def do_training(self, epochs=100, test_interval=10, show_interval=50):
errors = []
if test_interval: self.avg_vector_distances = []
self.current_session = sess = TFT.gen_initialized_session()
step = 0
for i in range(epochs):
error = 0
grabvars = [self.error]
for c in self.cases:
feeder = {self.input: [c[0]], self.target: [c[1]]}
_, grabvals, _ = self.run_one_step([self.trainer],
grabvars,
step=step,
show_interval=show_interval,
session=sess,
feed_dict=feeder)
error += grabvals[0]
step += 1
errors.append(error)
if (test_interval and i % test_interval == 0):
self.avg_vector_distances.append(
calc_avg_vect_dist(self.do_testing(sess, scatter=False)))
PLT.figure()
TFT.simple_plot(errors, xtitle="Epoch", ytitle="Error", title="")
if test_interval:
PLT.figure()
TFT.simple_plot(self.avg_vector_distances,
xtitle='Epoch',
ytitle='Avg Hidden-Node Vector Distance',
title='')
def run_one_step(self,
operators,
grabbed_vars=None,
probed_vars=None,
dir='probeview',
session=None,
feed_dict=None,
step=1,
show_interval=1):
""" Similar to the "quickrun" functions used earlier. """
sess = session if session else TFT.gen_initialized_session(dir=dir)
if probed_vars is not None:
results = sess.run([operators, grabbed_vars, probed_vars],
feed_dict=feed_dict)
sess.probe_stream.add_summary(results[2], global_step=step)
else:
results = sess.run([operators, grabbed_vars], feed_dict=feed_dict)
if show_interval and (step % show_interval == 0):
self.display_grabvars(results[1], grabbed_vars, step=step)
return results[0], results[1], sess
def run_one_step(self, operators, grabbed_vars=None, probed_vars=None, dir='probeview',
session=None, feed_dict=None, step=1):
sess = session if session else TFT.gen_initialized_session(dir=dir)
if probed_vars is not None:
results = sess.run([operators, grabbed_vars, probed_vars], feed_dict=feed_dict)
sess.probe_stream.add_summary(results[2], global_step=step)
else:
results = sess.run([operators, grabbed_vars], feed_dict=feed_dict)
return results[0], results[1], sess
def run_one_step(self, operators, grabbed_vars=None, probed_vars=None, dir="probeview",
session=None, feed_dict=None, step=1, display_interval=1, testing = False):
sess = session if session else TFT.gen_initialized_session(dir=dir)
if probed_vars is not None:
results = sess.run([operators, grabbed_vars, probed_vars], feed_dict=feed_dict)
sess.probe_stream.add_summary(results[2], global_step=step)
else:
results = sess.run([operators, grabbed_vars], feed_dict=feed_dict)
if display_interval and (step % display_interval == 0):
self.display_grabvars(results[1], grabbed_vars, step=step)
return results[0], results[1], sess
def training_session(self,
epochs,
sess=None,
dir="probeview",
continued=False):
self.roundup_probes()
session = sess if sess else TFT.gen_initialized_session(dir=dir)
self.current_session = session
self.do_training(session,
self.caseman.get_training_cases(),
epochs,
continued=continued)
def quickrun4(operators,
grabbed_vars=None,
dir='probeview',
session=None,
feed_dict=None,
step=1,
show_interval=1):
sess = session if session else TFT.gen_initialized_session(dir=dir)
results = sess.run([operators, grabbed_vars], feed_dict=feed_dict)
if (show_interval and step % show_interval) == 0:
TFT.show_results(results[1], grabbed_vars, dir)
return results[0], results[1], sess
def training_session(self,
steps,
sess=None,
dir="probeview",
continued=False):
session = sess if sess else TFT.gen_initialized_session(dir=dir)
self.current_session = session
self.roundup_probes(
) # this call must come AFTER the session is created, else graph is not in tensorboard.
self.do_training(session,
self.caseman.get_training_cases(),
steps,
continued=continued)
def do_training(self,
epochs=100,
test_interval=10,
show_interval=50,
mbs=100):
errors = []
self.val_error = []
self.train_error = []
if test_interval: self.avg_vector_distances = []
self.current_session = sess = TFT.gen_initialized_session()
for i in range(epochs):
self.current_epoch = i
error = 0
grabvars = [self.error]
step = self.global_step + i
ncases = len(self.training_cases)
nmb = math.ceil(ncases / mbs)
for cstart in range(0, ncases, mbs):
cend = min(ncases, cstart + mbs)
minibatch = self.training_cases[cstart:cend]
feeder = {
self.input: [c[0] for c in minibatch],
self.target: [c[1] for c in minibatch]
}
_, grabvals, _ = self.run_one_step([self.trainer],
grabvars,
step=step,
show_interval=show_interval,
session=sess,
feed_dict=feeder)
error += grabvals[0]
errors.append([i, error])
if (test_interval and i % test_interval == 0):
self.do_testing(sess,
scatter=False,
mbs=len(self.training_cases),
testset="training")
if (len(self.validation_cases) != 0):
self.do_testing(sess,
scatter=False,
mbs=len(self.validation_cases),
testset="validation")
if (len(self.testing_cases) != 0):
self.do_testing(sess, scatter=False, mbs=mbs, testset="testing")
#TFT.simple_plot(errors,xtitle="Epoch",ytitle="Error",title="")
TFT.plot_training_history(self.train_error, self.val_error)
def loadParams(self, layerDims, loadPath, globalStep):
self.anet = ANET(
layer_dims = layerDims,
softmax=True,
case_manager = CaseManager([]))
session = TFT.gen_initialized_session(dir="probeview")
self.anet.current_session = session
state_vars = []
for m in self.anet.layer_modules:
vars = [m.getvar('wgt'), m.getvar('bias')]
state_vars = state_vars + vars
self.anet.state_saver = tf.train.Saver(state_vars)
self.anet.state_saver.restore(self.anet.current_session, loadPath+"-"+str(globalStep))
def runOneStep(self, operators, grabbedVars = None, probedVars = None, dir = 'probeview', session = None, feedDict = None, step = 1, showInterval = 1): sess = session if session else tft.gen_initialized_session(dir=dir) if probed_vars is not None: results = sess.run([operators, grabbedVars, probedVars], feed_dict = feedDict) sess.probe_stream.add_summary(results[2], global_step = step) else: results = sess.run([operators, grabbedVars, probedVars], feed_dict = feedDict) if showInterval and (step % showInterval == 0): self.displayGrabvars(results[1], grabbedVars, step = step) return results[0], results[1], sess
def tfex8(size=5, steps=50, tvect=None, learning_rate=0.5, showint=10):
target = tvect if tvect else np.ones((1, size))
w = tf.Variable(np.random.uniform(-.1, .1, size=(size, size)),
name='weights') # weights applied to x.
b = tf.Variable(np.zeros((1, size)), name='bias') # bias terms
x = tf.placeholder(tf.float64, shape=(1, size), name='input')
y = tf.sigmoid(
tf.matmul(x, w) + b, name='out-sigmoid'
) # Gather all weighted inputs, then apply activation function
error = tf.reduce_mean(tf.square(target - y))
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
training_operator = optimizer.minimize(error)
feeder = {x: np.random.uniform(-1, 1, size=(1, size))}
sess = TFT.gen_initialized_session()
for step in range(steps):
quickrun4([training_operator], [w, b, y],
session=sess,
feed_dict=feeder,
step=step,
show_interval=showint)
TFT.close_session(sess)
def do_training(self,
epochs=100,
test_interval=10,
show_interval=50,
mbs=100):
errors = []
if test_interval: self.avg_vector_distances = []
self.current_session = sess = TFT.gen_initialized_session()
for i in range(epochs):
error = 0
grabvars = [self.error]
step = self.global_step + i
ncases = len(self.cases)
nmb = math.ceil(ncases / mbs)
for cstart in range(0, ncases, mbs):
cend = min(ncases, cstart + mbs)
minibatch = self.cases[cstart:cend]
feeder = {
self.input: [c[0] for c in minibatch],
self.target: [c[1] for c in minibatch]
}
_, grabvals, _ = self.run_one_step([self.trainer],
grabvars,
step=step,
show_interval=show_interval,
session=sess,
feed_dict=feeder)
error += grabvals[0]
errors.append(error)
if (test_interval and i % test_interval == 0):
self.do_testing(sess, scatter=False, mbs=mbs)
PLT.figure()
TFT.simple_plot(errors, xtitle="Epoch", ytitle="Error", title="")
if test_interval:
PLT.figure()
TFT.simple_plot(self.avg_vector_distances,
xtitle='Epoch',
ytitle='Avg Hidden-Node Vector Distance',
title='')
def run_one_step(self,
operators,
grabbed_vars=None,
probed_vars=None,
dir='probeview',
session=None,
feed_dict=None,
step=1,
show_interval=1,
mapping=False,
dendrogram=False,
onezero=False,
komma=False,
punktum=False,
decimals=1,
dendro_time=30,
leaf_font_size=None):
sess = session if session else TFT.gen_initialized_session(dir=dir)
if probed_vars is not None:
results = sess.run([operators, grabbed_vars, probed_vars],
feed_dict=feed_dict)
sess.probe_stream.add_summary(results[2], global_step=step)
else:
results = sess.run([operators, grabbed_vars], feed_dict=feed_dict)
if show_interval and (step % show_interval == 0):
self.display_grabvars(results[1], grabbed_vars, step=step)
if mapping:
self.display_grabvars(results[1], grabbed_vars, step=step)
if dendrogram:
self.display_dendrograms(results[1],
grabbed_vars,
step=step,
onezero=onezero,
komma=komma,
punktum=punktum,
decimals=decimals,
sleep_time=dendro_time,
leaf_font_size=leaf_font_size)
return results[0], results[1], sess
def setupSession(self, sess=None, dir="probeview"):
session = sess if sess else TFT.gen_initialized_session(dir=dir)
self.roundup_probes()
self.current_session = session
