def __init__(self, sess, network_params):
self._sess = sess
self._params = network_params
self._device = self._params['device']
self._tf_sumry_wrtr = None
if network_params['write_summary']:
if 'summary_dir' in network_params:
summary_dir = network_params['summary_dir']
else:
summary_dir = None
self._tf_sumry_wrtr = TfSummaryWriter(tf_session=sess,
summary_dir=summary_dir)
with tf.device(self._device):
if self._params['write_summary']:
tf.global_variables_initializer().run()
self._mdn = MixtureDensityNetwork(
network_params, tf_sumry_wrtr=self._tf_sumry_wrtr)
self._mdn._init_model()
self._net_ops = self._mdn._ops
self._init_op = tf.global_variables_initializer()
self._saver = tf.train.Saver()
def __init__(self, sess, network_params):
self._sess = sess
self._params = network_params
self._device = self._params['device']
self._tf_sumry_wrtr = None
self._optimiser = network_params['optimiser']
self._data_configured = False
if network_params['write_summary']:
if 'summary_dir' in network_params:
summary_dir = network_params['summary_dir']
else:
summary_dir = None
self._tf_sumry_wrtr = TfSummaryWriter(tf_session=sess,summary_dir=summary_dir)
cuda_path = '/usr/local/cuda/extras/CUPTI/lib64'
curr_ld_path = os.environ["LD_LIBRARY_PATH"]
if not cuda_path in curr_ld_path.split(os.pathsep):
print "Enviroment variable LD_LIBRARY_PATH does not contain %s"%cuda_path
print "Please add it, else the program will crash!"
raw_input("Press Ctrl+C")
# os.environ["LD_LIBRARY_PATH"] = curr_ld_path + ':'+cuda_path
with tf.device(self._device):
self._net_ops = tf_model(dim_input=network_params['dim_input'],
dim_output=network_params['dim_output'],
loss_type='quadratic',
cnn_params=network_params['cnn_params'],
fc_params=network_params['fc_params'],
optimiser_params=network_params['optimiser'],
tf_sumry_wrtr=self._tf_sumry_wrtr)
self._init_op = tf.initialize_all_variables()
self._saver = tf.train.Saver()
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import argparse
import sys
import tensorflow as tf
from aml_dl.utilities.tf_summary_writer import TfSummaryWriter
from tensorflow.examples.tutorials.mnist import input_data
FLAGS = None
sess = tf.InteractiveSession()
tf_sum_wrtr = TfSummaryWriter(tf_session=sess)
def train():
# Import data
mnist = input_data.read_data_sets(FLAGS.data_dir,
one_hot=True,
fake_data=FLAGS.fake_data)
# Create a multilayer model.
# Input placeholders
with tf.name_scope('input'):
x = tf.placeholder(tf.float32, [None, 784], name='x-input')
y_ = tf.placeholder(tf.float32, [None, 10], name='y-input')
with tf.name_scope('input_reshape'):
class MDN(object):
def __init__(self, sess, network_params):
self._sess = sess
self._params = network_params
self._device = self._params['device']
self._tf_sumry_wrtr = None
if network_params['write_summary']:
if 'summary_dir' in network_params:
summary_dir = network_params['summary_dir']
else:
summary_dir = None
self._tf_sumry_wrtr = TfSummaryWriter(tf_session=sess,
summary_dir=summary_dir)
with tf.device(self._device):
if self._params['write_summary']:
tf.global_variables_initializer().run()
self._mdn = MixtureDensityNetwork(
network_params, tf_sumry_wrtr=self._tf_sumry_wrtr)
self._mdn._init_model()
self._net_ops = self._mdn._ops
self._init_op = tf.global_variables_initializer()
self._saver = tf.train.Saver()
def init_model(self):
with tf.device(self._device):
self._sess.run(self._init_op)
if self._params['load_saved_model']:
self.load_model()
# def load_model(self):
# load_tf_check_point(session=self._sess, filename=self.get_model_path())
def save_model(self, path):
save_path = self._saver.save(self._sess, path)
print("Model saved in file: %s" % save_path)
def train(self, x_data, y_data, iterations=10000):
with tf.device(self._device):
# Keeping track of loss progress as we train
loss = np.zeros(iterations)
train_op = self._net_ops['train']
loss_op = self._net_ops['loss']
for i in range(iterations):
if self._params['write_summary']:
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
summary, loss[i] = self._sess.run(
fetches=[self._tf_sumry_wrtr._merged, train_op],
feed_dict={
self._net_ops['x']: x_data,
self._net_ops['y']: y_data
},
options=run_options,
run_metadata=run_metadata)
self._tf_sumry_wrtr.add_run_metadata(metadata=run_metadata,
itr=i)
self._tf_sumry_wrtr.add_summary(summary=summary, itr=i)
else:
_, loss[i] = self._sess.run([train_op, loss_op],
feed_dict={
self._net_ops['x']: x_data,
self._net_ops['y']: y_data
})
if self._tf_sumry_wrtr is not None:
self._tf_sumry_wrtr.close_writer()
return loss
# TODO: Fix
def sample_out(self, x_input, m_samples=10):
with tf.device(self._device):
out_pi, out_mu, out_sigma = self._sess.run(
[
self._net_ops['pi'], self._net_ops['mu'],
self._net_ops['sigma']
],
feed_dict={self._net_ops['x']: x_input})
samples = self._generate_mixture_samples(out_pi, out_mu, out_sigma,
m_samples)
return samples
# TODO: Fix
def sample_out_max_pi(self, x_input, m_samples=10):
with tf.device(self._device):
out_pi, out_mu, out_sigma = self._sess.run(
[
self._net_ops['pi'], self._net_ops['mu'],
self._net_ops['sigma']
],
feed_dict={self._net_ops['x']: x_input})
samples = self._generate_mixture_samples_from_max_pi(
out_pi, out_mu, out_sigma, m_samples)
return samples
# TODO: Fix
def expected_out(self, x_input, m_samples=10):
with tf.device(self._device):
samples = self.sample_out(x_input, m_samples)[0]
return np.mean(samples)
# TODO: Fix
def expected_out2(self, x_input, m_samples=10):
with tf.device(self._device):
samples = self.sample_out(x_input, m_samples)[0]
out = np.zeros(2)
for i in range(m_samples):
out += np.array([np.cos(samples[i]), np.sin(samples[i])])
out /= m_samples
out /= np.linalg.norm(out)
return out
# TODO: Fix
def expected_max_pi_out(self, x_input, m_samples=10):
with tf.device(self._device):
samples = self.sample_out_max_pi(x_input, m_samples)[0]
return np.mean(samples)
# TODO: Fix
def expected_max_pi_out2(self, x_input, m_samples=10):
with tf.device(self._device):
samples = self.sample_out_max_pi(x_input, m_samples)[0]
out = np.zeros(2)
for i in range(m_samples):
out += np.array([np.cos(samples[i]), np.sin(samples[i])])
out /= m_samples
out /= np.linalg.norm(out)
return out
def run_op(self, op_name, x_input):
with tf.device(self._device):
op = self._net_ops[op_name]
out = self._sess.run(op, feed_dict={self._net_ops['x']: x_input})
return out
def forward(self, xs):
out = self._mdn.forward(self._sess, xs)
return out
def predict(self, xs):
out = self._mdn.predict(self._sess, xs)
return out
def _sample_pi_idx(self, x, pdf):
N = pdf.size
acc = 0
for i in range(0, N):
acc += pdf[i]
if (acc >= x):
return i
print 'failed to sample mixture weight index'
return -1
def _max_pi_idx(self, pdf):
i = np.argmax(pdf, axis=1)
return i
# TODO: Fix
def _sample_gaussian(self, rn, mu, std):
return mu + rn * std
# TODO: Fix
def _generate_mixture_samples_from_max_pi(self,
out_pi,
out_mu,
out_sigma,
m_samples=10):
# Number of test inputs
N = out_mu.shape[0]
D = out_mu.shape[1] # mean dimension
M = m_samples
result = np.random.rand(N, D, M) # initially random [0, 1]
rn = np.random.randn(N, D, M) # normal random matrix (0.0, 1.0)
# Generates M samples from the mixture for each test input
for j in range(M):
for i in range(0, N):
idx = self._max_pi_idx(out_pi[i])
mu = out_mu[i, idx]
std = out_sigma[i, idx]
result[i, :, j] = self._sample_gaussian(rn[i, :, j], mu, std)
return result
# TODO: Fix
def _generate_mixture_samples(self,
out_pi,
out_mu,
out_sigma,
m_samples=10):
# Number of test inputs
N = out_mu.shape[0]
D = out_mu.shape[1] # mean dimension
M = m_samples
result = np.random.rand(N, D, M) # initially random [0, 1]
rn = np.random.randn(N, D, M) # normal random matrix (0.0, 1.0)
# Generates M samples from the mixture for each test input
for j in range(M):
for i in range(0, N):
idx = self._sample_pi_idx(np.random.rand(1)[0], out_pi[i])
mu = out_mu[i, :, idx]
std = out_sigma[i, idx]
result[i, :, j] = self._sample_gaussian(rn[i, :, j], mu, std)
return result
class NNPushFwdModel(object):
def __init__(self, sess, network_params):
self._sess = sess
self._params = network_params
self._device = self._params['device']
self._tf_sumry_wrtr = None
self._optimiser = network_params['optimiser']
self._data_configured = False
if network_params['write_summary']:
if 'summary_dir' in network_params:
summary_dir = network_params['summary_dir']
else:
summary_dir = None
self._tf_sumry_wrtr = TfSummaryWriter(tf_session=sess,summary_dir=summary_dir)
cuda_path = '/usr/local/cuda/extras/CUPTI/lib64'
curr_ld_path = os.environ["LD_LIBRARY_PATH"]
if not cuda_path in curr_ld_path.split(os.pathsep):
print "Enviroment variable LD_LIBRARY_PATH does not contain %s"%cuda_path
print "Please add it, else the program will crash!"
raw_input("Press Ctrl+C")
# os.environ["LD_LIBRARY_PATH"] = curr_ld_path + ':'+cuda_path
with tf.device(self._device):
self._net_ops = tf_model(dim_input=network_params['dim_input'],
dim_output=network_params['dim_output'],
loss_type='quadratic',
cnn_params=network_params['cnn_params'],
fc_params=network_params['fc_params'],
optimiser_params=network_params['optimiser'],
tf_sumry_wrtr=self._tf_sumry_wrtr)
self._init_op = tf.initialize_all_variables()
self._saver = tf.train.Saver()
def init_model(self):
with tf.device(self._device):
self._sess.run(self._init_op)
if self._params['load_saved_model']:
self.load_model()
def configure_data(self, data_x, data_y, batch_creator):
self._data_x = data_x
self._data_y = data_y
self._batch_creator = batch_creator
self._data_configured = True
def get_model_path(self):
if 'model_dir' in self._params:
model_dir = self._params['model_dir']
else:
model_path = './fwd/'
if not os.path.exists(model_dir):
os.makedirs(model_dir)
if 'model_name' in self._params:
model_name = self._params['model_name']
else:
model_name = 'fwd_model.ckpt'
return model_dir+model_name
def load_model(self):
load_tf_check_point(session=self._sess, filename=self.get_model_path())
def save_model(self):
save_path = self._saver.save(self._sess, self.get_model_path())
print("Model saved in file: %s" % save_path)
def get_data(self):
round_complete = None
if self._params['batch_params'] is not None:
if self._batch_creator is not None:
self._data_x, self._data_y, round_complete = self._batch_creator.get_batch(random_samples=self._params['batch_params']['use_random_batches'])
else:
raise Exception("Batch training chosen but batch_creator not configured")
if self._params['cnn_params'] is None:
feed_dict = {self._net_ops['x']:self._data_x, self._net_ops['y']:self._data_y}
else:
feed_dict = {self._net_ops['image_input']:self._data_x, self._net_ops['y']:self._data_y}
return feed_dict, round_complete
def train(self, epochs):
if not self._data_configured:
raise Exception("Data not configured, please configure..")
if self._params['write_summary']:
tf.global_variables_initializer().run()
loss = np.zeros(epochs)
feed_dict, _ = self.get_data()
if self._tf_sumry_wrtr is not None:
for i in range(epochs):
print "Starting epoch \t", i
round_complete = False
while not round_complete:
if self._params['batch_params'] is not None:
feed_dict, round_complete = self.get_data()
else:
#this is to take care of the case when we are not doing batch training.
round_complete = True
if round_complete:
print "Completed round"
if i % 100 == 99: # Record execution stats
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
summary, loss[i] = self._sess.run(fetches=[self._tf_sumry_wrtr._merged, self._net_ops['train_step']],
feed_dict=feed_dict,
options=run_options,
run_metadata=run_metadata)
self._tf_sumry_wrtr.add_run_metadata(metadata=run_metadata, itr=i)
self._tf_sumry_wrtr.add_summary(summary=summary, itr=i)
print('Adding run metadata for', i)
else: # Record a summary
summary, loss[i] = self._sess.run(fetches=[self._tf_sumry_wrtr._merged, self._net_ops['train_step']],
feed_dict=feed_dict)
self._tf_sumry_wrtr.add_summary(summary=summary, itr=i)
self._tf_sumry_wrtr.close_writer()
else:
with tf.device(self._device):
# Keeping track of loss progress as we train
train_step = self._net_ops['train_step']
loss_op = self._net_ops['cost']
for i in range(epochs):
_, loss[i] = self._sess.run([train_step, loss_op], feed_dict=feed_dict)
return loss
def run_op(self, op_name, x_input):
with tf.device(self._device):
op = self._net_ops[op_name]
out = self._sess.run(op, feed_dict={self._net_ops['x']: x_input})
return out
class MDNPushFwdModel(object):
def __init__(self, sess, network_params):
self._sess = sess
self._params = network_params
self._device = self._params['device']
self._tf_sumry_wrtr = None
self._optimiser = network_params['optimiser']
self._data_configured = False
if network_params['write_summary']:
if 'summary_dir' in network_params:
summary_dir = network_params['summary_dir']
else:
summary_dir = None
self._tf_sumry_wrtr = TfSummaryWriter(tf_session=sess,
summary_dir=summary_dir)
cuda_path = '/usr/local/cuda/extras/CUPTI/lib64'
curr_ld_path = os.environ["LD_LIBRARY_PATH"]
if not cuda_path in curr_ld_path.split(os.pathsep):
print "Enviroment variable LD_LIBRARY_PATH does not contain %s" % cuda_path
print "Please add it, else the program will crash!"
raw_input("Press Ctrl+C")
# os.environ["LD_LIBRARY_PATH"] = curr_ld_path + ':'+cuda_path
with tf.device(self._device):
self._mdn = MixtureDensityNetwork(
network_params, tf_sumry_wrtr=self._tf_sumry_wrtr)
self._mdn._init_model()
self._net_ops = self._mdn._ops
self._init_op = tf.initialize_all_variables()
self._saver = tf.train.Saver()
def init_model(self):
with tf.device(self._device):
self._sess.run(self._init_op)
if self._params['load_saved_model']:
self.load_model()
def configure_data(self, data_x, data_y, batch_creator):
self._data_x = data_x
self._data_y = data_y
self._batch_creator = batch_creator
self._data_configured = True
def get_model_path(self, subscript=None):
model_name_subscript = '_'
if subscript is not None:
if not isinstance(subscript, str):
subscript = str(subscript)
model_name_subscript = subscript + '_'
if 'model_dir' in self._params:
model_dir = self._params['model_dir']
else:
model_path = './siam/'
if not os.path.exists(model_dir):
os.makedirs(model_dir)
if 'model_name' in self._params:
model_name = model_name_subscript + self._params['model_name']
else:
model_name = model_name_subscript + 'siam_model.ckpt'
return model_dir + model_name
def load_model(self, epoch=None):
'''
question: is it better to give filename directly or give epoch number?
'''
load_tf_check_point(session=self._sess,
filename=self.get_model_path(epoch))
def save_model(self, epoch=None):
save_path = self._saver.save(self._sess, self.get_model_path(epoch))
print("Model saved in file: %s" % save_path)
def get_data(self):
round_complete = None
if self._params['batch_params'] is not None:
if self._batch_creator is not None:
self._data_x, self._data_y, round_complete = self._batch_creator.get_batch(
random_samples=self._params['batch_params']
['use_random_batches'])
else:
raise Exception(
"Batch training chosen but batch_creator not configured")
if self._params['cnn_params'] is None:
feed_dict = {
self._net_ops['x']: self._data_x,
self._net_ops['y']: self._data_y
}
else:
feed_dict = {
self._net_ops['image_input']: self._data_x,
self._net_ops['y']: self._data_y
}
return feed_dict, round_complete
def train(self, epochs, chk_pnt_save_invl=10):
if not self._data_configured:
raise Exception("Data not configured, please configure..")
if self._params['write_summary']:
tf.global_variables_initializer().run()
loss = np.zeros(epochs)
feed_dict, _ = self.get_data()
if self._tf_sumry_wrtr is not None:
for i in range(epochs):
print "Starting epoch \t", i
round_complete = False
while not round_complete:
if self._params['batch_params'] is not None:
feed_dict, round_complete = self.get_data()
else:
#this is to take care of the case when we are not doing batch training.
round_complete = True
if round_complete:
print "Completed round"
if i % 100 == 99: # Record execution stats
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
summary, loss[i] = self._sess.run(
fetches=[
self._tf_sumry_wrtr._merged,
self._net_ops['train']
],
feed_dict=feed_dict,
options=run_options,
run_metadata=run_metadata)
self._tf_sumry_wrtr.add_run_metadata(
metadata=run_metadata, itr=i)
self._tf_sumry_wrtr.add_summary(summary=summary, itr=i)
print('Adding run metadata for', i)
else: # Record a summary
summary, loss[i] = self._sess.run(fetches=[
self._tf_sumry_wrtr._merged, self._net_ops['train']
],
feed_dict=feed_dict)
self._tf_sumry_wrtr.add_summary(summary=summary, itr=i)
self._tf_sumry_wrtr.close_writer()
else:
with tf.device(self._device):
# Keeping track of loss progress as we train
train_step = self._net_ops['train']
loss_op = self._net_ops['loss']
for i in range(epochs):
print "Starting epoch \t", i
round_complete = False
batch_no = 0
while not round_complete:
if self._params['batch_params'] is not None:
feed_dict, round_complete = self.get_data()
else:
#this is to take care of the case when we are not doing batch training.
round_complete = True
print "Batch number \t", batch_no
batch_no += 1
_, loss[i] = self._sess.run([train_step, loss_op],
feed_dict=feed_dict)
if round_complete:
print "That was the last round of epoch %d" % i
if i % chk_pnt_save_invl == 0 and i != 0:
self.save_model(epoch=i)
np.savetxt('loss_values.txt', np.asarray(loss))
plt.figure()
plt.plot(loss)
plt.show()
return loss
def run_op(self, op_name, x_input):
with tf.device(self._device):
op = self._net_ops[op_name]
out = self._sess.run(op, feed_dict={self._net_ops['x']: x_input})
return out