def run_loop(self, n_unrolls):
ts = self._env.reset()
self._traj.reset()
self._traj.start(next_state=self._shell.next_state, **dict(ts._asdict()))
i = 0
system_logs = {}
while True:
if n_unrolls is not None:
if i == n_unrolls:
return
with U.Timer() as shell_step_timer:
step_output = self._shell.step(step_type=ts.step_type,
reward=ts.reward,
observation=ts.observation)
with U.Timer() as env_step_timer:
ts = self._env.step(step_output.action)
self._traj.add(step_output=step_output, **dict(ts._asdict()))
if len(self._traj) == self._traj_length + 1:
with U.Timer() as send_experience_timer:
exps = self._traj.debatch_and_stack()
self._traj.reset()
self._send_experiences(exps)
self._traj.start(next_state=self._shell.next_state, **dict(ts._asdict()))
system_logs['put_experience_async_sec'] = send_experience_timer.to_seconds()
for logger in self._system_loggers:
logger.write(
dict(shell_step_time_sec=shell_step_timer.to_seconds(),
env_step_time_sec=env_step_timer.to_seconds(),
**system_logs))
i += 1
def _scip_solve(self, solver):
"""solves a mip/lp using scip"""
if solver is None:
solver = Model()
solver.hideOutput()
if self.config.disable_maxcuts:
for param in [
'separating/maxcuts', 'separating/maxcutsroot', 'propagating/maxrounds',
'propagating/maxroundsroot', 'presolving/maxroundsroot'
]:
solver.setIntParam(param, 0)
solver.setBoolParam('conflict/enable', False)
solver.setPresolve(SCIP_PARAMSETTING.OFF)
solver.setBoolParam('randomization/permutevars', True)
# seed is set to 0 permanently.
solver.setIntParam('randomization/permutationseed', 0)
solver.setIntParam('randomization/randomseedshift', 0)
with U.Timer() as timer:
solver.optimize()
assert solver.getStatus() == 'optimal', solver.getStatus()
obj = float(solver.getObjVal())
ass = {var.name: solver.getVal(var) for var in solver.getVars()}
mip_stats = ConfigDict(mip_work=solver.getNNodes(),
n_cuts=solver.getNCuts(),
n_cuts_applied=solver.getNCutsApplied(),
n_lps=solver.getNLPs(),
solving_time=solver.getSolvingTime(),
pre_solving_time=solver.getPresolvingTime(),
time_elapsed=timer.to_seconds())
return ass, obj, mip_stats
def testUpdate(self):
global B, T
if not FLAGS.testUpdate:
return
self._setup()
env = _get_env()
shell = _create_shell(env)
learner = _create_learner(env, shell)
exps = self._sample_trajectory(env, shell)
batch = learner.batch_and_preprocess_trajs(exps)
print('***************')
print('Starting....')
print('***************')
update_times = []
N = 100
for _ in range(10):
learner.update(batch)
for i in trange(N):
with U.Timer() as timer:
learner.update(batch)
update_times.append(timer.to_seconds())
print(
f'\nPer update step time taken (First half): {np.mean(update_times[:N // 2])}'
)
print(
f'Per update step time taken (Second half): {np.mean(update_times[N // 2:])}'
)
print('Test complete!')
def main(argv):
env = make_env()
ts = env.reset()
with U.Timer() as timer:
for i in trange(500):
mask = ts.observation['mask']
act = np.random.choice(len(mask), env.k, replace=False, p=mask / sum(mask))
ts = env.step(act)
print('Total time taken: ', timer.to_seconds())
def main(_):
milp = get_sample('milp-cauction-300-filtered', 'train', 102)
mip = SCIPMIPInstance.fromMIPInstance(milp.mip)
times = []
for _ in range(10):
with U.Timer() as timer:
model = mip.get_scip_model()
times.append(timer.to_seconds())
print(f'Avg time to copy the model: {np.mean(times[5:])}')
for i in tqdm(range(20)):
fixed_ass = {
k: milp.feasible_solution[k]
for k in np.random.permutation(list(milp.feasible_solution.keys()))[:500]
}
ass, obj = fix_and_solve(model, fixed_ass)
print(obj)
model.freeTransform()
def run_branch_and_bound_scip(m, heuristic):
# m must be presolved before passing
m.includeHeur(
heuristic,
"PyEvalHeur",
"custom heuristic implemented in python to evaluate RL agent",
"Y",
timingmask=SCIP_HEURTIMING.BEFORENODE)
with U.Timer() as timer:
m.optimize()
# collect stats
results = ConfigDict(mip_work=m.getNNodes(),
n_cuts=m.getNCuts(),
n_cuts_applied=m.getNCutsApplied(),
n_lps=m.getNLPs(),
pre_solving_time=m.getPresolvingTime(),
solving_time=m.getSolvingTime(),
time_elapsed=timer.to_seconds())
# m.freeProb()
return results
def main(self):
for _ in range(self.config.n_train_steps):
system_logs = dict()
# fetch the next training batch
with U.Timer() as batch_timer:
batch = self._exp_fetcher.get()
with U.Timer() as step_timer:
# run update step on the sampled batch
feed_dict = {
ph: val
for ph, val in zip(nest.flatten(self._traj_phs),
nest.flatten(batch))
}
profile_kwargs = {}
if self.global_step == self._profile_step:
profile_kwargs = dict(options=tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE),
run_metadata=tf.RunMetadata())
log_vals = self._agent.update(self.sess, feed_dict,
profile_kwargs)
if profile_kwargs:
self._save_profile(**profile_kwargs)
with U.Timer() as log_timer:
for logger in self._loggers:
logger.write(log_vals)
# after first sess.run finishes send the metagraph.
if self.global_step == 1:
self._send_metagraph()
# publish the variables if required.
if self._publish_tracker.track_increment():
with U.Timer() as publish_timer:
self._publish_variables()
system_logs['publish_time_sec'] = publish_timer.to_seconds()
# Checkpoint if required
if self.global_step % self._checkpoint_every == 0:
with U.Timer() as ckpt_timer:
self._create_ckpt()
system_logs['ckpt_time_sec'] = ckpt_timer.to_seconds()
with U.Timer() as system_log_timer:
# log system profile
for logger in self._system_loggers:
logger.write(
dict(global_step=self.global_step,
sps=self._batch_size * self._traj_length /
float(step_timer.to_seconds()),
per_step_time_sec=step_timer.to_seconds(),
batch_fetch_time_sec=batch_timer.to_seconds(),
**system_logs))
system_logs['log_time_sec'] = log_timer.to_seconds(
) + system_log_timer.to_seconds()
self._publish_queue.put(None) # exit the thread once training ends.
