def run(dat, ac_config):
with tf.Graph().as_default() as g:
m = SCModel(ac_config)
s = tf.train.Saver(tf.global_variables())
# print("AC config hypnodensity path",ac_config.hypnodensity_model_dir)
with tf.Session(config=tf.ConfigProto(log_device_placement=False)) as session:
ckpt = tf.train.get_checkpoint_state(ac_config.hypnodensity_model_dir)
# For debugging
# pdb.set_trace()
s.restore(session, ckpt.model_checkpoint_path)
state = np.zeros([1, ac_config.num_hidden * 2])
dat, Nextra, prediction, num_batches = Hypnodensity.segment(dat, ac_config)
for i in range(num_batches):
x = dat[:, i * ac_config.eval_nseg_atonce * ac_config.segsize:(
i + 1) * ac_config.eval_nseg_atonce * ac_config.segsize,
:]
est, _ = session.run([m.logits, m.final_state], feed_dict={
m.features: x,
m.targets: np.ones([ac_config.eval_nseg_atonce * ac_config.segsize, 5]),
m.mask: np.ones(ac_config.eval_nseg_atonce * ac_config.segsize),
m.batch_size: np.ones([1]),
m.initial_state: state
})
prediction[i * ac_config.eval_nseg_atonce:(i + 1) * ac_config.eval_nseg_atonce, :] = est
prediction = prediction[:-int(Nextra / ac_config.segsize), :]
return prediction
def run(dat, ac_config):
++Hypnodensity.run_count
once = min(ac_config.eval_nseg_atonce, int((dat.shape[1]/ac_config.segsize)-1))
thread_count = max(multiprocessing.cpu_count() - 1, 1)
os.environ["MKL_NUM_THREADS"] = str(thread_count)
with tf.Graph().as_default() as g:
m = SCModel(ac_config)
s = tf.train.Saver(tf.global_variables())
print("AC config hypnodensity path",ac_config.hypnodensity_model_dir)
builder = tf.profiler.ProfileOptionBuilder
opts = builder(builder.time_and_memory()).order_by('micros').build()
profile_dir = gettempdir() + 'mlstages_profile_' + str(Hypnodensity.run_count)
with tf.contrib.tfprof.ProfileContext(profile_dir) as pctx:
config = tf.ConfigProto(log_device_placement=False,
intra_op_parallelism_threads=thread_count,
inter_op_parallelism_threads=thread_count)
with tf.Session(config=config) as session:
pctx.trace_next_step()
pctx.dump_next_step()
ckpt = tf.train.get_checkpoint_state(ac_config.hypnodensity_model_dir)
pctx.profiler.profile_operations(options=opts)
s.restore(session, ckpt.model_checkpoint_path)
state = np.zeros([1, ac_config.num_hidden * 2])
dat, Nextra, prediction, num_batches = Hypnodensity.segment(dat, ac_config)
for i in range(num_batches):
x = dat[:, i * once * ac_config.segsize:(i + 1) * once * ac_config.segsize,:]
pctx.trace_next_step()
pctx.dump_next_step()
est, _ = session.run([m.logits, m.final_state], feed_dict={
m.features: x,
m.targets: np.ones([once * ac_config.segsize, 5]),
m.mask: np.ones(once * ac_config.segsize),
m.batch_size: np.ones([1]),
m.initial_state: state
},
options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE))
pctx.profiler.profile_operations(options=opts)
prediction[i * once:(i + 1) * once, :] = est
prediction = prediction[:-int(Nextra / ac_config.segsize), :]
return prediction
def run(dat, ac_config):
with tf.compat.v1.Graph().as_default() as g:
m = SCModel(ac_config)
s = tf.compat.v1.train.Saver(tf.compat.v1.global_variables())
print("AC config hypnodensity path",
ac_config.hypnodensity_model_dir)
# config = tf.compat.v1.ConfigProto(log_device_placement=False, device_count={'GPU': 0}) # For cpu only operations
config = tf.compat.v1.ConfigProto(log_device_placement=False)
# config = tf.ConfigProto()
# config = tf.ConfigProto(log_device_placement=True) # Setting log_device_placement=True gives way too much output.
# config.gpu_options.allow_growth = True # See also: https://www.tensorflow.org/guide/using_gpu
# config.gpu_options.per_process_gpu_memory_fraction = 1.0
with tf.compat.v1.Session(config=config) as session:
ckpt = tf.compat.v1.train.get_checkpoint_state(
ac_config.hypnodensity_model_dir)
s.restore(session, ckpt.model_checkpoint_path)
state = np.zeros([1, ac_config.num_hidden * 2])
# state = m.initial_state
dat, Nextra, prediction, num_batches = Hypnodensity.segment(
dat, ac_config)
for i in range(num_batches):
x = dat[:, i * ac_config.eval_nseg_atonce *
ac_config.segsize:(i + 1) *
ac_config.eval_nseg_atonce * ac_config.segsize, :]
est, state = session.run(
[m.logits, m.final_state],
feed_dict={
m.features:
x,
m.targets:
np.ones([
ac_config.eval_nseg_atonce * ac_config.segsize,
5
]),
m.mask:
np.ones(ac_config.eval_nseg_atonce *
ac_config.segsize),
m.batch_size:
np.ones([1]),
m.initial_state:
state
})
prediction[i * ac_config.eval_nseg_atonce:(i + 1) *
ac_config.eval_nseg_atonce, :] = est
prediction = prediction[:-int(Nextra / ac_config.segsize), :]
return prediction