def main():
args = []
for line in fileinput.input():
args.append(line.rstrip())
input = args[0]
skew_array = iterative_skew(input)
utils.print_array(skew_array)
def main():
args = []
for line in fileinput.input():
args.append(line.rstrip())
input = args[0]
skew_array = iterative_skew(input)
utils.print_array(skew_array)
def main():
args = []
for line in fileinput.input():
args.append(line.rstrip())
input = args[0]
pattern = args[1]
count, indices = find_incidence(input, pattern)
utils.print_array(indices)
def main():
args = []
for line in fileinput.input():
elems = line.rstrip().split(' ')
for el in elems:
if el != 'Input':
args.append(el)
input = args[0]
k = int(args[1])
L = int(args[2])
t = int(args[3])
kmers = find_clumps(input, k, L, t)
utils.print_array(kmers)
def main():
args = []
for line in fileinput.input():
elems = line.rstrip().split(' ')
for el in elems:
if el != 'Input':
args.append(el)
input = args[0]
k = int(args[1])
L = int(args[2])
t = int(args[3])
kmers = find_clumps(input, k, L, t);
utils.print_array(kmers)
def simple_plan(self, c_start, c_goal, verbose=True, **kwargs):
"""
Generate a plan in observation space given start and goal states via interpolation.
:param c_start: bs x c_dim
:param c_goal: bs x c_dim
:return: rollout: horizon x bs x channel_dim x img_W x img_H
"""
with torch.no_grad():
rollout = []
_z = Variable(torch.randn(c_start.size()[0],
self.rand_z_dim)).cuda()
for t in range(self.plan_length):
c = c_start + (c_goal - c_start) * t / self.plan_length
c_next = c_start + (c_goal - c_start) * (t +
1) / self.plan_length
# _z = Variable(torch.randn(c.size()[0], self.rand_z_dim)).cuda()
_cur_img, _next_img = self.G(_z, c, c_next)
if t == 0:
rollout.append(_cur_img)
next_img = _next_img
rollout.append(next_img)
if verbose:
# import ipdb; ipdb.set_trace()
print("\t c_%d: %s" % (t, print_array(c[0].data)))
# print("\t Transition var: %s" % print_array(self.T.get_var(c_start[0, None]).data[0]))
# print("\t Direction: %s" % print_array((c_goal-c_start).data[0]/self.planning_horizon))
return rollout
def dump(self):
return print_array('Commands Registered',
commandFormat.format('Name', 'Min Permissions', 'Channel Type', 'Challonge', 'Aliases', 'Required Args', 'Optional Args'),
self._commands,
lambda c: commandFormat.format(c.name,
c.attributes.minPermissions.name,
c.attributes.channelRestrictions.name,
c.attributes.challongeAccess.name,
'-' if len(c.aliases) == 0 else '/'.join(c.aliases),
'-' if len(c.reqParams) == 0 else '/'.join(c.reqParams),
'-' if len(c.optParams) == 0 else '/'.join(c.optParams)))
def dump(self):
return print_array(
'Commands Registered',
commandFormat.format('Name', 'Min Permissions', 'Channel Type',
'Challonge', 'Aliases', 'Required Args',
'Optional Args'), self._commands,
lambda c: commandFormat.format(
c.name, c.attributes.minPermissions.name, c.attributes.
channelRestrictions.name, c.attributes.challongeAccess.name,
'-' if len(c.aliases) == 0 else '/'.join(c.aliases), '-'
if len(c.reqParams) == 0 else '/'.join(c.reqParams), '-'
if len(c.optParams) == 0 else '/'.join(c.optParams)))
def astar_plan(self, c_start, c_goal, verbose=True, **kwargs):
"""
Generate a plan in observation space given start and goal states via A* search.
:param c_start: bs x c_dim
:param c_goal: bs x c_dim
:return: rollout: horizon x bs x channel_dim x img_W x img_H
"""
with torch.no_grad():
rollout = []
# _z = Variable(torch.randn(c_start.size()[0], self.rand_z_dim)).cuda()
bs = c_start.size()[0]
traj = plan_traj_astar(
kwargs['start_obs'],
kwargs['goal_obs'],
start_state=c_start[0].data.cpu().numpy(),
goal_state=c_goal[0].data.cpu().numpy(),
transition_function=self.continuous_transition_function,
preprocess_function=self.preprocess_function,
discriminator_function=self.discriminator_function_np,
generator_function=self.conditional_generator_function)
for t, disc in enumerate(traj[:-1]):
state = undiscretize(disc.state, self.discretization_bins,
self.P.unif_range)
state_next = undiscretize(traj[t + 1].state,
self.discretization_bins,
self.P.unif_range)
c = from_numpy_to_var(state).repeat(bs, 1)
c_next = from_numpy_to_var(state_next).repeat(bs, 1)
_z = Variable(torch.randn(c.size()[0], self.rand_z_dim)).cuda()
_cur_img, _next_img = self.G(_z, c, c_next)
if t == 0:
rollout.append(_cur_img)
next_img = _next_img
rollout.append(next_img)
if verbose:
# import ipdb; ipdb.set_trace()
print("\t c_%d: %s" % (t, print_array(c[0].data)))
return rollout
def main():
args = []
for line in fileinput.input():
args.append(line.rstrip())
input = args[0]
utils.print_array(nucleiotide_incidence(input))
def closest_code(self,
obs,
n_trials,
use_second,
metric,
regress_bs,
verbose=True):
"""
Get the code that generates an image with closest distance to obs.
:param obs: 1 x channel_dim x img_W x img_H
:param n_trials: number of copies to search
:param use_second: bool, to measure distance using the second image
:param metric: str, choose either l2 or D to measure distance
:param regress_bs: int, regression batch size when 0 do just sampling.
:return: the best noise and codes
"""
if metric == 'L2':
f = lambda x, y: ((x - y)**2).view(n_trials, -1).sum(1)
elif metric == 'classifier':
f = lambda x, y: -self.classifier(x, y).view(-1) + (
(x - y)**2).view(n_trials, -1).sum(1) / 10
else:
assert metric == 'D'
# turned max into min using minus.
f = lambda x, y: -self.D(x, y).view(-1)
if regress_bs:
z_var = Variable(0.1 *
torch.randn(n_trials, self.rand_z_dim).cuda(),
requires_grad=True)
c_var = Variable(0.1 * torch.randn(n_trials, self.c_dim).cuda(),
requires_grad=True)
# c_var = Variable(self.Q.forward_soft(self.FE(obs.repeat(n_trials, 1, 1, 1))).data, requires_grad=True)
optimizer = optim.Adam([c_var, z_var], lr=1e-2)
n_iters = 1000
for i in range(n_iters):
optimizer.zero_grad()
if self.planner == self.astar_plan:
c = F.tanh(c_var.repeat(regress_bs, 1))
else:
c = c_var.repeat(regress_bs, 1)
_z = z_var.repeat(regress_bs, 1)
c_next = self.T(c)
o, o_next = self.G(_z, c, c_next)
if use_second:
out = o_next
else:
out = o
dist = f(obs.repeat(n_trials * regress_bs, 1, 1, 1),
out).sum(0) / regress_bs
if i % 100 == 0:
print("\t Closest code (%d/%d): %.3f" % (i, n_iters, dist))
dist.backward()
optimizer.step()
_z = z_var.detach()
if self.planner == self.astar_plan:
c = F.tanh(c_var.detach())
else:
c = c_var.detach()
else:
_z = Variable(torch.randn(n_trials, self.rand_z_dim)).cuda()
c = self.Q.forward_soft(self.FE(obs)).repeat(n_trials, 1)
# Select best c and c_next from different initializations.
if self.planner == self.astar_plan:
c_next = torch.clamp(self.T(c), -1 + 1e-3, 1 - 1e-3)
else:
c_next = self.T(c)
o, o_next = self.G(_z, c, c_next)
if use_second:
out = o_next
else:
out = o
dist = f(obs.repeat(n_trials, 1, 1, 1), out)
min_dist, min_idx = dist.min(0)
if verbose:
# import ipdb; ipdb.set_trace()
print("\t best_c: %s" % print_array(c[min_idx.item()].data))
print("\t best_c_next: %s" %
print_array(c_next[min_idx.item()].data))
print('\t %s measure: %.3f' % (metric, min_dist))
return _z[min_idx].detach(), c[min_idx].detach(
), c_next[min_idx].detach(), out[min_idx].detach()
