-
Notifications
You must be signed in to change notification settings - Fork 0
/
trpo.py
177 lines (154 loc) · 7.95 KB
/
trpo.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
import tensorflow as tf
import numpy as np
from tensorpack import *
from tensorpack.utils.concurrency import ensure_proc_terminate, start_proc_mask_signal
from tensorpack.utils.serialize import dumps
from tensorpack.tfutils.gradproc import MapGradient, SummaryGradient
from tensorpack.utils.gpu import get_nr_gpu
from tensorpack.tfutils.summary import add_moving_summary
from tensorpack.tfutils import get_current_tower_context
from tensorflow.python.eager import context
from tensorflow.python.framework import ops
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import resource_variable_ops
from tensorflow.python.ops import state_ops
from tensorflow.python.training import optimizer
from tensorflow.python.training import training_ops
from tensorflow.python.util.tf_export import tf_export
class ConjugateGradientOptimizer(optimizer.Optimizer):
# cg_iter: conjugate gradient iteration
# ls_max_iter: line search max iteration
# back_trace_ratio: backward line search ratio per iteration
def __init__(self, policy, cost, delta=0.01, cg_iter=10, ls_max_iter=15, back_trace_ratio=0.8, use_locking=False, name="ConjGradient"):
super(ConjugateGradientOptimizer, self).__init__(use_locking, name)
self._cg_iter = cg_iter
self._delta = delta
self._ls_max_iter = ls_max_iter
self._back_trace_ratio = back_trace_ratio
# self._actions = actions
self._policy = policy
self._cost_before = cost
self._mean_KL = tf.reduce_mean(tf.reduce_sum(tf.stop_gradient(self._policy) * tf.log(tf.stop_gradient(self._policy) / (self._policy + 1e-8) + 1e-8), 1))
# Tensor versions of the constructor arguments, created in _prepare().
self._lr_t = None
def _prepare(self):
self._cg_iter_t = ops.convert_to_tensor(self._cg_iter, name='trpo_cg_iter')
self._ls_max_iter_t = ops.convert_to_tensor(self._ls_max_iter, name='trpo_max_iter')
self._delta_t = ops.convert_to_tensor(self._delta, name="trpo_delta")
self._back_trace_ratio_t = ops.convert_to_tensor(self._back_trace_ratio, name='trpo_back_trace_ratio')
def _create_slots(self, var_list):
pass
def cg(self, Hx_fn, g):
r = tf.stop_gradient(g)
p = tf.stop_gradient(r)
x = tf.zeros_like(g)
for i in range(self._cg_iter):
Ap = Hx_fn(p)
rr = tf.reduce_sum(r * r)
alpha = rr / tf.reduce_sum(p * Ap)
x = tf.cond(tf.norm(r) > 1e-5, lambda: x + alpha * p, lambda: tf.identity(x))
r_old = r
r = r_old - alpha * Ap
# r = tf.Print(r, [tf.reduce_sum(r * r_old)])
# due to numerical error, the following check does not pass
# with tf.control_dependencies([tf.assert_equal(tf.reduce_sum(r * r_old), 0.)]):
beta = tf.reduce_sum(r * r) / rr
p_old = p
p = r + beta * p_old
return x
def apply_gradients(self, grads_and_vars, global_step=None, name=None):
# No DistributionStrategy case.
grads_and_vars = tuple(grads_and_vars) # Make sure repeat iteration works.
if not grads_and_vars:
raise ValueError("No variables provided.")
converted_grads_and_vars = []
for g, v in grads_and_vars:
if g is not None:
try:
# Convert the grad to Tensor or IndexedSlices if necessary.
g = ops.convert_to_tensor_or_indexed_slices(g)
except TypeError:
raise TypeError(
"Gradient must be convertible to a Tensor"
" or IndexedSlices, or None: %s" % g)
if not isinstance(g, (ops.Tensor, ops.IndexedSlices)):
raise TypeError(
"Gradient must be a Tensor, IndexedSlices, or None: %s" % g)
# p = _get_processor(v)
converted_grads_and_vars.append((g, v))
# converted_grads_and_vars = tuple(converted_grads_and_vars)
converted_grads_and_vars = tuple([(g, v) for g, v in converted_grads_and_vars if g is not None])
var_list = [v for g, v in converted_grads_and_vars]
cache_var_list = []
for v in var_list:
for c in self.cache_vars:
if c.op.name == v.op.name + 'cache':
cache_var_list.append(c)
break
assert len(var_list) == len(cache_var_list)
if not var_list:
raise ValueError("No gradients provided for any variable: %s." %
([v.name for _, v in converted_grads_and_vars],))
with ops.init_scope():
self._create_slots(var_list)
var_shapes = [v.shape for _, v in converted_grads_and_vars]
slice_idx = np.concatenate([[0], np.cumsum([np.prod(vs) for vs in var_shapes])], 0)
# print(var_shapes)
# print(slice_idx)
with ops.name_scope(name, self._name) as name:
self._prepare()
grad_flatten = tf.concat([tf.reshape(grad, [-1]) for grad, _ in converted_grads_and_vars], 0)
KL_grad = tf.gradients(self._mean_KL, var_list)
KL_grad_flatten = tf.concat([tf.reshape(g, [-1]) for g in KL_grad], 0)
# calculate Hessian * x
def Hx_fn(m):
grads = tf.gradients(tf.reduce_sum(KL_grad_flatten * tf.stop_gradient(m)), var_list)
return tf.concat([tf.reshape(g, [-1]) for g in grads], 0) + 1e-5
x = self.cg(Hx_fn, grad_flatten)
xHx = tf.reduce_sum(tf.transpose(x) * Hx_fn(x))
beta = tf.sqrt(2 * self._delta_t / (xHx + 1e-8))
def get_KL(policy):
return tf.reduce_mean(tf.reduce_sum(tf.stop_gradient(self._policy) * tf.log(
tf.stop_gradient(self._policy) / (policy + 1e-8) + 1e-8), 1))
i = tf.constant(0)
def c(i, beta):
with tf.control_dependencies([control_flow_ops.group(
[state_ops.assign(var, cache_var_list[i] - beta * tf.reshape(x[slice_idx[i]:slice_idx[i + 1]], var_shapes[i])) for
i, (_, var) in enumerate(grads_and_vars)])]):
kl = get_KL(self.policy_fn())
cost = self.cost_fn()
return tf.logical_and(i < self._ls_max_iter_t, tf.logical_or(kl > self._delta_t,
cost > self._cost_before))
b = lambda i, beta: [i + 1, self._back_trace_ratio_t * beta]
i, _ = tf.while_loop(c, b, loop_vars=[i, beta], back_prop=False)
var_update = tf.cond(tf.logical_or(tf.equal(i, self._ls_max_iter_t), tf.logical_not(tf.reduce_any(tf.is_nan(x)))),
lambda: self.cache2var,
lambda: self.var2cache)
if not context.executing_eagerly():
if isinstance(var_update, ops.Tensor):
var_update = var_update.op
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
if var_update not in train_op:
train_op.append(var_update)
return var_update
class TestModel(ModelDesc):
def inputs(self):
return [tf.placeholder_with_default(tf.ones([3, 5]), [3, 5], name='x')]
def build_graph(self, x):
action = FullyConnected('fc', x, 5, activation=tf.nn.softmax)
self.action = tf.identity(action, name='action')
loss = tf.reduce_sum(self.action * tf.Variable(np.expand_dims(np.arange(5), 0), trainable=False, dtype=tf.float32))
add_moving_summary(self.action[-1, -1])
return loss
def optimizer(self):
return ConjugateGradientOptimizer(self.action, 0.1)
if __name__ == '__main__':
trainer = SimpleTrainer()
config = TrainConfig(
model=TestModel(),
dataflow=FakeData(shapes=[[3, 5]]),
steps_per_epoch=500,
max_epoch=1
)
launch_train_with_config(config, trainer)