start_new_model=False):
"""Performs training on the currently defined Tensorflow graph.
Returns:
A tuple of the training Hit@1 and the training PERR.
"""
if self.is_master and start_new_model:
self.remove_training_directory(self.train_dir)
if not os.path.exists(self.train_dir):
os.makedirs(self.train_dir)
model_flags_dict = {
"model": FLAGS.model,
"feature_sizes": FLAGS.feature_sizes,
"feature_names": FLAGS.feature_names,
"frame_features": FLAGS.frame_features,
"label_loss": FLAGS.label_loss,
}
flags_json_path = os.path.join(FLAGS.train_dir, "model_flags.json")
if os.path.exists(flags_json_path):
existing_flags = json.load(open(flags_json_path))
if existing_flags != model_flags_dict:
logging.error("Model flags do not match existing file %s. Please "
"delete the file, change --train_dir, or pass flag "
"--start_new_model",
flags_json_path)
logging.error("Ran model with flags: %s", str(model_flags_dict))
logging.error("Previously ran with flags: %s", str(existing_flags))
exit(1)
else:
# Write the file.
with open(flags_json_path, "w") as fout:
fout.write(json.dumps(model_flags_dict))
target, device_fn = self.start_server_if_distributed()
meta_filename = self.get_meta_filename(start_new_model, self.train_dir)
with tf.Graph().as_default() as graph:
if meta_filename:
saver = self.recover_model(meta_filename)
with tf.device(device_fn):
if not meta_filename:
saver = self.build_model(self.model, self.reader)
global_step = tf.get_collection("global_step")[0]
loss = tf.get_collection("loss")[0]
predictions = tf.get_collection("predictions")[0]
labels = tf.get_collection("labels")[0]
train_op = tf.get_collection("train_op")[0]
init_op = tf.global_variables_initializer()
if FLAGS.ema_source:
# Here the variables still exsist
ema_op = tf.get_collection("ema_op")[0]
def_vars = tf.get_collection("updatable_vars")
ema_vars = tf.get_collection("ema_vars")
sv = tf.train.Supervisor(
graph,
logdir=self.train_dir,
init_op=init_op,
is_chief=self.is_master,
global_step=global_step,
save_model_secs=15 * 60,
save_summaries_secs=120,
saver=saver)
logging.info("%s: Starting managed session.", task_as_string(self.task))
with sv.managed_session(target, config=self.config) as sess:
if FLAGS.ema_source:
logging.info("%s: Entering training loop.", task_as_string(self.task))
sess.graph._unsafe_unfinalize()
ckpt_reader = pywrap_tensorflow.NewCheckpointReader(FLAGS.ema_source)
for xtensor, ematensor in zip(def_vars, ema_vars):
src_tensor = ckpt_reader.get_tensor(xtensor.name.split(":")[0])
# Loading does not take up graoh space
xtensor.load(src_tensor, session=sess)
ematensor.load(src_tensor, session=sess)
# sess.run(tf.assign(xtensor, src_tensor))
# sess.run(tf.assign(ematensor, src_tensor))
try:
logging.info("%s: Entering training loop.", task_as_string(self.task))
while (not sv.should_stop()) and (not self.max_steps_reached):
batch_start_time = time.time()
_, global_step_val, loss_val, predictions_val, labels_val = sess.run(
[train_op, global_step, loss, predictions, labels])
if FLAGS.ema_source: # Update EMA if needed
_ = sess.run(ema_op)
seconds_per_batch = time.time() - batch_start_time
examples_per_second = labels_val.shape[0] / seconds_per_batch
if self.max_steps and self.max_steps <= global_step_val:
self.max_steps_reached = True
if self.is_master and global_step_val % 10 == 0 and self.train_dir:
eval_start_time = time.time()
hit_at_one = eval_util.calculate_hit_at_one(predictions_val, labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(predictions_val,
labels_val)
gap = eval_util.calculate_gap(predictions_val, labels_val)
eval_end_time = time.time()
eval_time = eval_end_time - eval_start_time
logging.info("training step " + str(global_step_val) + " | Loss: " + ("%.2f" % loss_val) +
" Examples/sec: " + ("%.2f" % examples_per_second) + " | Hit@1: " +
("%.2f" % hit_at_one) + " PERR: " + ("%.2f" % perr) +
" GAP: " + ("%.2f" % gap))
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Hit@1", hit_at_one),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Perr", perr), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_GAP", gap), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("global_step/Examples/Second",
examples_per_second), global_step_val)
sv.summary_writer.flush()
# Exporting the model every x steps
time_to_export = ((self.last_model_export_step == 0) or
(global_step_val - self.last_model_export_step
>= self.export_model_steps))
if self.is_master and time_to_export:
# self.export_model(global_step_val, sv.saver, sv.save_path, sess)
self.last_model_export_step = global_step_val
else:
logging.info("training step " + str(global_step_val) + " | Loss: " +
("%.2f" % loss_val) + " Examples/sec: " + ("%.2f" % examples_per_second))
except tf.errors.OutOfRangeError:
save_name = "{0}/model.ckpt".format(self.train_dir)
saver.save(sess, save_name, global_step_val)
logging.info("Final model export.")
logging.info("%s: Done training -- epoch limit reached.",
task_as_string(self.task))
logging.info("%s: Exited training loop.", task_as_string(self.task))
sv.Stop()
def train():
with tf.Graph().as_default(), tf.device('/cpu:0'):
global_step = tf.Variable(0, trainable=False)
video_id, labels, rgb, audio, num_frames = readers.input(True)
coord = tf.train.Coordinator()
lr = tf.train.exponential_decay(FLAGS.lr,
global_step,
FLAGS.decay_steps,
FLAGS.learning_decay_rate,
staircase=True)
tf.summary.scalar('learning_rate', lr)
opt = tf.train.AdamOptimizer(lr,
beta1=FLAGS.beta1,
beta2=FLAGS.beta2,
epsilon=1e-08,
use_locking=False,
name='Adam')
grads = inference(rgb, audio, num_frames, label=labels, train=True)
loss = tf.get_collection("loss")[0]
predict = tf.get_collection("predict")[0]
tvars = tf.trainable_variables()
for var in tvars:
tf.summary.histogram(var.op.name, var)
for grad, var in grads:
print var.op.name
if grad is not None and type(grad) is not tf.IndexedSlices:
tf.summary.histogram(var.op.name + '/gradients', grad)
elif type(grad) is tf.IndexedSlices:
print "This is a indexslice gradient"
print grad.dense_shape
else:
print "There is a None gradient"
apply_gradient_op = opt.apply_gradients(grads, global_step=global_step)
variable_averages = tf.train.ExponentialMovingAverage(
FLAGS.moving_average_decay, global_step)
variables_averages_op = variable_averages.apply(
tf.trainable_variables())
train_op = tf.group(apply_gradient_op, variables_averages_op)
saver = tf.train.Saver(tf.global_variables())
summary_op = tf.summary.merge_all()
init = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
config = tf.ConfigProto(log_device_placement=False,
allow_soft_placement=True)
config.intra_op_parallelism_threads = 10
config.inter_op_parallelism_threads = 16
sess = tf.Session(config=config)
sess.run(init)
tf.train.start_queue_runners(sess=sess, coord=coord)
summary_writer = tf.summary.FileWriter(FLAGS.train_dir, sess.graph)
ckpt = tf.train.get_checkpoint_state(FLAGS.train_dir)
if not FLAGS.new_model:
saver.restore(sess, ckpt.model_checkpoint_path)
#loader.restore(sess, ckpt.model_checkpoint_path)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
_, loss_value, predict_value, labels_value, num_frames_value = sess.run(
[train_op, loss, predict, labels, num_frames])
duration = time.time() - start_time
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step % 10 == 0:
num_examples_per_step = FLAGS.batch_size
examples_per_sec = num_examples_per_step / duration
sec_per_batch = float(duration)
hit_at_one = eval_util.calculate_hit_at_one(
predict_value, labels_value)
perr = eval_util.calculate_precision_at_equal_recall_rate(
predict_value, labels_value)
gap = eval_util.calculate_gap(predict_value, labels_value)
format_str = (
'%s: step %d, loss = %.2f, hit@one = %.2f, perr = %.2f, gap = %.2f, (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str %
(datetime.now(), step, loss_value, hit_at_one, perr, gap,
examples_per_sec, sec_per_batch))
if step % 100 == 0:
summary_str = sess.run(summary_op)
summary_writer.add_summary(summary_str, step)
summary_writer.add_summary(
utils.MakeSummary("Hit@1", hit_at_one), step)
summary_writer.add_summary(utils.MakeSummary("Perr", perr),
step)
summary_writer.add_summary(utils.MakeSummary("Gap", gap), step)
summary_writer.add_summary(
utils.MakeSummary("example per second", examples_per_sec),
step)
if (step % 1000 == 0
and step != 0) or (step + 1) == FLAGS.max_steps:
checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
def run(self, start_new_model=False):
"""Performs training on the currently defined Tensorflow graph.
Returns:
A tuple of the training Hit@1 and the training PERR.
"""
if self.is_master and start_new_model:
self.remove_training_directory(self.train_dir)
target, device_fn = self.start_server_if_distributed()
meta_filename = self.get_meta_filename(start_new_model, self.train_dir)
with tf.Graph().as_default() as graph:
if meta_filename:
saver = self.recover_model(meta_filename)
with tf.device(device_fn):
if not meta_filename:
saver = self.build_model()
global_step = tf.get_collection("global_step")[0]
loss = tf.get_collection("loss")[0]
predictions = tf.get_collection("predictions")[0]
labels = tf.get_collection("labels")[0]
train_op = tf.get_collection("train_op")[0]
init_op = tf.global_variables_initializer()
sv = tf.train.Supervisor(graph,
logdir=self.train_dir,
init_op=init_op,
is_chief=self.is_master,
global_step=global_step,
save_model_secs=15 * 60,
save_summaries_secs=120,
saver=saver)
logging.info("%s: Starting managed session.",
task_as_string(self.task))
with sv.managed_session(target, config=self.config) as sess:
try:
logging.info("%s: Entering training loop.",
task_as_string(self.task))
while not sv.should_stop():
batch_start_time = time.time()
_, global_step_val, loss_val, predictions_val, labels_val = sess.run(
[train_op, global_step, loss, predictions, labels])
seconds_per_batch = time.time() - batch_start_time
if self.is_master:
examples_per_second = labels_val.shape[
0] / seconds_per_batch
hit_at_one = eval_util.calculate_hit_at_one(
predictions_val, labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(
predictions_val, labels_val)
gap = eval_util.calculate_gap(predictions_val,
labels_val)
logging.info(
"%s: training step " + str(global_step_val) +
"| Hit@1: " + ("%.2f" % hit_at_one) + " PERR: " +
("%.2f" % perr) + " GAP: " +
("%.2f" % gap) + " Loss: " + str(loss_val),
task_as_string(self.task))
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Hit@1",
hit_at_one), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Perr", perr),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_GAP", gap),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("global_step/Examples/Second",
examples_per_second),
global_step_val)
sv.summary_writer.flush()
except tf.errors.OutOfRangeError:
logging.info("%s: Done training -- epoch limit reached.",
task_as_string(self.task))
logging.info("%s: Exited training loop.", task_as_string(self.task))
sv.Stop()
def train_loop(train_dir=None,
saver=None,
is_chief=True,
master="",
start_supervisor_services=True):
"""Performs training on the currently defined tensorflow graph.
Args:
train_dir: Where to save the model checkpoints.
saver: The class to use for serializing the graph variables.
is_chief: Whether this worker is the primary worker (which is responsible
for writing checkpoints and summaries), or an anonymous member of the flock.
master: Which Tensorflow master to listen to.
start_supervisor_services: Whether to start threads for writing summaries
and checkpoints.
Returns:
A tuple of the training Hit@1 and the training PERR.
"""
global_step = tf.get_collection("global_step")[0]
loss = tf.get_collection("loss")[0]
predictions = tf.get_collection("predictions")[0]
labels = tf.get_collection("labels")[0]
train_op = tf.get_collection("train_op")[0]
sv = tf.train.Supervisor(logdir=train_dir,
is_chief=is_chief,
global_step=global_step,
save_model_secs=60,
save_summaries_secs=60,
saver=saver)
sess = sv.prepare_or_wait_for_session(
master,
start_standard_services=start_supervisor_services,
config=tf.ConfigProto(log_device_placement=False))
logging.info("prepared session")
sv.start_queue_runners(sess)
logging.info("started queue runners")
try:
logging.info("entering training loop")
while not sv.should_stop():
batch_start_time = time.time()
_, global_step_val, loss_val, predictions_val, labels_val = sess.run(
[train_op, global_step, loss, predictions, labels])
seconds_per_batch = time.time() - batch_start_time
examples_per_second = labels_val.shape[0] / seconds_per_batch
hit_at_one = eval_util.calculate_hit_at_one(
predictions_val, labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(
predictions_val, labels_val)
gap = eval_util.calculate_gap(predictions_val, labels_val)
logging.info("training step " + str(global_step_val) +
"| Hit@1: " + ("%.2f" % hit_at_one) + " PERR: " +
("%.2f" % perr) + " GAP: " + ("%.2f" % gap) +
" Loss: " + str(loss_val))
if is_chief and global_step_val % 10 == 0 and train_dir:
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Hit@1", hit_at_one),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Perr", perr),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("global_step/Examples/Second",
examples_per_second), global_step_val)
sv.summary_writer.flush()
except tf.errors.OutOfRangeError:
logging.info("Done training -- epoch limit reached")
logging.info("exited training loop")
sv.Stop()
return hit_at_one, perr
def eval():
g1 = tf.Graph()
with g1.as_default():
config = tf.ConfigProto(
allow_soft_placement=True,
log_device_placement=False) # 如果有gpu，则优先用gpu，否则走cpu资源
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
data_object = DataProcessing(args)
valid_batch_example = data_object.input_frame_data(
frame_path=args.valid_path, batch_size=160, num_epoch=1)
model = creat_model(sess, args, isTraining=False) # 构建模型计算图
# app_tag_model函数： 模型计算图的具体实现
ckpt = tf.train.get_checkpoint_state(args.model_dir)
# 判断是否存在模型
if ckpt and tf.train.checkpoint_exists(ckpt.model_checkpoint_path):
# print("Reloading model parameters..")
print("valid step: Reading model parameters from {}".format(
ckpt.model_checkpoint_path))
saver = tf.train.Saver(tf.global_variables())
saver.restore(sess=sess, save_path=ckpt.model_checkpoint_path
) # 调用saver接口，将各个tensor变量的值赋给对应的tensor
print(sess.run(model.global_step))
else:
if not os.path.exists(args.model_dir):
os.makedirs(args.model_dir)
print("valid step: Created new model parameters..")
sess.run(tf.global_variables_initializer())
valid_cate2_perr_list = []
valid_cate2_gap_list = []
valid_total_loss_list = []
try:
while True:
context_parsed, sequence_parsed = sess.run(
valid_batch_example)
batch_origin_labels = [
np.nonzero(row)[0].tolist()
for row in context_parsed['labels']
]
cate1_multilabel, cate2_multilabel, batch_origin_cate1, batch_origin_cate2 = data_object.get_cate1_cate2_label(
batch_origin_labels)
batch_vid_name = np.asarray(context_parsed['id'])
batch_cate1_label_multiHot = np.asarray(
cate1_multilabel) # batch,cate1_nums
batch_cate2_label_multiHot = np.asarray(
cate2_multilabel) # batch,cate2_nums
batch_rgb_fea_float_list = np.asarray(
sequence_parsed['rgb']) # batch,max_frame,1024
batch_audio_fea_float_list = np.asarray(
sequence_parsed['audio']) # batch,max_frame,128
batch_num_audio_rgb_true_frame = np.asarray(
context_parsed['num_audio_rgb_true_frame'])
feed = dict(
zip([
model.input_video_vidName,
model.input_cate1_multilabel,
model.input_cate2_multilabel,
model.input_video_RGB_feature,
model.input_video_Audio_feature,
model.input_rgb_audio_true_frame,
model.dropout_keep_prob
], [
batch_vid_name, batch_cate1_label_multiHot,
batch_cate2_label_multiHot,
batch_rgb_fea_float_list,
batch_audio_fea_float_list,
batch_num_audio_rgb_true_frame, 1.0
]))
cate2_probs, total_loss = sess.run(
[model.cate2_probs, model.total_loss], feed)
cate2_perr = eval_util.calculate_precision_at_equal_recall_rate(
cate2_probs, batch_cate2_label_multiHot)
cate2_gap = eval_util.calculate_gap(
cate2_probs, batch_cate2_label_multiHot)
valid_cate2_perr_list.append(cate2_perr)
valid_cate2_gap_list.append(cate2_gap)
valid_total_loss_list.append(total_loss)
except tf.errors.OutOfRangeError:
print("end!")
valid_cate2_perr_aver_loss = 1.0 * np.sum(
valid_cate2_perr_list) / len(valid_cate2_perr_list)
valid_cate2_gap_aver_loss = 1.0 * np.sum(
valid_cate2_gap_list) / len(valid_cate2_gap_list)
valid_total_valid_aver_loss = 1.0 * np.sum(
valid_total_loss_list) / len(valid_total_loss_list)
print('total valid cate2_perr_aver_loss: %0.4f' %
valid_cate2_perr_aver_loss)
print('total valid cate2_gap_aver_loss: %0.4f' %
valid_cate2_gap_aver_loss)
print('***********************')
print('total valid total_valid_aver_loss: %0.4f' %
valid_total_valid_aver_loss)
return valid_cate2_gap_aver_loss
def run(self, start_new_model=False):
"""Performs training on the currently defined Tensorflow graph.
Returns:
A tuple of the training Hit@1 and the training PERR.
"""
if self.is_master and start_new_model:
self.remove_training_directory(self.train_dir)
target, device_fn = self.start_server_if_distributed()
# meta_filename = self.get_meta_filename(start_new_model, self.train_dir)
with tf.Graph().as_default() as graph:
# if meta_filename:
# saver = self.recover_model(meta_filename)
with tf.device(device_fn):
# if not meta_filename:
self.build_model(self.model, self.reader, self.downloader)
global_step = tf.get_collection("global_step")[0]
loss = tf.get_collection("loss")[0]
predictions = tf.get_collection("predictions")[0]
labels = tf.get_collection("labels")[0]
train_op = tf.get_collection("train_op")[0]
# init_op = tf.global_variables_initializer()
#
saver = tf.train.Saver()
#
config = tf.ConfigProto(
allow_soft_placement=True, log_device_placement=False)
config.gpu_options.allow_growth = True
with tf.Session(config=config,graph=graph) as sess:
enqueue_thread = threading.Thread(target=self.downloader.enqueuing, args=(sess,),
kwargs={'num_epochs': FLAGS.num_epochs})
enqueue_thread.daemon = True
enqueue_thread.start()
latest_checkpoint = tf.train.latest_checkpoint(self.train_dir)
if latest_checkpoint:
logging.info("Loading checkpoint for train: " + latest_checkpoint)
# Restores from checkpoint
saver.restore(sess, latest_checkpoint)
# Assuming model_checkpoint_path looks something like:
# /my-favorite-path/yt8m_train/model.ckpt-0, extract global_step from it.
# global_step_val = latest_checkpoint.split("/")[-1].split("-")[-1]
else:
logging.info("No checkpoint file found.")
sess.run([tf.local_variables_initializer()])
logging.info("%s: Entering training loop.", task_as_string(self.task))
while (not self.max_steps_reached):
batch_start_time = time.time()
_, global_step_val, loss_val, predictions_val, labels_val = sess.run(
[train_op, global_step, loss, predictions, labels])
seconds_per_batch = time.time() - batch_start_time
examples_per_second = labels_val.shape[0] / seconds_per_batch
if self.max_steps and self.max_steps <= global_step_val:
self.max_steps_reached = True
if self.is_master and global_step_val % self.disp_batches == 0 and self.train_dir:
eval_start_time = time.time()
hit_at_one = eval_util.calculate_hit_at_one(predictions_val, labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(predictions_val,
labels_val)
gap = eval_util.calculate_gap(predictions_val, labels_val)
eval_end_time = time.time()
eval_time = eval_end_time - eval_start_time
logging.info("training step " + str(global_step_val) + " | Loss: " + ("%.2f" % loss_val) +
" Examples/sec: " + ("%.2f" % examples_per_second) + " | Hit@1: " +
("%.4f" % hit_at_one) + " PERR: " + ("%.4f" % perr) +
" GAP: " + ("%.4f" % gap))
# sv.summary_writer.add_summary(
# utils.MakeSummary("model/Training_Hit@1", hit_at_one),
# global_step_val)
# sv.summary_writer.add_summary(
# utils.MakeSummary("model/Training_Perr", perr), global_step_val)
# sv.summary_writer.add_summary(
# utils.MakeSummary("model/Training_GAP", gap), global_step_val)
# sv.summary_writer.add_summary(
# utils.MakeSummary("global_step/Examples/Second",
# examples_per_second), global_step_val)
# sv.summary_writer.flush()
# Exporting the model every x steps
time_to_export = ((self.last_model_export_step == 0) or
(global_step_val - self.last_model_export_step
>= self.export_model_steps))
if self.is_master and time_to_export:
self.export_model(global_step_val, saver, self.train_dir, sess)
self.last_model_export_step = global_step_val
# else:
logging.info("training step " + str(global_step_val) + " | Loss: " +
("%.4f" % loss_val) + " Examples/sec: " + ("%.4f" % examples_per_second))
self.downloader.stop()
logging.info('Donwloader stopped')
logging.info("%s: Exited training loop.", task_as_string(self.task))
def run(self, start_new_model=False):
"""Performs training on the currently defined Tensorflow graph.
Returns:
A tuple of the training Hit@1 and the training PERR.
"""
if self.is_master and start_new_model:
self.remove_training_directory(self.train_dir)
if not os.path.exists(self.train_dir):
os.makedirs(self.train_dir)
model_flags_dict = {
"model": FLAGS.model,
"feature_sizes": FLAGS.feature_sizes,
"feature_names": FLAGS.feature_names,
"frame_features": FLAGS.frame_features,
"label_loss": FLAGS.label_loss,
}
flags_json_path = os.path.join(FLAGS.train_dir, "model_flags.json")
if os.path.exists(flags_json_path):
existing_flags = json.load(open(flags_json_path))
if existing_flags != model_flags_dict:
logging.error("Model flags do not match existing file %s. Please "
"delete the file, change --train_dir, or pass flag "
"--start_new_model",
flags_json_path)
logging.error("Ran model with flags: %s", str(model_flags_dict))
logging.error("Previously ran with flags: %s", str(existing_flags))
exit(1)
else:
# Write the file.
with open(flags_json_path, "w") as fout:
fout.write(json.dumps(model_flags_dict))
target, device_fn = self.start_server_if_distributed()
meta_filename = self.get_meta_filename(start_new_model, self.train_dir)
with tf.Graph().as_default() as graph:
if meta_filename:
saver = self.recover_model(meta_filename)
with tf.device(device_fn):
if not meta_filename:
saver = self.build_model(self.model, self.reader)
global_step = tf.get_collection("global_step")[0]
loss = tf.get_collection("loss")[0]
predictions = tf.get_collection("predictions")[0]
labels = tf.get_collection("labels")[0]
train_op = tf.get_collection("train_op")[0]
init_op = tf.global_variables_initializer()
sv = tf.train.Supervisor(
graph,
logdir=self.train_dir,
init_op=init_op,
is_chief=self.is_master,
global_step=global_step,
save_model_secs=15 * 60,
save_summaries_secs=120,
saver=saver)
logging.info("%s: Starting managed session.", task_as_string(self.task))
with sv.managed_session(target, config=self.config) as sess:
try:
logging.info("%s: Entering training loop.", task_as_string(self.task))
while (not sv.should_stop()) and (not self.max_steps_reached):
batch_start_time = time.time()
_, global_step_val, loss_val, predictions_val, labels_val = sess.run(
[train_op, global_step, loss, predictions, labels])
seconds_per_batch = time.time() - batch_start_time
examples_per_second = labels_val.shape[0] / seconds_per_batch
if self.max_steps and self.max_steps <= global_step_val:
self.max_steps_reached = True
if self.is_master and global_step_val % 10 == 0 and self.train_dir:
eval_start_time = time.time()
hit_at_one = eval_util.calculate_hit_at_one(predictions_val, labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(predictions_val,
labels_val)
gap = eval_util.calculate_gap(predictions_val, labels_val)
eval_end_time = time.time()
eval_time = eval_end_time - eval_start_time
logging.info("training step " + str(global_step_val) + " | Loss: " + ("%.2f" % loss_val) +
" Examples/sec: " + ("%.2f" % examples_per_second) + " | Hit@1: " +
("%.2f" % hit_at_one) + " PERR: " + ("%.2f" % perr) +
" GAP: " + ("%.2f" % gap))
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Hit@1", hit_at_one),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Perr", perr), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_GAP", gap), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("global_step/Examples/Second",
examples_per_second), global_step_val)
sv.summary_writer.flush()
# Exporting the model every x steps
time_to_export = ((self.last_model_export_step == 0) or
(global_step_val - self.last_model_export_step
>= self.export_model_steps))
if self.is_master and time_to_export:
self.export_model(global_step_val, sv.saver, sv.save_path, sess)
self.last_model_export_step = global_step_val
else:
logging.info("training step " + str(global_step_val) + " | Loss: " +
("%.2f" % loss_val) + " Examples/sec: " + ("%.2f" % examples_per_second))
except tf.errors.OutOfRangeError:
logging.info("%s: Done training -- epoch limit reached.",
task_as_string(self.task))
logging.info("%s: Exited training loop.", task_as_string(self.task))
sv.Stop()
def run(self, start_new_model=False):
"""Performs training on the currently defined Tensorflow graph.
Returns:
A tuple of the training Hit@1 and the training PERR.
"""
if self.is_master and start_new_model:
self.remove_training_directory(self.train_dir)
target, device_fn = self.start_server_if_distributed()
meta_filename = self.get_meta_filename(start_new_model, self.train_dir)
with tf.Graph().as_default() as graph:
if meta_filename:
saver = self.recover_model(meta_filename)
with tf.device(device_fn):
if not meta_filename:
saver = self.build_model(self.model, self.reader)
global_step = tf.get_collection("global_step")[0]
loss = tf.get_collection("loss")[0]
predictions = tf.get_collection("predictions")[0]
labels = tf.get_collection("labels")[0]
train_op = tf.get_collection("train_op")[0]
input_batch_raw = tf.get_collection("input_batch_raw")[0]
is_neg = tf.get_collection("is_negative")[0]
init_op = tf.global_variables_initializer()
if FLAGS.model == "EmbeddingModel":
hidden_layer_activations = tf.get_collection("hidden_layer_activations")[0]
sv = tf.train.Supervisor(
graph,
logdir=self.train_dir,
init_op=init_op,
is_chief=self.is_master,
global_step=global_step,
save_model_secs=15 * 60,
save_summaries_secs=120,
saver=saver)
logging.info("%s: Starting managed session.", task_as_string(self.task))
with sv.managed_session(target, config=self.config) as sess:
try:
logging.info("%s: Entering training loop.", task_as_string(self.task))
batch_counter = 0
while (not sv.should_stop()) and (not self.max_steps_reached):
batch_counter += 1
batch_start_time = time.time()
if FLAGS.model == "EmbeddingModel":
_, global_step_val, loss_val, predictions_val, labels_val, input_batch_raw_val, embeddings, is_neg_val = sess.run(
[train_op, global_step, loss, predictions, labels, input_batch_raw,
hidden_layer_activations, is_neg])
else:
_, global_step_val, loss_val, predictions_val, labels_val, input_batch_raw_val = sess.run(
[train_op, global_step, loss, predictions, labels, input_batch_raw])
seconds_per_batch = time.time() - batch_start_time
if self.max_steps and self.max_steps <= global_step_val:
self.max_steps_reached = True
if self.is_master:
k = 10
examples_per_second = labels_val.shape[0] / seconds_per_batch
predictions_val = predictions_val[:, 0:4716]
hit_at_one = eval_util.calculate_hit_at_one(predictions_val,
labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(
predictions_val, labels_val)
gap = eval_util.calculate_gap(predictions_val, labels_val)
if FLAGS.model == "EmbeddingModel" \
"":
logging.info(is_neg_val[1])
hit_emb = eval_util.calculate_hit_at_k_embedding(embeddings, k)
logging.info(numpy.sum(numpy.multiply(embeddings[1, 0:FLAGS.embedding_size], embeddings[1, FLAGS.embedding_size:2 * FLAGS.embedding_size])))
logging.info("%s Training step " + str(global_step_val) + "| Hit@1: " +
("%.2f" % hit_at_one) + " HitEmbedding@" + ("%.0f: " % k) + (
"%.2f" % hit_emb) + " GAP: " +
("%.2f" % gap) + " Loss: " + str(loss_val), task_as_string(self.task))
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_HitEmbedding@10", hit_emb),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Hit@1", hit_at_one),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Perr", perr), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_GAP", gap), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("global_step/Examples/Second",
examples_per_second), global_step_val)
sv.summary_writer.flush()
# Exporting the model every x steps
time_to_export = ((self.last_model_export_step == 0) or
(global_step_val - self.last_model_export_step
>= self.export_model_steps))
if self.is_master and time_to_export:
self.export_model(global_step_val, sv.saver, sv.save_path, sess)
self.last_model_export_step = global_step_val
if FLAGS.model == "EmbeddingModel":
if FLAGS.image_server & (batch_counter == 9000):
pred_audio = np.asarray(predictions_val[1, 0:FLAGS.embedding_size])
pred_frames = np.asarray(predictions_val[1, FLAGS.embedding_size:2 * FLAGS.embedding_size])
# plt.bar(range(1, 129), pred_audio / np.linalg.norm(pred_audio))
# plt.savefig("embedding_audio2.png")
# plt.cla()
# plt.bar(range(1, 129), pred_frames / np.linalg.norm(pred_frames))
# plt.savefig("embedding_frames2.png")
# plt.cla()
# plt.bar(range(1, 129),
# pred_frames / np.linalg.norm(pred_frames) - pred_audio / np.linalg.norm(pred_audio))
# plt.savefig("embedding_diferencia2.png")
# plt.cla()
# logging.info("Imatges guardades")
# Exporting the final model
if self.is_master:
self.export_model(global_step_val, sv.saver, sv.save_path, sess)
except tf.errors.OutOfRangeError:
logging.info("%s: Done training -- epoch limit reached.",
task_as_string(self.task))
logging.info("%s: Exited training loop.", task_as_string(self.task))
sv.Stop()
print("Hem acabat")
def run(self, start_new_model=False):
"""Performs training on the currently defined Tensorflow graph.
Returns:
A tuple of the training Hit@1 and the training PERR.
"""
if self.is_master and start_new_model:
self.remove_training_directory(self.train_dir)
if not os.path.exists(self.train_dir):
os.makedirs(self.train_dir)
model_flags_dict = {
"model": FLAGS.model,
"feature_sizes": FLAGS.feature_sizes,
"feature_names": FLAGS.feature_names,
"frame_features": FLAGS.frame_features,
"label_loss": FLAGS.label_loss,
}
flags_json_path = os.path.join(FLAGS.train_dir, "model_flags.json")
if file_io.file_exists(flags_json_path):
existing_flags = json.load(
file_io.FileIO(flags_json_path, mode="r"))
if existing_flags != model_flags_dict:
logging.error(
"Model flags do not match existing file %s. Please "
"delete the file, change --train_dir, or pass flag "
"--start_new_model", flags_json_path)
logging.error("Ran model with flags: %s",
str(model_flags_dict))
logging.error("Previously ran with flags: %s",
str(existing_flags))
exit(1)
else:
# Write the file.
with file_io.FileIO(flags_json_path, mode="w") as fout:
fout.write(json.dumps(model_flags_dict))
target, device_fn = self.start_server_if_distributed()
meta_filename = self.get_meta_filename(start_new_model, self.train_dir)
with tf.Graph().as_default() as graph:
if meta_filename:
saver = self.recover_model(meta_filename)
with tf.device(device_fn):
if not meta_filename:
saver = self.build_model()
global_step = tf.get_collection("global_step")[0]
loss = tf.get_collection("loss")[0]
predictions = tf.get_collection("predictions")[0]
labels = tf.get_collection("labels")[0]
train_op = tf.get_collection("train_op")[0]
init_op = tf.global_variables_initializer()
if FLAGS.dropout:
keep_prob_tensor = tf.get_collection("keep_prob")[0]
if FLAGS.noise_level > 0:
noise_level_tensor = tf.get_collection("noise_level")[0]
if FLAGS.reweight:
weights_input, weights_assignment = None, None
if len(tf.get_collection("weights_input")) > 0:
weights_input = tf.get_collection("weights_input")[0]
weights_assignment = tf.get_collection(
"weights_assignment")[0]
sv = tf.train.Supervisor(
graph,
logdir=self.train_dir,
init_op=init_op,
is_chief=self.is_master,
global_step=global_step,
save_model_secs=FLAGS.keep_checkpoint_interval * 60,
save_summaries_secs=120,
saver=saver)
logging.info("%s: Starting managed session.",
task_as_string(self.task))
with sv.managed_session(target, config=self.config) as sess:
# re-assign weights
if FLAGS.reweight:
optional_assign_weights(sess, weights_input,
weights_assignment)
steps = 0
try:
logging.info("%s: Entering training loop.",
task_as_string(self.task))
while not sv.should_stop():
steps += 1
batch_start_time = time.time()
custom_feed = {}
if FLAGS.dropout:
custom_feed[keep_prob_tensor] = FLAGS.keep_prob
if FLAGS.noise_level > 0:
custom_feed[noise_level_tensor] = FLAGS.noise_level
_, global_step_val, loss_val, predictions_val, labels_val = sess.run(
[train_op, global_step, loss, predictions, labels],
feed_dict=custom_feed)
seconds_per_batch = time.time() - batch_start_time
if self.is_master:
examples_per_second = labels_val.shape[
0] / seconds_per_batch
hit_at_one = eval_util.calculate_hit_at_one(
predictions_val, labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(
predictions_val, labels_val)
recall = "N/A"
if False:
recall = eval_util.calculate_recall_at_n(
predictions_val, labels_val, FLAGS.recall_at_n)
sv.summary_writer.add_summary(
utils.MakeSummary(
"model/Training_Recall@%d" %
FLAGS.recall_at_n, recall),
global_step_val)
recall = "%.2f" % recall
gap = eval_util.calculate_gap(predictions_val,
labels_val)
logging.info(
"%s: training step " + str(global_step_val) +
"| Hit@1: " + ("%.2f" % hit_at_one) + " PERR: " +
("%.2f" % perr) + " GAP: " + ("%.2f" % gap) +
" Recall@%d: " % FLAGS.recall_at_n +
recall + " Loss: " + str(loss_val),
task_as_string(self.task))
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Hit@1",
hit_at_one), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Perr", perr),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_GAP", gap),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("global_step/Examples/Second",
examples_per_second),
global_step_val)
sv.summary_writer.flush()
if FLAGS.max_steps is not None and steps > FLAGS.max_steps:
logging.info(
"%s: Done training -- max_steps limit reached.",
task_as_string(self.task))
break
except tf.errors.OutOfRangeError:
logging.info("%s: Done training -- epoch limit reached.",
task_as_string(self.task))
logging.info("%s: Exited training loop.", task_as_string(self.task))
sv.Stop()
def run(self, start_new_model=False):
"""Performs training on the currently defined Tensorflow graph.
Returns:
A tuple of the training Hit@1 and the training PERR.
"""
if self.is_master and start_new_model:
self.remove_training_directory(self.train_dir)
if not os.path.exists(self.train_dir):
os.makedirs(self.train_dir)
logging.info(
'############## PARAMETERS ##############################')
logging.info("feature_names {}".format(FLAGS.feature_names))
logging.info("feature_sizes {}".format(FLAGS.feature_sizes))
logging.info("frame_features {}".format(FLAGS.frame_features))
logging.info("model {}".format(FLAGS.model))
logging.info("start_new_model {}".format(FLAGS.start_new_model))
logging.info("num_gpu {}".format(FLAGS.num_gpu))
logging.info("batch_size {}".format(FLAGS.batch_size))
logging.info("label_loss {}".format(FLAGS.label_loss))
logging.info("regularization_penalty {}".format(
FLAGS.regularization_penalty))
logging.info("base_learning_rate {}".format(FLAGS.base_learning_rate))
logging.info("learning_rate_decay {}".format(
FLAGS.learning_rate_decay))
logging.info("learning_rate_decay_examples {}".format(
FLAGS.learning_rate_decay_examples))
logging.info("num_epochs {}".format(FLAGS.num_epochs))
logging.info("max_steps {}".format(FLAGS.max_steps))
logging.info("export_model_steps {}".format(FLAGS.export_model_steps))
logging.info("num_readers {}".format(FLAGS.num_readers))
logging.info("optimizer {}".format(FLAGS.optimizer))
logging.info("clip_gradient_norm {}".format(FLAGS.clip_gradient_norm))
logging.info(
'########################################################')
logging.info(' '.join([x for x in sys.argv]))
model_flags_dict = {
"model": FLAGS.model,
"feature_sizes": FLAGS.feature_sizes,
"feature_names": FLAGS.feature_names,
"frame_features": FLAGS.frame_features,
"label_loss": FLAGS.label_loss,
}
flags_json_path = os.path.join(FLAGS.train_dir, "model_flags.json")
if os.path.exists(flags_json_path):
existing_flags = json.load(open(flags_json_path))
if existing_flags != model_flags_dict:
logging.error(
"Model flags do not match existing file %s. Please "
"delete the file, change --train_dir, or pass flag "
"--start_new_model", flags_json_path)
logging.error("Ran model with flags: %s",
str(model_flags_dict))
logging.error("Previously ran with flags: %s",
str(existing_flags))
exit(1)
else:
# Write the file.
with open(flags_json_path, "w") as fout:
fout.write(json.dumps(model_flags_dict))
target, device_fn = self.start_server_if_distributed()
meta_filename = self.get_meta_filename(start_new_model, self.train_dir)
with tf.Graph().as_default() as graph:
with tf.device(device_fn):
saver = self.build_model(self.model, self.reader)
global_step = tf.get_collection("global_step")[0]
loss = tf.get_collection("loss")[0]
predictions = tf.get_collection("predictions")[0]
labels = tf.get_collection("labels")[0]
train_op = tf.get_collection("train_op")[0]
init_op, init_fn = None, None
if meta_filename:
saver = tf.train.Saver(tf.global_variables(),
max_to_keep=0,
keep_checkpoint_every_n_hours=0.25)
def init_fn(sess):
return saver.restore(sess, meta_filename)
else:
init_op = tf.global_variables_initializer()
sv = tf.train.Supervisor(graph,
logdir=self.train_dir,
init_op=init_op,
init_fn=init_fn,
is_chief=self.is_master,
global_step=global_step,
save_model_secs=40 * 60,
save_summaries_secs=120,
saver=saver)
logging.info("%s: Starting managed session.",
task_as_string(self.task))
with sv.managed_session(target, config=self.config) as sess:
try:
logging.info("%s: Entering training loop.",
task_as_string(self.task))
while (not sv.should_stop()) and (not self.max_steps_reached):
batch_start_time = time.time()
_, global_step_val, loss_val, predictions_val, labels_val = sess.run(
[train_op, global_step, loss, predictions, labels])
seconds_per_batch = time.time() - batch_start_time
examples_per_second = labels_val.shape[
0] / seconds_per_batch
if self.max_steps and self.max_steps <= global_step_val:
self.max_steps_reached = True
if self.is_master and global_step_val % 10 == 0 and self.train_dir:
eval_start_time = time.time()
hit_at_one = eval_util.calculate_hit_at_one(
predictions_val, labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(
predictions_val, labels_val)
gap = eval_util.calculate_gap(predictions_val,
labels_val)
eval_end_time = time.time()
eval_time = eval_end_time - eval_start_time
logging.info("training step " + str(global_step_val) +
" | Loss: " + ("%.2f" % loss_val) +
" Examples/sec: " +
("%.2f" % examples_per_second) +
" | Hit@1: " + ("%.2f" % hit_at_one) +
" PERR: " + ("%.2f" % perr) + " GAP: " +
("%.2f" % gap))
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Hit@1",
hit_at_one), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Perr", perr),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_GAP", gap),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("global_step/Examples/Second",
examples_per_second),
global_step_val)
sv.summary_writer.flush()
# Exporting the model every x steps
time_to_export = (
(self.last_model_export_step == 0)
or (global_step_val - self.last_model_export_step
>= self.export_model_steps))
if self.is_master and time_to_export:
self.export_model(global_step_val, sv.saver,
sv.save_path, sess)
self.last_model_export_step = global_step_val
else:
logging.info("training step " + str(global_step_val) +
" | Loss: " + ("%.2f" % loss_val) +
" Examples/sec: " +
("%.2f" % examples_per_second))
except tf.errors.OutOfRangeError:
logging.info("%s: Done training -- epoch limit reached.",
task_as_string(self.task))
logging.info("%s: Exited training loop.", task_as_string(self.task))
sv.Stop()
def main():
env = json.loads(os.environ.get("TF_CONFIG", "{}"))
task_data = env.get("task", None) or {"type": "master", "index": 0}
task = type("TaskSpec", (object, ), task_data)
logging.set_verbosity(tf.logging.INFO)
logging.info("%s: Tensorflow version: %s.", task_as_string(task),
tf.__version__)
video_ids, video_features, video_labels, video_frames = gen_input(
data_pattern,
reader_batch_size=reader_batch_size,
num_classes=num_classes,
num_readers=num_readers,
mini_batch_size=mini_batch_size)
result = gen_model(model_input=video_features,
vocab_size=num_classes,
labels=video_labels,
num_frames=video_frames)
predictions = result["predictions"]
global_step = tf.Variable(0, trainable=False, name="global_step")
label_loss = label_loss_fn.calculate_loss(predictions, video_labels)
if "regularization_loss" in result.keys():
reg_loss = result["regularization_loss"]
else:
reg_loss = tf.constant(0.0)
reg_losses = tf.losses.get_regularization_losses()
if reg_losses:
reg_loss += tf.add_n(reg_losses)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
if "update_ops" in result.keys():
update_ops += result["update_ops"]
if update_ops:
with tf.control_dependencies(update_ops):
barrier = tf.no_op(name="gradient_barrier")
with tf.control_dependencies([barrier]):
label_loss = tf.identity(label_loss)
final_loss = regularization_penalty * reg_loss + label_loss
learning_rate = tf.train.exponential_decay(base_learning_rate,
global_step * mini_batch_size *
num_towers,
learning_rate_decay_examples,
learning_rate_decay,
staircase=True)
tf.summary.scalar('learning_rate', learning_rate)
optimizer = optimizer_class(learning_rate)
gradients = optimizer.compute_gradients(final_loss,
colocate_gradients_with_ops=False)
tf.summary.scalar("label_loss", label_loss)
tf.summary.scalar("reg_loss", reg_loss)
if clip_gradient_norm > 0:
with tf.name_scope('clip_grads'):
gradients = utils.clip_gradient_norms(gradients,
clip_gradient_norm)
train_op = optimizer.apply_gradients(gradients, global_step=global_step)
with tf.Session() as sess:
tf.global_variables_initializer().run()
tf.local_variables_initializer().run()
#init_local_op = tf.local_variables_initializer()
#sess.run(init_local_op)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
total_step = 0
try:
while total_step < 100000:
batch_start_time = time.time()
# v_ids, v_features, v_labels, v_frames = sess.run([video_ids, video_features, video_labels, video_frames])
_, global_step_val, loss_val, predictions_val, labels_val = sess.run(
[
train_op, global_step, label_loss, predictions,
tf.cast(video_labels, tf.float32)
])
seconds_per_batch = time.time() - batch_start_time
examples_per_second = labels_val.shape[0] / seconds_per_batch
# if max_steps <= global_step_val:
# max_steps_reached = True
# print(v_features.shape)
# print(v_ids)
if total_step % 10 == 0:
eval_start_time = time.time()
hit_at_one = eval_util.calculate_hit_at_one(
predictions_val, labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(
predictions_val, labels_val)
gap = eval_util.calculate_gap(predictions_val, labels_val)
eval_end_time = time.time()
eval_time = eval_end_time - eval_start_time
logging.info("training step " + str(global_step_val) +
" | Loss: " + ("%.2f" % loss_val) +
" Examples/sec: " +
("%.2f" % examples_per_second) +
" | Hit@1: " + ("%.2f" % hit_at_one) +
" PERR: " + ("%.2f" % perr) + " GAP: " +
("%.2f" % gap))
else:
logging.info("training step " + str(global_step_val) +
" | Loss: " + ("%.2f" % loss_val) +
" Examples/sec: " +
("%.2f" % examples_per_second))
total_step = total_step + 1
except tf.errors.OutOfRangeError:
logging.info("%s: Done training -- epoch limit reached.",
task_as_string(task))
coord.request_stop()
coord.join(threads)
def run(self, start_new_model=False):
"""Performs training on the currently defined Tensorflow graph.
Returns:
A tuple of the training Hit@1 and the training PERR.
"""
if self.is_master and start_new_model:
self.remove_training_directory(self.train_dir)
target, device_fn = self.start_server_if_distributed()
meta_filename = self.get_meta_filename(start_new_model, self.train_dir)
with tf.Graph().as_default() as graph:
if meta_filename:
saver = self.recover_model(meta_filename)
with tf.device(device_fn):
if not meta_filename:
saver = self.build_model(self.model, self.reader)
global_step = tf.get_collection("global_step")[0]
loss = tf.get_collection("loss")[0]
input_batch_raw = tf.get_collection("input_batch_raw")[0]
input_batch = tf.get_collection("input_batch")[0]
model_input_raw_ph = tf.get_collection("model_input_raw_ph")[0]
predictions = tf.get_collection("predictions")[0]
labels = tf.get_collection("labels")[0]
train_op = tf.get_collection("train_op")[0]
num_frames = tf.get_collection("num_frames")[0]
num_frames_ph = tf.get_collection("num_frames_ph")[0]
learning_rate = tf.get_collection("learning_rate")[0]
init_op = tf.global_variables_initializer()
sv = tf.train.Supervisor(graph,
logdir=self.train_dir,
init_op=init_op,
is_chief=self.is_master,
global_step=global_step,
save_model_secs=15 * 60,
save_summaries_secs=120,
saver=saver)
logging.info("%s: Starting managed session.",
task_as_string(self.task))
with sv.managed_session(target, config=self.config) as sess:
try:
logging.info("%s: Entering training loop.",
task_as_string(self.task))
while (not sv.should_stop()) and (not self.max_steps_reached):
batch_start_time = time.time()
model_input_raw_val, num_frames_val, learning_rate_val = sess.run(
[input_batch_raw, num_frames, learning_rate])
pr_feature = []
pr_num = []
for i in range(model_input_raw_val.shape[0]):
if num_frames_val[i] / FLAGS.max_scene <= 2:
num_tmp = num_frames_val[i] * np.ceil(
1 + FLAGS.max_scene / num_frames_val[i])
input_tmp = model_input_raw_val[
i][:num_frames_val[i]]
input_tmp = np.repeat(
input_tmp,
np.ceil(1 +
FLAGS.max_scene / num_frames_val[i]),
0)
else:
num_tmp = num_frames_val[i]
input_tmp = model_input_raw_val[i][:num_tmp]
numvec = (input_tmp[1:] *
input_tmp[:-1]).sum(axis=1) / (np.sqrt(
(input_tmp[1:]**2).sum(1)) * (np.sqrt(
(input_tmp[:-1]**2).sum(1))))
idx = np.sort(
numvec.argpartition(FLAGS.max_scene -
1)[:FLAGS.max_scene - 1] + 1)
example_splits = np.split(input_tmp, idx, 0)
example_splits_mean = [
np.mean(example_split, 0)
for example_split in example_splits
]
example_splits_mean = np.stack(example_splits_mean, 0)
pr_num.append(FLAGS.max_scene)
pr_feature.append(example_splits_mean)
pr_feature = np.stack(pr_feature, 0)
pr_num = np.stack(pr_num, 0)
_, global_step_val, loss_val, predictions_val, labels_val = sess.run(
[train_op, global_step, loss, predictions, labels],
feed_dict={
model_input_raw_ph: pr_feature,
num_frames_ph: pr_num
})
seconds_per_batch = time.time() - batch_start_time
examples_per_second = labels_val.shape[
0] / seconds_per_batch
if self.max_steps and self.max_steps <= global_step_val:
self.max_steps_reached = True
if self.is_master and global_step_val % 10 == 0 and self.train_dir:
eval_start_time = time.time()
hit_at_one = eval_util.calculate_hit_at_one(
predictions_val, labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(
predictions_val, labels_val)
gap = eval_util.calculate_gap(predictions_val,
labels_val)
eval_end_time = time.time()
eval_time = eval_end_time - eval_start_time
logging.info("training step " + str(global_step_val) +
" | Loss: " + ("%.2f" % loss_val) +
" Examples/sec: " +
("%.2f" % examples_per_second) +
" | Hit@1: " + ("%.2f" % hit_at_one) +
" PERR: " + ("%.2f" % perr) + " GAP: " +
("%.2f" % gap))
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Hit@1",
hit_at_one), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Perr", perr),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_GAP", gap),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("global_step/Examples/Second",
examples_per_second),
global_step_val)
sv.summary_writer.flush()
# Exporting the model every x steps
time_to_export = (
(self.last_model_export_step == 0)
or (global_step_val - self.last_model_export_step
>= self.export_model_steps))
if self.is_master and time_to_export:
self.export_model(global_step_val, sv.saver,
sv.save_path, sess)
self.last_model_export_step = global_step_val
else:
logging.info("training step " + str(global_step_val) +
" | Loss: " + ("%.2f" % loss_val) +
" Examples/sec: " +
("%.2f" % examples_per_second))
except tf.errors.OutOfRangeError:
logging.info("%s: Done training -- epoch limit reached.",
task_as_string(self.task))
logging.info("%s: Exited training loop.", task_as_string(self.task))
sv.Stop()
def run(self, start_new_model=False):
"""Performs training on the currently defined Tensorflow graph.
Returns:
A tuple of the training Hit@1 and the training PERR.
"""
if self.is_master and start_new_model:
self.remove_training_directory(self.train_dir)
if not os.path.exists(self.train_dir):
os.makedirs(self.train_dir)
model_flags_dict = {
"model": FLAGS.model,
"feature_sizes": FLAGS.feature_sizes,
"feature_names": FLAGS.feature_names,
"frame_features": FLAGS.frame_features,
"label_loss": FLAGS.label_loss,
}
flags_json_path = os.path.join(FLAGS.train_dir, "model_flags.json")
if os.path.exists(flags_json_path):
existing_flags = json.load(open(flags_json_path))
if existing_flags != model_flags_dict:
logging.error("Model flags do not match existing file %s. Please "
"delete the file, change --train_dir, or pass flag "
"--start_new_model",
flags_json_path)
logging.error("Ran model with flags: %s", str(model_flags_dict))
logging.error("Previously ran with flags: %s", str(existing_flags))
exit(1)
else:
# Write the file.
with open(flags_json_path, "w") as fout:
fout.write(json.dumps(model_flags_dict))
target, device_fn = self.start_server_if_distributed()
meta_filename = self.get_meta_filename(start_new_model, self.train_dir)
with tf.Graph().as_default() as graph:
if meta_filename:
saver = self.recover_model(meta_filename)
with tf.device(device_fn):
if not meta_filename:
saver = self.build_model(self.model, self.reader)
global_step = tf.get_collection("global_step")[0]
loss = tf.get_collection("loss")[0]
predictions = tf.get_collection("predictions")[0]
labels = tf.get_collection("labels")[0]
labels_origin = tf.get_collection("original_labels")[0]
sampling_distribution = tf.get_collection("sampling_probs")[0]
train_op = tf.get_collection("train_op")[0]
init_op = tf.global_variables_initializer()
sv = tf.train.Supervisor(
graph,
logdir=self.train_dir,
init_op=init_op,
is_chief=self.is_master,
global_step=global_step,
#save_model_secs=15 * 60,
save_model_secs=int(FLAGS.save_checkpoint_every_n_hour * 3600),
#save_summaries_secs=120,
save_summaries_secs=int(FLAGS.save_checkpoint_every_n_hour * 3600),
saver=saver)
logging.info("%s: Starting managed session.", task_as_string(self.task))
with sv.managed_session(target, config=self.config) as sess:
try:
logging.info("%s: Entering training loop.", task_as_string(self.task))
while (not sv.should_stop()) and (not self.max_steps_reached):
batch_start_time = time.time()
_, global_step_val, loss_val, predictions_val, labels_val, labels_origin_val, sampling_distribution_val = sess.run(
[train_op, global_step, loss, predictions, labels, labels_origin, sampling_distribution])
seconds_per_batch = time.time() - batch_start_time
examples_per_second = labels_val.shape[0] / seconds_per_batch
## check on the correctness of label sampling
#if not np.all(labels_val[0]==labels_origin_val[0]):
# print(np.where(labels_val[0])[0], " --- ", np.where(labels_origin_val[0])[0])
#print(np.all(labels_val == labels_origin_val, 1).mean())
#temp_bool = np.sum(labels_val, 1)<1
#if np.any(temp_bool):
# print("\n\n\nFOUND!!", np.where(labels_val[temp_bool])[0], "\n\n\n")
#print(sampling_distribution_val[:5, :5])
if self.max_steps and self.max_steps <= global_step_val:
self.max_steps_reached = True
#if self.is_master and global_step_val % 10 == 0 and self.train_dir:
if self.is_master and global_step_val % FLAGS.validate_every_n_training_steps == 0 and self.train_dir:
eval_start_time = time.time()
hit_at_one = eval_util.calculate_hit_at_one(predictions_val, labels_origin_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(predictions_val,
labels_origin_val)
gap = eval_util.calculate_gap(predictions_val, labels_origin_val)
eval_end_time = time.time()
eval_time = eval_end_time - eval_start_time
logging.info("training step " + str(global_step_val) + " | Loss: " + ("%.2f" % loss_val) +
" Examples/sec: " + ("%.2f" % examples_per_second) + " | Hit@1: " +
("%.2f" % hit_at_one) + " PERR: " + ("%.2f" % perr) +
" GAP: " + ("%.2f" % gap))
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Hit@1", hit_at_one),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Perr", perr), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_GAP", gap), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("global_step/Examples/Second",
examples_per_second), global_step_val)
sv.summary_writer.flush()
with open(FLAGS.train_dir + '/global_step_{%d}_training_GAP_{%.6f}.txt' % (global_step_val, gap), 'w') as f:
f.write('\n')
# Exporting the model every x steps
time_to_export = ((self.last_model_export_step == 0) or
(global_step_val - self.last_model_export_step
>= self.export_model_steps))
if self.is_master and time_to_export:
self.export_model(global_step_val, sv.saver, sv.save_path, sess)
self.last_model_export_step = global_step_val
else:
#logging.info("training step " + str(global_step_val) + " | Loss: " +
#("%.2f" % loss_val) + " Examples/sec: " + ("%.2f" % examples_per_second))
continue
except tf.errors.OutOfRangeError:
logging.info("%s: Done training -- epoch limit reached.",
task_as_string(self.task))
logging.info("%s: Exited training loop.", task_as_string(self.task))
sv.Stop()
def train():
'''
模型，训练任务
:return:
'''
valid_max_accuracy = -9999
config = tf.ConfigProto(
allow_soft_placement=True,
log_device_placement=False) # 如果有gpu，则优先用gpu，否则走cpu资源
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
data_object = DataProcessing(args)
train_batch_example = data_object.input_frame_data(
frame_path=args.train_path,
batch_size=args.batch_size,
num_epoch=args.epoch)
model = creat_model(sess, args, isTraining=True) # 构建模型计算图
saver = tf.train.Saver(tf.global_variables(),
max_to_keep=3) # max_to_keep 表征只保留最好的3个模型
print("Begin training..")
train_cate1_perr_list = []
train_cate1_gap_list = []
train_cate2_perr_list = []
train_cate2_gap_list = []
train_total_loss_list = []
try:
while True:
context_parsed, sequence_parsed = sess.run(train_batch_example)
batch_origin_labels = [
np.nonzero(row)[0].tolist()
for row in context_parsed['labels']
]
cate1_multilabel, cate2_multilabel, batch_origin_cate1, batch_origin_cate2 = data_object.get_cate1_cate2_label(
batch_origin_labels)
batch_vid_name = np.asarray(context_parsed['id'])
batch_num_audio_rgb_true_frame = np.asarray(
context_parsed['num_audio_rgb_true_frame'])
batch_cate1_label_multiHot = np.asarray(
cate1_multilabel) # batch,cate1_nums
batch_cate2_label_multiHot = np.asarray(
cate2_multilabel) # batch,cate2_nums
batch_rgb_fea_float_list = np.asarray(
sequence_parsed['rgb']) # batch,max_frame,1024
batch_audio_fea_float_list = np.asarray(
sequence_parsed['audio']) # batch,max_frame,1
# print("batch_vid_name.shape:",batch_vid_name.shape)
# print("batch_cate1_label_multiHot.shape: ",batch_cate1_label_multiHot.shape)
# print("batch_cate2_label_multiHot.shape: ",batch_cate2_label_multiHot.shape)
# print("batch_rgb_fea_float_list.shape: ",batch_rgb_fea_float_list.shape)
# print("batch_audio_fea_float_list.shape: ",batch_audio_fea_float_list.shape)
# print("batch_num_audio_rgb_true_frame: ",batch_num_audio_rgb_true_frame)
# print("batch_num_audio_rgb_true_frame: ", np.asarray(batch_num_audio_rgb_true_frame).shape)
# assert 1==2
feed = dict(
zip([
model.input_video_vidName,
model.input_cate1_multilabel,
model.input_cate2_multilabel,
model.input_video_RGB_feature,
model.input_video_Audio_feature,
model.input_rgb_audio_true_frame,
model.dropout_keep_prob
], [
batch_vid_name, batch_cate1_label_multiHot,
batch_cate2_label_multiHot, batch_rgb_fea_float_list,
batch_audio_fea_float_list,
batch_num_audio_rgb_true_frame, 0.5
]))
cate1_probs, cate2_probs, total_loss, _ = sess.run([
model.cate1_probs, model.cate2_probs, model.total_loss,
model.optimizer
], feed)
train_cate1_perr = eval_util.calculate_precision_at_equal_recall_rate(
cate1_probs, batch_cate1_label_multiHot)
train_cate1_gap = eval_util.calculate_gap(
cate1_probs, batch_cate1_label_multiHot)
train_cate2_perr = eval_util.calculate_precision_at_equal_recall_rate(
cate2_probs, batch_cate2_label_multiHot)
train_cate2_gap = eval_util.calculate_gap(
cate2_probs, batch_cate2_label_multiHot)
train_cate1_perr_list.append(train_cate1_perr)
train_cate1_gap_list.append(train_cate1_gap)
train_cate2_perr_list.append(train_cate2_perr)
train_cate2_gap_list.append(train_cate2_gap)
train_total_loss_list.append(total_loss)
if model.global_step.eval() % args.report_freq == 0:
print("report_freq: ", args.report_freq)
print(
'cate1_train: Step:{} ; aver_train_cate1_perr:{} ; aver_train_cate1_gap_list:{} ; aver_total_loss:{}'
.format(
model.global_step.eval(),
1.0 * np.sum(train_cate1_perr_list) /
len(train_cate1_perr_list),
1.0 * np.sum(train_cate1_gap_list) /
len(train_cate1_gap_list),
1.0 * np.sum(train_total_loss_list) /
len(train_total_loss_list)))
print(
'cate2_train: Step:{} ; aver_train_cate2_perr:{} ; aver_train_cate2_gap_list:{} ; aver_total_loss:{}'
.format(
model.global_step.eval(),
1.0 * np.sum(train_cate2_perr_list) /
len(train_cate2_perr_list),
1.0 * np.sum(train_cate2_gap_list) /
len(train_cate2_gap_list),
1.0 * np.sum(train_total_loss_list) /
len(train_total_loss_list)))
train_cate1_perr_list = []
train_cate1_gap_list = []
train_cate2_perr_list = []
train_cate2_gap_list = []
train_total_loss_list = []
if model.global_step.eval(
) > 1 and model.global_step.eval() % args.valid_freq == 0:
# 统计验证集的准确率
print("valid infer is process 111!")
print('model.global_step.eval(): ',
model.global_step.eval())
valid_cate2_gap_aver_loss = eval()
# 保存当前验证集合准确率最高的模型
if valid_cate2_gap_aver_loss > valid_max_accuracy:
print("save the model, step= : ",
model.global_step.eval())
valid_max_accuracy = valid_cate2_gap_aver_loss
checkpoint_path = os.path.join(args.model_dir,
'model.ckpt')
saver.save(sess=sess,
save_path=checkpoint_path,
global_step=model.global_step.eval())
except tf.errors.OutOfRangeError:
print("train processing is finished!")
def run(self, start_new_model=False):
"""Performs training on the currently defined Tensorflow graph.
Returns:
A tuple of the training Hit@1 and the training PERR.
"""
if self.is_master and start_new_model:
self.remove_training_directory(self.train_dir)
target, device_fn = self.start_server_if_distributed()
meta_filename = []
for filename in self.train_dir.split(','):
logging.info("filename:%s", str(filename))
meta_filename.append(
self.get_meta_filename(start_new_model, filename))
label_loss_fn = find_class_by_name(FLAGS.label_loss, [losses])()
optimizer_class = find_class_by_name(FLAGS.optimizer, [tf.train])
local_device_protos = device_lib.list_local_devices()
gpus = [x.name for x in local_device_protos if x.device_type == 'GPU']
num_gpus = len(gpus)
if num_gpus > 0:
logging.info("Using the following GPUs to train: " + str(gpus))
num_towers = num_gpus
device_string = '/gpu:%d'
else:
logging.info("No GPUs found. Training on CPU.")
num_towers = 1
device_string = '/cpu:%d'
# build_graph_retrain(
# reader=self.reader,
# model=self.model,
# train_data_pattern=FLAGS.train_data_pattern,
# label_loss_fn=label_loss_fn,
# num_readers=FLAGS.num_readers,
# batch_size=FLAGS.batch_size)
# with tf.variable_scope("net2"):
####
global_step = tf.Variable(0, trainable=False, name="global_step")
learning_rate = tf.train.exponential_decay(
FLAGS.base_learning_rate,
global_step * FLAGS.batch_size * num_towers,
FLAGS.learning_rate_decay_examples,
FLAGS.learning_rate_decay,
staircase=True)
tf.summary.scalar('learning_rate', learning_rate)
video_id_batch, model_input_raw, labels_batch, num_frames = get_input_data_tensors(
# pylint: disable=g-line-too-long
self.reader,
FLAGS.train_data_pattern,
batch_size=FLAGS.batch_size,
num_readers=FLAGS.num_readers)
tf.summary.histogram("model_input_raw", model_input_raw)
feature_dim = len(model_input_raw.get_shape()) - 1
# Normalize input features.
model_input = tf.nn.l2_normalize(model_input_raw, feature_dim)
# with tf.variable_scope("net1"):
with tf.variable_scope("tower"):
result1 = self.model[0].create_model(
model_input,
num_frames=num_frames,
vocab_size=self.reader.num_classes,
is_training=False)
#####
result1 = tf.stop_gradient(result1)
result2 = self.model[1].create_model(
model_input,
num_frames=num_frames,
vocab_size=self.reader.num_classes,
labels=labels_batch,
is_training=False)
result2 = tf.stop_gradient(result2)
all_vars = tf.global_variables()
# for v in all_vars:
# print v.name
# for i in v_vars:
# logging.info(str(i))
for i, v in enumerate(all_vars):
logging.info(str(v.name))
if 'rnn' in v.name:
vars1 = all_vars[:i]
vars2 = all_vars[i:]
break
# v_vars0 = [v for v in all_vars if v.name == 'tower/input_bn/beta:0'
# or v.name == 'tower/input_bn/gamma:0'
# or v.name == 'tower/input_bn/beta:0'
# or v.name == 'tower/rnn/multi_rnn_cell/cell_1/basic_lstm_cell/biases:0']
# v_vars = [v for v in all_vars if v.name == 'tower/rnn/multi_rnn_cell/cell_0/basic_lstm_cell/weights:0'
# or v.name == 'tower/rnn/multi_rnn_cell/cell_0/basic_lstm_cell/biases:0'
# or v.name == 'tower/rnn/multi_rnn_cell/cell_1/basic_lstm_cell/weights:0'
# or v.name == 'tower/rnn/multi_rnn_cell/cell_1/basic_lstm_cell/biases:0']
result1 = tf.nn.l2_normalize(result1, dim=1)
result2 = tf.nn.l2_normalize(result2, dim=1)
embeddings = tf.concat([result1, result2], axis=1)
model_concat = find_class_by_name('MoeModel',
[video_level_models])()
result = model_concat.create_model(
embeddings, vocab_size=self.reader.num_classes, num_mixtures=4)
predictions = result["predictions"]
# predictions=(result1["predictions"]+result2["predictions"])/2
tf.summary.histogram("model_activations", predictions)
# if "loss" in result.keys():
# label_loss = result["loss"]
# else:
label_loss = label_loss_fn.calculate_loss(predictions,
labels_batch)
tf.summary.scalar("label_loss", label_loss)
if "regularization_loss" in result.keys():
reg_loss = result["regularization_loss"]
reg_losses = tf.losses.get_regularization_losses()
if "regularization_loss" in result.keys():
reg_loss = result["regularization_loss"]
else:
reg_loss = tf.constant(0.0)
final_loss = FLAGS.regularization_penalty * reg_loss + label_loss
optimizer = optimizer_class(learning_rate)
gradients = optimizer.compute_gradients(
final_loss, colocate_gradients_with_ops=False)
with tf.name_scope('clip_grads'):
merged_gradients = utils.clip_gradient_norms(gradients, 1.0)
train_op = optimizer.apply_gradients(merged_gradients,
global_step=global_step)
tf.add_to_collection("global_step", global_step)
tf.add_to_collection("loss", label_loss)
tf.add_to_collection("predictions", predictions)
tf.add_to_collection("input_batch", model_input)
tf.add_to_collection("video_id_batch", video_id_batch)
tf.add_to_collection("num_frames", num_frames)
tf.add_to_collection("labels", tf.cast(labels_batch, tf.float32))
tf.add_to_collection("summary_op", tf.summary.merge_all())
tf.add_to_collection("train_op", train_op)
video_id_batch = tf.get_collection("video_id_batch")[0]
prediction_batch = tf.get_collection("predictions")[0]
label_batch = tf.get_collection("labels")[0]
loss = tf.get_collection("loss")[0]
summary_op = tf.get_collection("summary_op")[0]
# saver = tf.train.Saver(tf.global_variables())
# saver=tf.train.Saver(result1)
summary_writer = tf.summary.FileWriter(
FLAGS.ensemble_dir, graph=tf.get_default_graph())
config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
train_dirs = FLAGS.train_dir.split(',')
latest_checkpoint0 = tf.train.latest_checkpoint(train_dirs[0])
latest_checkpoint1 = tf.train.latest_checkpoint(train_dirs[1])
sess.run(tf.global_variables_initializer())
if latest_checkpoint0:
logging.info("Loading checkpoint for eval: " +
latest_checkpoint0)
saver1 = tf.train.Saver(vars1)
saver1.restore(sess, latest_checkpoint0)
if latest_checkpoint1:
saver2 = tf.train.Saver(vars2)
logging.info("Loading checkpoint for eval: " +
latest_checkpoint1)
saver2.restore(sess, latest_checkpoint1)
saver = tf.train.Saver()
fetches = [
learning_rate, global_step, train_op, video_id_batch,
prediction_batch, label_batch, loss, summary_op
]
coord = tf.train.Coordinator()
threads = []
for qr in tf.get_collection(tf.GraphKeys.QUEUE_RUNNERS):
threads.extend(
qr.create_threads(sess,
coord=coord,
daemon=True,
start=True))
while not coord.should_stop():
# batch_start_time = time.time()
learning_rate_val, global_step_val, _, vid_val, predictions_val, labels_val, loss_val, summary_val = sess.run(
fetches)
# hit_at_one = eval_util.calculate_hit_at_one(predictions_val, labels_val)
# perr = eval_util.calculate_precision_at_equal_recall_rate(predictions_val,
# labels_val)
# gap = eval_util.calculate_gap(predictions_val, labels_val)
# logging.info( "training step " + str(global_step_val)+" | Loss: " + ("%.2f" % loss_val) +" | Hit@1: " +
# ("%.4f" % hit_at_one) + " PERR: " + ("%.4f" % perr) +
# " GAP: " + ("%.4f" % gap))
if self.is_master and global_step_val % self.disp_batches == 0 and self.train_dir:
eval_start_time = time.time()
hit_at_one = eval_util.calculate_hit_at_one(
predictions_val, labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(
predictions_val, labels_val)
gap = eval_util.calculate_gap(predictions_val,
labels_val)
eval_end_time = time.time()
eval_time = eval_end_time - eval_start_time
logging.info("training step " + str(global_step_val) +
"| learning rate: " +
("%.4f" % learning_rate_val) +
" | Loss: " + ("%.2f" % loss_val) +
" | Hit@1: " + ("%.4f" % hit_at_one) +
" PERR: " + ("%.4f" % perr) + " GAP: " +
("%.4f" % gap))
summary_writer.add_summary(
utils.MakeSummary("model/Training_Hit@1",
hit_at_one), global_step_val)
summary_writer.add_summary(
utils.MakeSummary("model/Training_Perr", perr),
global_step_val)
summary_writer.add_summary(
utils.MakeSummary("model/Training_GAP", gap),
global_step_val)
summary_writer.add_summary(
utils.MakeSummary("model/loss", loss_val),
global_step_val)
summary_writer.add_summary(
utils.MakeSummary("model/lr", learning_rate_val),
global_step_val)
summary_writer.flush()
if global_step_val % FLAGS.export_model_steps == 0:
saver.save(sess,
FLAGS.ensemble_dir,
global_step=global_step_val)
coord.request_stop()
coord.join(threads, stop_grace_period_secs=10)
def run(self, start_new_model=False):
"""Performs training on the currently defined Tensorflow graph.
Returns:
A tuple of the training Hit@1 and the training PERR.
"""
if self.is_master and start_new_model:
self.remove_training_directory(self.train_dir)
target, device_fn = self.start_server_if_distributed()
meta_filename = self.get_meta_filename(start_new_model, self.train_dir)
with tf.Graph().as_default() as graph:
if meta_filename:
saver = self.recover_model(meta_filename)
with tf.device(device_fn):
if not meta_filename:
saver = self.build_model()
global_step = tf.get_collection("global_step")[0]
loss = tf.get_collection("loss")[0]
predictions = tf.get_collection("predictions")[0]
labels = tf.get_collection("labels")[0]
train_op = tf.get_collection("train_op")[0]
init_op = tf.global_variables_initializer()
if FLAGS.dropout:
keep_prob_tensor = tf.get_collection("keep_prob")[0]
if FLAGS.noise_level > 0:
noise_level_tensor = tf.get_collection("noise_level")[0]
if FLAGS.reweight:
weights_input, weights_assignment = None, None
if len(tf.get_collection("weights_input")) > 0:
weights_input = tf.get_collection("weights_input")[0]
weights_assignment = tf.get_collection(
"weights_assignment")[0]
sv = tf.train.Supervisor(graph,
logdir=self.train_dir,
init_op=init_op,
is_chief=self.is_master,
global_step=global_step,
save_model_secs=6 * 60,
save_summaries_secs=120,
saver=saver)
logging.info("%s: Starting managed session.",
task_as_string(self.task))
with sv.managed_session(target, config=self.config) as sess:
# re-assign weights
if FLAGS.reweight:
optional_assign_weights(sess, weights_input,
weights_assignment)
try:
logging.info("%s: Entering training loop.",
task_as_string(self.task))
while not sv.should_stop():
batch_start_time = time.time()
custom_feed = {}
if FLAGS.dropout:
custom_feed[keep_prob_tensor] = FLAGS.keep_prob
if FLAGS.noise_level > 0:
custom_feed[noise_level_tensor] = FLAGS.noise_level
_, global_step_val, loss_val, predictions_val, labels_val = sess.run(
[train_op, global_step, loss, predictions, labels],
feed_dict=custom_feed)
seconds_per_batch = time.time() - batch_start_time
if self.is_master:
examples_per_second = labels_val.shape[
0] / seconds_per_batch
hit_at_one = eval_util.calculate_hit_at_one(
predictions_val, labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(
predictions_val, labels_val)
recall = "N/A"
if False:
recall = eval_util.calculate_recall_at_n(
predictions_val, labels_val, FLAGS.recall_at_n)
sv.summary_writer.add_summary(
utils.MakeSummary(
"model/Training_Recall@%d" %
FLAGS.recall_at_n, recall),
global_step_val)
recall = "%.2f" % recall
gap = eval_util.calculate_gap(predictions_val,
labels_val)
logging.info(
"%s: training step " + str(global_step_val) +
"| Hit@1: " + ("%.2f" % hit_at_one) + " PERR: " +
("%.2f" % perr) + " GAP: " + ("%.2f" % gap) +
" Recall@%d: " % FLAGS.recall_at_n +
recall + " Loss: " + str(loss_val),
task_as_string(self.task))
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Hit@1",
hit_at_one), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Perr", perr),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_GAP", gap),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("global_step/Examples/Second",
examples_per_second),
global_step_val)
sv.summary_writer.flush()
except tf.errors.OutOfRangeError:
logging.info("%s: Done training -- epoch limit reached.",
task_as_string(self.task))
logging.info("%s: Exited training loop.", task_as_string(self.task))
sv.Stop()
def run(self, start_new_model=False):
"""Performs training on the currently defined Tensorflow graph.
Returns:
A tuple of the training Hit@1 and the training PERR.
"""
if self.is_master and start_new_model:
self.remove_training_directory(self.train_dir)
target, device_fn = self.start_server_if_distributed()
meta_filename = self.get_meta_filename(start_new_model, self.train_dir)
with tf.Graph().as_default() as graph:
if meta_filename:
saver = self.recover_model(meta_filename)
with tf.device(device_fn):
if not meta_filename:
saver = self.build_model(self.model, self.reader)
global_step = tf.get_collection("global_step")[0]
loss = tf.get_collection("loss")[0]
predictions = tf.get_collection("predictions")[0]
labels = tf.get_collection("labels")[0]
train_op = tf.get_collection("train_op")[0]
init_op = tf.global_variables_initializer()
sv = tf.train.Supervisor(
graph,
logdir=self.train_dir,
init_op=init_op,
is_chief=self.is_master,
global_step=global_step,
save_model_secs=15 * 60,
save_summaries_secs=120,
saver=saver)
logging.info("%s: Starting managed session.", task_as_string(self.task))
with sv.managed_session(target, config=self.config) as sess:
try:
logging.info("%s: Entering training loop.", task_as_string(self.task))
while (not sv.should_stop()) and (not self.max_steps_reached):
batch_start_time = time.time()
_, global_step_val, loss_val, predictions_val, labels_val = sess.run(
[train_op, global_step, loss, predictions, labels])
seconds_per_batch = time.time() - batch_start_time
if self.max_steps and self.max_steps <= global_step_val:
self.max_steps_reached = True
if self.is_master:
examples_per_second = labels_val.shape[0] / seconds_per_batch
hit_at_one = eval_util.calculate_hit_at_one(predictions_val,
labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(
predictions_val, labels_val)
gap = eval_util.calculate_gap(predictions_val, labels_val)
logging.info(
"%s: training step " + str(global_step_val) + "| Hit@1: " +
("%.2f" % hit_at_one) + " PERR: " + ("%.2f" % perr) + " GAP: " +
("%.2f" % gap) + " Loss: " + str(loss_val),
task_as_string(self.task))
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Hit@1", hit_at_one),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Perr", perr), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_GAP", gap), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("global_step/Examples/Second",
examples_per_second), global_step_val)
sv.summary_writer.flush()
# Exporting the model every x steps
time_to_export = ((self.last_model_export_step == 0) or
(global_step_val - self.last_model_export_step
>= self.export_model_steps))
if self.is_master and time_to_export:
self.export_model(global_step_val, sv.saver, sv.save_path, sess)
self.last_model_export_step = global_step_val
# Exporting the final model
if self.is_master:
self.export_model(global_step_val, sv.saver, sv.save_path, sess)
except tf.errors.OutOfRangeError:
logging.info("%s: Done training -- epoch limit reached.",
task_as_string(self.task))
logging.info("%s: Exited training loop.", task_as_string(self.task))
sv.Stop()
x_length_valid = np.ones(dtype=np.int32,
shape=(frame_fea_valid.shape[0])) * 45
audio_fea_valid = audio_feat_process(audio_fea_valid)
title_fea_valid = text_feat_propress(title_fea_valid)
desc_fea_valid = text_feat_propress(desc_fea_valid)
tags_valid = tags_process(tags_valid)
batch_size_in_valid = frame_fea_valid.shape[0]
feed_dict = {
frame_fea_placeholder: frame_fea_valid,
x_length_placeholder: x_length_valid,
audio_fea_placeholder: audio_fea_valid,
tags_placeholder: tags_valid,
batch_size_in_placeholder: batch_size_in_valid,
title_placeholder: title_fea_valid,
desc_placeholder: desc_fea_valid,
image_fea_placeholder: image_fea_valid
}
predict_result, loss_valid = sess.run(
[valid_nets['confidence'], valid_nets['loss']],
feed_dict=feed_dict)
if i == 0:
print(predict_result[0, :])
score = eval_util.calculate_gap(predict_result, tags_valid)
score_average += score
loss_average += loss_valid
print('step:{}, score_average:{:.4f}, loss_average:{:.4f}'.format(
step, score_average / 10, loss_average / 10))
def run(self, start_new_model=False):
"""Performs training on the currently defined Tensorflow graph.
Returns:
A tuple of the training Hit@1 and the training PERR.
"""
if self.is_master and start_new_model:
self.remove_training_directory(self.train_dir)
target, device_fn = self.start_server_if_distributed()
latest_checkpoint, meta_filename = self.get_meta_filename(
start_new_model, self.train_dir)
with tf.Graph().as_default() as graph:
if meta_filename:
if not FLAGS.change_file:
saver = self.recover_model(meta_filename)
with tf.device(device_fn):
if not meta_filename:
saver = self.build_model(self.model, self.reader)
if FLAGS.change_file:
saver = self.build_model(self.model, self.reader)
global_step = tf.get_collection("global_step")[0]
loss = tf.get_collection("loss")[0]
predictions = tf.get_collection("predictions")[0]
labels = tf.get_collection("labels")[0]
train_op = tf.get_collection("train_op")[0]
init_op = tf.global_variables_initializer()
sv = tf.train.Supervisor(graph,
logdir=self.train_dir,
init_op=init_op,
is_chief=self.is_master,
global_step=global_step,
save_model_secs=60 * FLAGS.time_to_save_model,
save_summaries_secs=120,
saver=saver)
logging.info("%s: Starting managed session.",
task_as_string(self.task))
with sv.managed_session(target, config=self.config) as sess:
try:
logging.info("%s: Entering training loop.",
task_as_string(self.task))
logging.info("TANG:restoring")
saver.restore(sess, latest_checkpoint)
while (not sv.should_stop()) and (not self.max_steps_reached):
batch_start_time = time.time()
_, global_step_val, loss_val, predictions_val, labels_val = sess.run(
[train_op, global_step, loss, predictions, labels])
seconds_per_batch = time.time() - batch_start_time
examples_per_second = labels_val.shape[
0] / seconds_per_batch
if self.max_steps and self.max_steps <= global_step_val:
self.max_steps_reached = True
if self.is_master and global_step_val % FLAGS.eval_loop == 0 and self.train_dir:
eval_start_time = time.time()
hit_at_one = eval_util.calculate_hit_at_one(
predictions_val, labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(
predictions_val, labels_val)
gap = eval_util.calculate_gap(predictions_val,
labels_val)
eval_end_time = time.time()
eval_time = eval_end_time - eval_start_time
logging.info("training step " + str(global_step_val) +
" | Loss: " + ("%.2f" % loss_val) +
" Examples/sec: " +
("%.2f" % examples_per_second) +
" | Hit@1: " + ("%.2f" % hit_at_one) +
" PERR: " + ("%.2f" % perr) + " GAP: " +
("%.2f" % gap))
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Hit@1",
hit_at_one), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Perr", perr),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_GAP", gap),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("global_step/Examples/Second",
examples_per_second),
global_step_val)
sv.summary_writer.flush()
# Exporting the model every x steps
time_to_export = (
(self.last_model_export_step == 0)
or (global_step_val - self.last_model_export_step
>= self.export_model_steps))
if self.is_master and time_to_export:
self.export_model(global_step_val, sv.saver,
sv.save_path, sess)
self.last_model_export_step = global_step_val
else:
logging.info("training step " + str(global_step_val) +
" | Loss: " + ("%.2f" % loss_val) +
" Examples/sec: " +
("%.2f" % examples_per_second))
except tf.errors.OutOfRangeError:
logging.info("%s: Done training -- epoch limit reached.",
task_as_string(self.task))
logging.info("%s: Exited training loop.", task_as_string(self.task))
sv.Stop()
def run(self, start_new_model=False):
"""Performs training on the currently defined Tensorflow graph.
Returns:
A tuple of the training Hit@1 and the training PERR.
"""
if self.is_master and start_new_model:
self.remove_training_directory(self.train_dir)
if not os.path.exists(self.train_dir):
os.makedirs(self.train_dir)
model_flags_dict = {
"model": FLAGS.model,
"feature_sizes": FLAGS.feature_sizes,
"feature_names": FLAGS.feature_names,
"frame_features": FLAGS.frame_features,
"label_loss": FLAGS.label_loss,
"netvlad_cluster_size": FLAGS.netvlad_cluster_size,
"netvlad_hidden_size": FLAGS.netvlad_hidden_size,
"moe_l2": FLAGS.moe_l2,
"iterations": FLAGS.iterations,
"netvlad_relu": FLAGS.netvlad_relu,
"gating": FLAGS.gating,
"moe_num_mixtures": FLAGS.moe_num_mixtures,
"moe_prob_gating": FLAGS.moe_prob_gating,
}
# model_flags_dict = {}
# for k in FLAGS.__flags:
# model_flags_dict[k] = FLAGS[k].value
flags_json_path = os.path.join(FLAGS.train_dir, "model_flags.json")
if os.path.exists(flags_json_path):
existing_flags = json.load(open(flags_json_path))
if existing_flags != model_flags_dict:
logging.error(
"Model flags do not match existing file %s. Please "
"delete the file, change --train_dir, or pass flag "
"--start_new_model", flags_json_path)
logging.error("Ran model with flags: %s",
str(model_flags_dict))
logging.error("Previously ran with flags: %s",
str(existing_flags))
exit(1)
else:
# Write the file.
with open(flags_json_path, "w") as fout:
fout.write(json.dumps(model_flags_dict))
target, device_fn = self.start_server_if_distributed()
meta_filename = self.get_meta_filename(start_new_model, self.train_dir)
with tf.Graph().as_default() as graph:
if meta_filename:
saver = self.recover_model(meta_filename)
with tf.device(device_fn):
if not meta_filename:
saver = self.build_model(self.model, self.reader)
global_step = tf.get_collection("global_step")[0]
loss = tf.get_collection("loss")[0]
predictions = tf.get_collection("predictions")[0]
labels = tf.get_collection("labels")[0]
train_op = tf.get_collection("train_op")[0]
init_op = tf.global_variables_initializer()
sv = tf.train.Supervisor(graph,
logdir=self.train_dir,
init_op=init_op,
is_chief=self.is_master,
global_step=global_step,
save_model_secs=15 * 60,
save_summaries_secs=120,
saver=saver)
logging.info("%s: Starting managed session.",
task_as_string(self.task))
with sv.managed_session(target, config=self.config) as sess:
try:
logging.info("%s: Entering training loop.",
task_as_string(self.task))
while (not sv.should_stop()) and (not self.max_steps_reached):
batch_start_time = time.time()
_, global_step_val, loss_val, predictions_val, labels_val = sess.run(
[train_op, global_step, loss, predictions, labels])
seconds_per_batch = time.time() - batch_start_time
examples_per_second = labels_val.shape[
0] / seconds_per_batch
if self.max_steps and self.max_steps <= global_step_val:
self.max_steps_reached = True
if self.is_master and global_step_val % 10 == 0 and self.train_dir:
eval_start_time = time.time()
hit_at_one = eval_util.calculate_hit_at_one(
predictions_val, labels_val)
perr = eval_util.calculate_precision_at_equal_recall_rate(
predictions_val, labels_val)
gap = eval_util.calculate_gap(predictions_val,
labels_val)
eval_end_time = time.time()
eval_time = eval_end_time - eval_start_time
logging.info("training step " + str(global_step_val) +
" | Loss: " + ("%.2f" % loss_val) +
" Examples/sec: " +
("%.2f" % examples_per_second) +
" | Hit@1: " + ("%.2f" % hit_at_one) +
" PERR: " + ("%.2f" % perr) + " GAP: " +
("%.2f" % gap))
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Hit@1",
hit_at_one), global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_Perr", perr),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("model/Training_GAP", gap),
global_step_val)
sv.summary_writer.add_summary(
utils.MakeSummary("global_step/Examples/Second",
examples_per_second),
global_step_val)
sv.summary_writer.flush()
# Exporting the model every x steps
time_to_export = (
(self.last_model_export_step == 0)
or (global_step_val - self.last_model_export_step
>= self.export_model_steps))
if self.is_master and time_to_export:
self.export_model(global_step_val, sv.saver,
sv.save_path, sess)
self.last_model_export_step = global_step_val
else:
logging.info("training step " + str(global_step_val) +
" | Loss: " + ("%.2f" % loss_val) +
" Examples/sec: " +
("%.2f" % examples_per_second))
except tf.errors.OutOfRangeError:
logging.info("%s: Done training -- epoch limit reached.",
task_as_string(self.task))
logging.info("%s: Exited training loop.", task_as_string(self.task))
sv.Stop()
def evaluation_loop(video_id_batch, prediction_batch, label_batch, loss,
summary_op, saver, summary_writer, evl_metrics,
last_global_step_val):
"""Run the evaluation loop once.
Args:
video_id_batch: a tensor of video ids mini-batch.
prediction_batch: a tensor of predictions mini-batch.
label_batch: a tensor of label_batch mini-batch.
loss: a tensor of loss for the examples in the mini-batch.
summary_op: a tensor which runs the tensorboard summary operations.
saver: a tensorflow saver to restore the model.
summary_writer: a tensorflow summary_writer
evl_metrics: an EvaluationMetrics object.
last_global_step_val: the global step used in the previous evaluation.
Returns:
The global_step used in the latest model.
"""
global_step_val = -1
with tf.Session(config=config) as sess:
if FLAGS.check_point < 0:
latest_checkpoint = get_latest_checkpoint()
else:
latest_checkpoint = FLAGS.train_dir + "model.ckpt-" + str(FLAGS.check_point)
if latest_checkpoint:
print(f'loading: {latest_checkpoint}')
logging.info("Loading checkpoint for eval: " + latest_checkpoint)
# Restores from checkpoint
saver.restore(sess, latest_checkpoint)
# Assuming model_checkpoint_path looks something like:
# /my-favorite-path/yt8m_train/model.ckpt-0, extract global_step from it.
global_step_val = os.path.basename(latest_checkpoint).split("-")[-1]
# Save model
saver.save(sess, os.path.join(FLAGS.train_dir, "inference_model"))
else:
logging.info("No checkpoint file found.")
return global_step_val
if global_step_val == last_global_step_val:
logging.info("skip this checkpoint global_step_val=%s "
"(same as the previous one).", global_step_val)
return global_step_val
sess.run([tf.local_variables_initializer()])
# Start the queue runners.
fetches = [video_id_batch, prediction_batch, label_batch, loss, summary_op]
coord = tf.train.Coordinator()
try:
threads = []
for qr in tf.get_collection(tf.GraphKeys.QUEUE_RUNNERS):
threads.extend(qr.create_threads(
sess, coord=coord, daemon=True,
start=True))
logging.info("enter eval_once loop global_step_val = %s. ",
global_step_val)
evl_metrics.clear()
examples_processed = 0
while not coord.should_stop():
batch_start_time = time.time()
_, predictions_val, labels_val, loss_val, summary_val = sess.run(
fetches)
seconds_per_batch = time.time() - batch_start_time
example_per_second = labels_val.shape[0] / seconds_per_batch
examples_processed += labels_val.shape[0]
iteration_info_dict = evl_metrics.accumulate(predictions_val,
labels_val, loss_val)
iteration_info_dict["examples_per_second"] = example_per_second
# iterinfo = utils.AddGlobalStepSummary(
# summary_writer,
# global_step_val,
# iteration_info_dict,
# summary_scope="Eval")
gap = eval_util.calculate_gap(predictions_val, labels_val)
logging.info("examples_processed: %d | %s", examples_processed,
iterinfo)
logging.info(f"GAP: {gap}",)
except tf.errors.OutOfRangeError as e:
logging.info(
"Done with batched inference. Now calculating global performance "
"metrics.")
# calculate the metrics for the entire epoch
epoch_info_dict = evl_metrics.get()
epoch_info_dict["epoch_id"] = global_step_val
# summary_writer.add_summary(summary_val, global_step_val)
# epochinfo = utils.AddEpochSummary(
# summary_writer,
# global_step_val,
# epoch_info_dict,
# summary_scope="Eval")
# logging.info(epochinfo)
evl_metrics.clear()
except Exception as e: # pylint: disable=broad-except
logging.info("Unexpected exception: " + str(e))
coord.request_stop(e)
coord.request_stop()
coord.join(threads, stop_grace_period_secs=10)
return global_step_val