Bart/bullet3
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

add yapf style and apply yapf to format all Python files

Browse Source 
This recreates pull request #2192
This commit is contained in:
Erwin Coumans
2019-04-27 07:31:15 -07:00
parent c591735042
commit ef9570c315
347 changed files with 70304 additions and 22752 deletions
Download Patch File Download Diff File Expand all files Collapse all files

157
examples/pybullet/gym/pybullet_envs/deep_mimic/learning/solvers/mpi_solver.py
View File

@@ -8,96 +8,97 @@ from pybullet_utils.logger import Logger
from pybullet_envs.deep_mimic.learning.solvers.solver import Solver
class MPISolver(Solver):
CHECK_SYNC_ITERS = 1000
CHECK_SYNC_ITERS = 1000
def __init__(self, sess, optimizer, vars):
super().__init__(vars)
self.sess = sess
self.optimizer = optimizer
self._build_grad_feed(vars)
self._update = optimizer.apply_gradients(zip(self._grad_tf_list, self.vars))
self._set_flat_vars = TFUtil.SetFromFlat(sess, self.vars)
self._get_flat_vars = TFUtil.GetFlat(sess, self.vars)
def __init__(self, sess, optimizer, vars):
super().__init__(vars)
self.sess = sess
self.optimizer = optimizer
self._build_grad_feed(vars)
self._update = optimizer.apply_gradients(zip(self._grad_tf_list, self.vars))
self._set_flat_vars = TFUtil.SetFromFlat(sess, self.vars)
self._get_flat_vars = TFUtil.GetFlat(sess, self.vars)
self.iter = 0
grad_dim = self._calc_grad_dim()
self._flat_grad = np.zeros(grad_dim, dtype=np.float32)
self._global_flat_grad = np.zeros(grad_dim, dtype=np.float32)
return
self.iter = 0
grad_dim = self._calc_grad_dim()
self._flat_grad = np.zeros(grad_dim, dtype=np.float32)
self._global_flat_grad = np.zeros(grad_dim, dtype=np.float32)
def get_stepsize(self):
return self.optimizer._learning_rate_tensor.eval()
return
def update(self, grads=None, grad_scale=1.0):
if grads is not None:
self._flat_grad = MathUtil.flatten(grads)
else:
self._flat_grad.fill(0)
return self.update_flatgrad(self._flat_grad, grad_scale)
def get_stepsize(self):
return self.optimizer._learning_rate_tensor.eval()
def update_flatgrad(self, flat_grad, grad_scale=1.0):
if self.iter % self.CHECK_SYNC_ITERS == 0:
assert self.check_synced(), Logger.print2('Network parameters desynchronized')
if grad_scale != 1.0:
flat_grad *= grad_scale
def update(self, grads=None, grad_scale=1.0):
if grads is not None:
self._flat_grad = MathUtil.flatten(grads)
else:
self._flat_grad.fill(0)
return self.update_flatgrad(self._flat_grad, grad_scale)
MPI.COMM_WORLD.Allreduce(flat_grad, self._global_flat_grad, op=MPI.SUM)
self._global_flat_grad /= MPIUtil.get_num_procs()
def update_flatgrad(self, flat_grad, grad_scale=1.0):
if self.iter % self.CHECK_SYNC_ITERS == 0:
assert self.check_synced(), Logger.print2('Network parameters desynchronized')
self._load_flat_grad(self._global_flat_grad)
self.sess.run([self._update], self._grad_feed)
self.iter += 1
if grad_scale != 1.0:
flat_grad *= grad_scale
return
MPI.COMM_WORLD.Allreduce(flat_grad, self._global_flat_grad, op=MPI.SUM)
self._global_flat_grad /= MPIUtil.get_num_procs()
def sync(self):
vars = self._get_flat_vars()
MPIUtil.bcast(vars)
self._set_flat_vars(vars)
return
self._load_flat_grad(self._global_flat_grad)
self.sess.run([self._update], self._grad_feed)
self.iter += 1
def check_synced(self):
synced = True
if self._is_root():
vars = self._get_flat_vars()
MPIUtil.bcast(vars)
else:
vars_local = self._get_flat_vars()
vars_root = np.empty_like(vars_local)
MPIUtil.bcast(vars_root)
synced = (vars_local == vars_root).all()
return synced
return
def _is_root(self):
return MPIUtil.is_root_proc()
def _build_grad_feed(self, vars):
self._grad_tf_list = []
self._grad_buffers = []
for v in self.vars:
shape = v.get_shape()
grad = np.zeros(shape)
grad_tf = tf.placeholder(tf.float32, shape=shape)
self._grad_buffers.append(grad)
self._grad_tf_list.append(grad_tf)
def sync(self):
vars = self._get_flat_vars()
MPIUtil.bcast(vars)
self._set_flat_vars(vars)
return
self._grad_feed = dict({g_tf: g for g_tf, g in zip(self._grad_tf_list, self._grad_buffers)})
return
def check_synced(self):
synced = True
if self._is_root():
vars = self._get_flat_vars()
MPIUtil.bcast(vars)
else:
vars_local = self._get_flat_vars()
vars_root = np.empty_like(vars_local)
MPIUtil.bcast(vars_root)
synced = (vars_local == vars_root).all()
return synced
def _calc_grad_dim(self):
grad_dim = 0
for grad in self._grad_buffers:
grad_dim += grad.size
return grad_dim
def _is_root(self):
return MPIUtil.is_root_proc()
def _load_flat_grad(self, flat_grad):
start = 0
for g in self._grad_buffers:
size = g.size
np.copyto(g, np.reshape(flat_grad[start:start + size], g.shape))
start += size
return
def _build_grad_feed(self, vars):
self._grad_tf_list = []
self._grad_buffers = []
for v in self.vars:
shape = v.get_shape()
grad = np.zeros(shape)
grad_tf = tf.placeholder(tf.float32, shape=shape)
self._grad_buffers.append(grad)
self._grad_tf_list.append(grad_tf)
self._grad_feed = dict({g_tf: g for g_tf, g in zip(self._grad_tf_list, self._grad_buffers)})
return
def _calc_grad_dim(self):
grad_dim = 0
for grad in self._grad_buffers:
grad_dim += grad.size
return grad_dim
def _load_flat_grad(self, flat_grad):
start = 0
for g in self._grad_buffers:
size = g.size
np.copyto(g, np.reshape(flat_grad[start:start + size], g.shape))
start += size
return

18
examples/pybullet/gym/pybullet_envs/deep_mimic/learning/solvers/solver.py
View File

@@ -1,15 +1,17 @@
from abc import abstractmethod
import sys, abc
if sys.version_info >= (3, 4):
ABC = abc.ABC
ABC = abc.ABC
else:
ABC = abc.ABCMeta('ABC', (), {})
ABC = abc.ABCMeta('ABC', (), {})
class Solver(ABC):
def __init__(self, vars):
self.vars = vars
return
@abstractmethod
def update(self, grads):
pass
def __init__(self, vars):
self.vars = vars
return
@abstractmethod
def update(self, grads):
pass