def new_episode(self):
self._make_sure_inited()
# initialize the state
rgb = self.images.next()
# TODO: reshape
self.y = UncertainImage(rgb)
id_environment = self.id_image_stream
id_episode = unique_timestamp_string()
self.timestamp = time.time()
self.commands = np.array([0], dtype='int')
self.commands_source = 'rest'
return EpisodeDesc(id_episode, id_environment, extra=None)
def obs2ui(self, y):
""" Converts observatiosn from outside in an UncertainImage
of the right dimensions to deal with this DDS. """
shape = self.discdds.get_shape()
y = (y * 255).astype('uint8')
y = resize(y, width=shape[1], height=shape[0])
ui = UncertainImage(y)
return ui
def update(self):
rgb0 = self.input.rgb
H, W = self.discdds.get_shape()
rgb = resize(rgb0, width=W, height=H)
y0 = UncertainImage(rgb)
y1 = self.action.predict(y0)
pred = y1.get_rgba_fill()
pred2 = resize(pred, height=rgb0.shape[0], width=rgb0.shape[1])
self.output.prediction = pred2[:, :, :3]
def make_logcases(id_stream, n, delta, id_tcs, id_discdds, discdds, seed):
stream = get_conftools_streams().instance(id_stream)
stream_data = stream.read_all()
miniplans = iterate_testcases(stream_data, delta)
sampled = reservoir_sample(miniplans, N=n, seed=seed)
for i, m in enumerate(sampled):
id_tc = id_tcs[i]
I0 = UncertainImage(m.y0)
I1 = UncertainImage(m.y1)
assert len(m.u) == delta
plan = discdds.commands_to_indices(m.u)
tc = TestCase(id_tc=id_tc,
id_discdds=id_discdds,
y0=I0,
y1=I1,
true_plan=plan)
# TODO: add field desc
# desc = 'Sampled from stream %s, i=%s for seed %s.' % (i, seed)
yield tc
def test_distance_neighbor_dist2_repeated():
# A constant pattern
a = np.ones((100, 100, 3))
# Let's get some translated slices
w = 80
N = 10
A = []
for i in range(N):
A.append(a[i:i + w, 0:0 + w, ...])
print()
areas = [0.01, 0.03, 0.05, 0.07]
Ds = [DistanceNeighborDist((a, a)) for a in areas]
L2 = DistanceNorm(2)
for i in range(N):
y0 = UncertainImage(A[0])
y1 = UncertainImage(A[i])
vl2 = L2.distance(y0, y1)
ds = [x.distance(y0, y1) for x in Ds]
assert_allclose(ds, 0) # < this must be zero
s = ";".join("%1.5f" % x for x in ds)
print('- step %d L2: %1.5f %s' % (i, vl2, s))
def test_distance_neighbor_dist2_rgb():
# A random pattern
a = np.random.rand(100, 100, 3)
# Let's get some translated slices
w = 80
N = 10
A = []
for i in range(N):
A.append(a[i:i + w, 0:0 + w, ...])
print()
areas = [0.01, 0.03, 0.05, 0.07]
Ds = [DistanceNeighborDist((a, a)) for a in areas]
L2 = DistanceNorm(2)
for i in range(N):
y0 = UncertainImage(A[0])
y1 = UncertainImage(A[i])
vl2 = L2.distance(y0, y1)
ds = [x.distance(y0, y1) for x in Ds]
s = ";".join("%1.5f" % x for x in ds)
print('- step %d L2: %1.5f %s' % (i, vl2, s))
def compute_predstats(id_discdds, id_stream, delta, id_distances):
dds = get_conftools_discdds().instance(id_discdds)
stream = get_conftools_streams().instance(id_stream)
distances_library = get_conftools_uncertain_image_distances()
distances = dict(
map(lambda x: (x, distances_library.instance(x)), id_distances))
dtype = [(x, 'float32') for x in id_distances]
results = []
for logitem in iterate_testcases(stream.read_all(), delta):
assert_allclose(len(logitem.u), delta)
y0 = UncertainImage(logitem.y0)
y1 = UncertainImage(logitem.y1)
py0 = dds.predict(y0, dds.commands_to_indices(logitem.u))
ds = []
for name in id_distances:
d = distances[name].distance(y1, py0)
# d0 = distances[name].distance(y1, y0)
ds.append(d)
a = np.array(tuple(ds), dtype=dtype)
results.append(a)
return results
def discdds_make_test_cases(config, outdir, id_discdds, id_image, num_cases,
plan_length):
# Get the DiffeoSystem
discdds = config.discdds.instance(id_discdds)
# Get the RGB image
rgb = config.images.instance(id_image)
shape = discdds.get_shape()
rgb = resize(rgb, shape[1], shape[0])
assert rgb.shape[0] == shape[0]
assert rgb.shape[1] == shape[1]
image = UncertainImage(rgb)
for i in range(num_cases):
tcname = 'tc_%s_%s_d%d_%03d' % (id_discdds, id_image, plan_length, i)
discdds_make_test_case(outdir, tcname, id_discdds, discdds, image,
plan_length)
def update(self):
rgb0 = self.input.rgb
pad = self.config.pad
H, W = self.discdds.get_shape()
rgb = resize(rgb0, width=W, height=H)
y0 = UncertainImage(rgb)
uimages = [a.predict(y0) for a in self.actions]
images = [x.get_rgba_fill()[..., :3] for x in uimages]
images = [
resize(x, height=rgb0.shape[0], width=rgb0.shape[1])
for x in images
]
canvas = make_images_grid(images, pad=pad)
self.output.prediction = canvas
def go(self):
outdir = self.options.output
which = self.options.get_extra()
diffeo2dds_config = get_diffeo2dds_config()
if not which:
todo = diffeo2dds_config.discdds.keys()
else:
todo = diffeo2dds_config.discdds.expand_names(which)
id_image = self.options.image
image = UncertainImage(diffeo2dds_config.images.instance(id_image))
for id_dds in todo:
self.info('Writing %s' % id_dds)
dds = diffeo2dds_config.discdds.instance(id_dds)
report = Report(id_dds)
dds.display(report, image=image)
write_report_files(report, basename=os.path.join(outdir, id_dds))
class DiffeoSystemSimulation(RobotInterface):
"""
Simulation of a DiffeoSystem as a RobotInterface.
"""
@contract(image_stream='str|code_spec',
discdds='str|code_spec')
def __init__(self, image_stream, discdds):
'''
:param discdds: Which DDS dynamics to use.
:param image_stream: What ImageStream to use.
Each image is used to start a new episode.
'''
diffeo2ddslearn_config = get_diffeo2ddslearn_config()
self.id_image_stream, self.image_stream = \
diffeo2ddslearn_config.image_streams.instance_smarter(image_stream)
if self.id_image_stream is None:
self.id_image_stream = 'image_stream'
diffeo2dds_config = get_diffeo2dds_config()
self.id_discdds, self.discdds = \
diffeo2dds_config.discdds.instance_smarter(discdds)
if self.id_discdds is None:
self.id_discdds = 'discdds'
self.shape = None
def _make_sure_inited(self):
# don't do this in the constructor, because we cannot pickle
# generators
if self.shape is not None:
return
# get the first image, look how it is to get shape
self.images = self.image_stream.read_all()
x0 = self.images.next()
self.shape = x0.shape[:2]
@contract(returns=BootSpec)
def get_spec(self):
self._make_sure_inited()
ncmds = self.discdds.get_num_commands()
cmd = make_streamels_finite_commands(ncommands=ncmds, default=0)
cmd_spec = StreamSpec(id_stream=None, streamels=cmd, extra={})
obs_spec = StreamSpec(id_stream=None,
streamels=make_streamels_rgb_float(self.shape),
extra={})
return BootSpec(obs_spec, cmd_spec)
@contract(returns=EpisodeDesc)
def new_episode(self):
self._make_sure_inited()
# initialize the state
rgb = self.images.next()
# TODO: reshape
self.y = UncertainImage(rgb)
id_environment = self.id_image_stream
id_episode = unique_timestamp_string()
self.timestamp = time.time()
self.commands = np.array([0], dtype='int')
self.commands_source = 'rest'
return EpisodeDesc(id_episode, id_environment, extra=None)
@contract(commands='array', commands_source='str')
def set_commands(self, commands, commands_source): # @UnusedVariable
u = int(commands[0])
warnings.warn('understand what was going on')
self.y = self.discdds.predict(self.y, plan=[u])
self.commands = commands
self.commands_source = commands_source
self.timestamp += 1
@contract(returns=RobotObservations)
def get_observations(self):
rgb = self.y.get_values()
timestamp = self.timestamp
observations = rgb
commands = self.commands
commands_source = self.commands_source
episode_end = False
robot_pose = None
obs = RobotObservations(timestamp, observations,
commands, commands_source,
episode_end, robot_pose)
return obs