我想在Keras中实现SPSA优化器。以前，我从https://github.com/fraunhofer-iais/tensorflow_spsa使用TF的SPSA实现
我发现我们可以在model.compile(...)期间直接向Keras提供TF优化器。但是，要使优化程序正常工作，它应该坚持使用标准的优化程序实施，并定义正确的compute_gradients()和apply gradients()。
我尝试如下修改代码：
import tensorflow as tf
from tensorflow.python.ops import control_flow_ops
from tensorflow.contrib import graph_editor as ge
from tensorflow.contrib.distributions import Bernoulli
class SimultaneousPerturbationOptimizer(tf.train.Optimizer):
def __init__(self, inputs=None, graph=None, a=0.01, c=0.01, alpha=1.0, gamma=0.4, use_locking=False, name="SPSA"):
super(SimultaneousPerturbationOptimizer, self).__init__(use_locking, name)
self.work_graph = tf.get_default_graph() if graph is None else graph # the graph to work on
self.tvars = [var.name.split(':')[0] for var in tf.trainable_variables()] # list of names of trainable variables
self.inputs = inputs
self.num_params = 0
self.loss = None
# optimizer parameters
self.a = tf.constant(a, dtype=tf.float32, name = "SPSA_a" )
self.c = tf.constant(c, dtype=tf.float32, name = "SPSA_c" )
self.alpha = tf.constant(alpha, dtype=tf.float32, name = "SPSA_alpha" )
self.gamma = tf.constant(gamma, dtype=tf.float32, name = "SPSA_gamma" )
self.global_step_tensor = tf.Variable(0, name='global_step', trainable=False)
# print( "SPSA: a = {:4.3f} c = {:4.3f} alpha = {:4.3f} gamma = {:4.3f}".format(a,c,alpha,gamma) )
# print( "Trainable Variables:" )
# for var in tf.trainable_variables():
# var_name = var.name.split(':')[0]
# print( " {} => {}".format(var.name,var_name) )
def _clone_model(self, model, perturbations, dst_scope):
''' make a copy of model and connect the resulting sub-graph to
input ops of the original graph and parameter assignments by
perturbator.
'''
def not_placeholder_or_trainvar_filter(op):
if op.type == 'Placeholder': # evaluation sub-graphs will be fed from original placeholders
return False
for var_name in self.tvars:
# print( var_name, type(var_name) )
if op.name.startswith(var_name): # remove Some/Var/(read,assign,...) -- will be replaced with perturbations
return False
return True
ops_without_inputs = ge.filter_ops(model.ops, not_placeholder_or_trainvar_filter)
# remove init op from clone if already present
try:
ops_without_inputs.remove(self.work_graph.get_operation_by_name("init"))
except:
pass
clone_sgv = ge.make_view(ops_without_inputs)
clone_sgv = clone_sgv.remove_unused_ops(control_inputs=True)
input_replacements = {}
for t in clone_sgv.inputs:
if t.name in perturbations.keys(): # input from trainable var --> replace with perturbation
input_replacements[t] = perturbations[t.name]
else: # otherwise take input from original graph
input_replacements[t] = self.work_graph.get_tensor_by_name(t.name)
return ge.copy_with_input_replacements(clone_sgv, input_replacements, dst_scope=dst_scope)
def _mul_dims(self, shape):
n = 1
for d in shape:
n *= d.value
return n
# def minimize( self, loss, var_list = None, global_step = None ):
# orig_graph_view = None
# trainable_vars = var_list if var_list != None else tf.trainable_variables()
# if self.inputs is not None:
# seed_ops = [t.op for t in self.inputs]
# result = list(seed_ops)
# wave = set(seed_ops)
# while wave: # stolen from grap_editor.select
# new_wave = set()
# for op in wave:
# for new_t in op.outputs:
# if new_t == loss:
# continue
# for new_op in new_t.consumers():
# #if new_op not in result and is_within(new_op):
# if new_op not in result:
# new_wave.add(new_op)
# for op in new_wave:
# if op not in result:
# result.append(op)
# wave = new_wave
# orig_graph_view = ge.sgv(result)
# else:
# orig_graph_view = ge.sgv(self.work_graph)
# self.global_step_tensor = tf.Variable(0, name='global_step', trainable=False) if global_step is None else global_step
# # Perturbations
# deltas = {}
# n_perturbations = {}
# p_perturbations = {}
# with tf.name_scope("Perturbator"):
# self.c_t = tf.div( self.c, tf.pow(tf.add(tf.cast(self.global_step_tensor, tf.float32),
# tf.constant(1, dtype=tf.float32)), self.gamma), name = "SPSA_ct" )
# # self.c_t = 0.00 #MOD
# for var in trainable_vars:
# self.num_params += self._mul_dims(var.get_shape())
# var_name = var.name.split(':')[0]
# random = Bernoulli(tf.fill(var.get_shape(), 0.5), dtype=tf.float32)
# deltas[var] = tf.subtract( tf.constant(1, dtype=tf.float32),
# tf.scalar_mul(tf.constant(2, dtype=tf.float32),random.sample(1)[0]), name = "SPSA_delta" )
# c_t_delta = tf.scalar_mul( tf.reshape(self.c_t, []), deltas[var] )
# n_perturbations[var_name+'/read:0'] = tf.subtract( var, c_t_delta, name = "perturb_n" )
# p_perturbations[var_name+'/read:0'] = tf.add(var, c_t_delta, name = "perturb_p" )
# # print("{} parameters".format(self.num_params))
# # Evaluator
# with tf.name_scope("Evaluator"):
# _, self.ninfo = self._clone_model(orig_graph_view, n_perturbations, 'N_Eval')
# _, self.pinfo = self._clone_model(orig_graph_view, p_perturbations, 'P_Eval')
# # Weight Updater
# optimizer_ops = []
# with tf.control_dependencies([loss]):
# with tf.name_scope('Updater'):
# a_t = self.a / (tf.pow(tf.add(tf.cast(self.global_step_tensor, tf.float32),
# tf.constant(1, dtype=tf.float32)), self.alpha))
# # a_t = 0.00 #MOD
# for var in trainable_vars:
# l_pos = self.pinfo.transformed( loss )
# l_neg = self.ninfo.transformed( loss )
# # print( "l_pos: ", l_pos)
# # print( "l_neg: ", l_neg)
# ghat = (l_pos - l_neg) / (tf.constant(2, dtype=tf.float32) * self.c_t * deltas[var])
# optimizer_ops.append(tf.assign_sub(var, a_t*ghat))
# grp = control_flow_ops.group(*optimizer_ops)
# with tf.control_dependencies([grp]):
# tf.assign_add(self.global_step_tensor, tf.constant(1, dtype=self.global_step_tensor.dtype))
# return grp
def compute_gradients( self, loss = None, var_list = None ):
# print( "this" )
orig_graph_view = None
trainable_vars = var_list if var_list != None else tf.trainable_variables()
if self.inputs is not None:
seed_ops = [t.op for t in self.inputs]
result = list(seed_ops)
wave = set(seed_ops)
while wave: # stolen from grap_editor.select
new_wave = set()
for op in wave:
for new_t in op.outputs:
if new_t == loss:
continue
for new_op in new_t.consumers():
#if new_op not in result and is_within(new_op):
if new_op not in result:
new_wave.add(new_op)
for op in new_wave:
if op not in result:
result.append(op)
wave = new_wave
orig_graph_view = ge.sgv(result)
else:
orig_graph_view = ge.sgv(self.work_graph)
self.global_step_tensor = tf.Variable(0, name='global_step', trainable=False)
# Perturbations
deltas = {}
n_perturbations = {}
p_perturbations = {}
with tf.name_scope("Perturbator"):
self.c_t = tf.div( self.c, tf.pow(tf.add(tf.cast(self.global_step_tensor, tf.float32),
tf.constant(1, dtype=tf.float32)), self.gamma), name = "SPSA_ct" )
# self.c_t = 0.00 #MOD
for var in trainable_vars:
self.num_params += self._mul_dims(var.get_shape())
var_name = var.name.split(':')[0]
random = Bernoulli(tf.fill(var.get_shape(), 0.5), dtype=tf.float32)
deltas[var] = tf.subtract( tf.constant(1, dtype=tf.float32),
tf.scalar_mul(tf.constant(2, dtype=tf.float32),random.sample(1)[0]), name = "SPSA_delta" )
c_t_delta = tf.scalar_mul( tf.reshape(self.c_t, []), deltas[var] )
n_perturbations[var_name+'/read:0'] = tf.subtract( var, c_t_delta, name = "perturb_n" )
p_perturbations[var_name+'/read:0'] = tf.add(var, c_t_delta, name = "perturb_p" )
# print("{} parameters".format(self.num_params))
# Evaluator
with tf.name_scope("Evaluator"):
_, self.ninfo = self._clone_model(orig_graph_view, n_perturbations, 'N_Eval')
_, self.pinfo = self._clone_model(orig_graph_view, p_perturbations, 'P_Eval')
# Weight Updater
optimizer_ops = []
grads_op = []
with tf.control_dependencies([loss]):
with tf.name_scope('Updater'):
a_t = self.a / (tf.pow(tf.add(tf.cast(self.global_step_tensor, tf.float32),
tf.constant(1, dtype=tf.float32)), self.alpha))
# a_t = 0.00 #MOD
for var in trainable_vars:
l_pos = self.pinfo.transformed( loss )
l_neg = self.ninfo.transformed( loss )
# print( "l_pos: ", l_pos)
# print( "l_neg: ", l_neg)
ghat = (l_pos - l_neg) / (tf.constant(2, dtype=tf.float32) * self.c_t * deltas[var])
grads_op.append( ghat )
# optimizer_ops.append(tf.assign_sub(var, a_t*ghat))
# grp = control_flow_ops.group(*optimizer_ops)
# with tf.control_dependencies([grp]):
# tf.assign_add(self.global_step_tensor, tf.constant(1, dtype=self.global_step_tensor.dtype))
# return grp
grads_and_vars = list( zip(grads_op,trainable_vars) )
return grads_and_vars
def apply_gradients(self, grad_and_vars, global_step=None):
self.global_step_tensor = tf.Variable(0, name='global_step', trainable=False) if global_step is None else global_step
a_t = self.a / (tf.pow(tf.add(tf.cast(self.global_step_tensor, tf.float32),
tf.constant(1, dtype=tf.float32)), self.alpha))
optimizer_ops = []
for (g,v) in grad_and_vars:
optimizer_ops.append(tf.assign_sub(v,a_t*g))
grp = control_flow_ops.group(*optimizer_ops)
with tf.control_dependencies([grp]):
tf.assign_add(self.global_step_tensor, tf.constant(1, dtype=self.global_step_tensor.dtype))
return grp
def get_clones(self, op):
return (self.ninfo.transformed(op), self.pinfo.transformed(op))
def get_clones(self, op):
return (self.ninfo.transformed(op), self.pinfo.transformed(op))
但是我遇到了这个错误：
---------------------------------------------------------------------------
FailedPreconditionError Traceback (most recent call last)
<ipython-input-16-fde0f2b12519> in <module>()
----> 1 train_log = model2.fit(x_train, y_train, batch_size=batch_size, epochs=epochs, validation_data=(x_test,y_test))
/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, max_queue_size, workers, use_multiprocessing, **kwargs)
878 initial_epoch=initial_epoch,
879 steps_per_epoch=steps_per_epoch,
--> 880 validation_steps=validation_steps)
881
882 def evaluate(self,
/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training_arrays.py in model_iteration(model, inputs, targets, sample_weights, batch_size, epochs, verbose, callbacks, val_inputs, val_targets, val_sample_weights, shuffle, initial_epoch, steps_per_epoch, validation_steps, mode, validation_in_fit, **kwargs)
327
328 # Get outputs.
--> 329 batch_outs = f(ins_batch)
330 if not isinstance(batch_outs, list):
331 batch_outs = [batch_outs]
/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/backend.py in __call__(self, inputs)
3074
3075 fetched = self._callable_fn(*array_vals,
-> 3076 run_metadata=self.run_metadata)
3077 self._call_fetch_callbacks(fetched[-len(self._fetches):])
3078 return nest.pack_sequence_as(self._outputs_structure,
/usr/local/lib/python3.6/dist-packages/tensorflow/python/client/session.py in __call__(self, *args, **kwargs)
1437 ret = tf_session.TF_SessionRunCallable(
1438 self._session._session, self._handle, args, status,
-> 1439 run_metadata_ptr)
1440 if run_metadata:
1441 proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)
/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/errors_impl.py in __exit__(self, type_arg, value_arg, traceback_arg)
526 None, None,
527 compat.as_text(c_api.TF_Message(self.status.status)),
--> 528 c_api.TF_GetCode(self.status.status))
529 # Delete the underlying status object from memory otherwise it stays alive
530 # as there is a reference to status from this from the traceback due to
FailedPreconditionError: Attempting to use uninitialized value training/TFOptimizer/global_step
[[{{node training/TFOptimizer/global_step/read}}]]
我对tensorflow低级别的经验很少，无法理解问题所在。
我该如何解决？