def _select_evo(name):
nd.evo = None
for evo in _evolution_manager.evolutions():
nd.log('check', evo.name(), name)
if evo.name() == name:
nd.log('<<< determined current evolution: %s >>>' % name)
nd.evo = evo
if nd.evo is None:
raise BaseException('evolution %s not found' % name)
def add_evolution(self, evolution):
idx = len(self.evolutions())
evolution.set_index(idx)
if evolution.name() is None:
evolution.set_name('%02d__%s__%s' %
(idx, evolution.training_dataset(),
evolution.schedule().name()))
evolution.make_folder()
self._evolutions.append(evolution)
ds_name = evolution.training_dataset()
if isinstance(evolution.training_dataset(), dict):
ds_name = ''
for key in evolution.training_dataset():
if ds_name != '': ds_name += '_'
ds_name += key
nd.log('evolution <%s>: %s' % (evolution.name(), str(evolution)))
def get(self, full_name):
for member in self._members:
member_name = self._member_to_str(member)
nd.log("mem: {} ".format(member_name))
if self.is_data(member) and member_name == full_name:
return self[member]
elif self.is_struct(member) and full_name.startswith(
member_name) and (full_name[len(member_name)]
in ['.', '[']):
sub_struct = self[member]
sub_name = full_name[len(member_name):]
if sub_name.startswith('.'): sub_name = sub_name[1:]
return sub_struct.get(sub_name)
elif member_name == full_name:
return self[member]
raise KeyError
def make_graph(self, data, include_losses=False):
# hypNet
pred_config = nd.PredConfig()
pred_config.add(nd.PredConfigId(type='flow_hyp', dir='fwd', offset=0, channels=2, scale=self._scale, array_length=self._num_hypotheses))
pred_config.add(nd.PredConfigId(type='iul_b_hyp_log', dir='fwd', offset=0, channels=2, scale=self._scale, array_length=self._num_hypotheses, mod_func=self._log_sigmoid))
nd.log('pred_config:')
nd.log(pred_config)
with nd.Scope('hypNet', shared_batchnorm=False, correlation_leaky_relu=True, **self.scope_args()):
arch = Architecture_C(
num_outputs=pred_config.total_channels(),
disassembling_function=pred_config.disassemble,
conv_upsample=True,
loss_function= None,
channel_factor=self._channel_factor,
feature_channels=self._feature_channels
)
out_hyp = arch.make_graph(data.img[0], data.img[1])
# mergeNet
pred_config = nd.PredConfig()
pred_config.add(nd.PredConfigId(type='flow', dir='fwd', offset=0, channels=2, scale=self._scale, dist=1))
pred_config.add(nd.PredConfigId(type='iul_b_log', dir='fwd', offset=0, channels=2, scale=self._scale, dist=1, mod_func=self.iul_b_log_sigmoid))
nd.log('pred_config:')
nd.log(pred_config)
hyps = [nd.ops.resample(hyp, reference=data.img[0], antialias=False, type='LINEAR') for hyp in [out_hyp.final.flow_hyp[0].fwd[i] for i in range(self._num_hypotheses)]]
uncertainties = [nd.ops.resample(unc, reference=data.img[0], antialias=False, type='LINEAR') for unc in [out_hyp.final.iul_b_hyp_log[0].fwd[i] for i in range(self._num_hypotheses)]]
img_warped = [nd.ops.warp(data.img[1], hyp) for hyp in hyps]
with nd.Scope('mergeNet', shared_batchnorm=False, **self.scope_args()):
input = nd.ops.concat([data.img[0]] + [data.img[1]] + hyps + uncertainties + img_warped)
arch = Architecture_S(
num_outputs=pred_config.total_channels(),
disassembling_function=pred_config.disassemble,
conv_upsample=True,
loss_function= None,
channel_factor=self._channel_factor
)
out_merge = arch.make_graph(input)
return out_merge
def make_graph(self, data, net_graph_constructor, divisor=64., scale=1.0):
self._scale = scale
self._width = data.width
self._height = data.height
self._temp_width = ceil(self._width*self._scale/divisor) * divisor
self._temp_height = ceil(self._height*self._scale/divisor) * divisor
self._rescale_coeff_x = self._width / self._temp_width
self._rescale_coeff_y = self._height / self._temp_height
nd.log('data:')
nd.log(data)
input = nd.Struct()
data.map('img', self.input_image_resample, input)
data.map('flow', self.input_resample_flow, input)
data.map('disp', self.input_resample_disp, input)
data.map('occ', self.input_resample_binary, input)
data.map('mb', self.input_resample_binary, input)
data.map('db', self.input_resample_binary, input)
pred = net_graph_constructor(input)
nd.log('pred:')
nd.log(pred)
if isinstance(pred, (tuple)):
pred = pred[0]
if isinstance(pred, (list)):
output = []
for prediction in pred:
out = nd.Struct()
self.map_output(prediction.final, out)
output.append(out)
else:
output = nd.Struct()
self.map_output(pred.final, output)
nd.log('output:')
nd.log(output)
return output
def make_graph(self, data, include_losses=True):
pred_config = nd.PredConfig()
pred_config.add(
nd.PredConfigId(type='flow',
dir='fwd',
offset=0,
channels=2,
scale=self._scale))
pred_config.add(
nd.PredConfigId(type='occ',
dir='fwd',
offset=0,
channels=2,
scale=self._scale))
nd.log('pred_config:')
nd.log(pred_config)
#### Net 1 ####
with nd.Scope('net1', learn=False, **self.scope_args()):
arch1 = Architecture_C(
num_outputs=pred_config.total_channels(),
disassembling_function=pred_config.disassemble,
loss_function=None,
conv_upsample=self._conv_upsample)
out1 = arch1.make_graph(data.img[0], data.img[1])
#### Net 2 ####
flow_fwd = out1.final.flow[0].fwd
upsampled_flow_fwd = nd.ops.differentiable_resample(
flow_fwd, reference=data.img[0])
warped = nd.ops.warp(data.img[1], upsampled_flow_fwd)
# prepare data for second net
occ_fwd = self.resample_occ(out1.final.occ[0].fwd, data.img[0])
input2 = nd.ops.concat(data.img[0], data.img[1],
nd.ops.scale(upsampled_flow_fwd, 0.05), warped,
occ_fwd)
pred_config[0].mod_func = lambda x: nd.ops.add(
x,
nd.ops.resample(
flow_fwd, reference=x, type='LINEAR', antialias=False))
pred_config[1].mod_func = lambda x: nd.ops.add(
x,
nd.ops.resample(
occ_fwd, reference=x, type='LINEAR', antialias=False))
with nd.Scope('net2', learn=False, **self.scope_args()):
arch2 = Architecture_S(
num_outputs=pred_config.total_channels(),
disassembling_function=pred_config.disassemble,
loss_function=None,
conv_upsample=self._conv_upsample)
out2 = arch2.make_graph(input2)
#### Net 3 ####
flow_fwd = out2.final.flow[0].fwd
upsampled_flow_fwd = nd.ops.differentiable_resample(
flow_fwd, reference=data.img[0])
warped = nd.ops.warp(data.img[1], upsampled_flow_fwd)
# prepare data for third net
occ_fwd = self.resample_occ(out2.final.occ[0].fwd, data.img[0])
input3 = nd.ops.concat(data.img[0], data.img[1],
nd.ops.scale(upsampled_flow_fwd, 0.05), warped,
occ_fwd)
pred_config.add(
nd.PredConfigId(type='mb',
dir='fwd',
offset=0,
channels=2,
scale=self._scale))
pred_config[0].mod_func = lambda x: nd.ops.add(
x,
nd.ops.resample(
flow_fwd, reference=x, type='LINEAR', antialias=False))
pred_config[1].mod_func = lambda x: nd.ops.add(
x,
nd.ops.resample(
occ_fwd, reference=x, type='LINEAR', antialias=False))
with nd.Scope('net3', learn=True, **self.scope_args()):
arch3 = Architecture_S(
num_outputs=pred_config.total_channels(),
disassembling_function=pred_config.disassemble,
loss_function=None,
conv_upsample=self._conv_upsample,
exit_after=0,
)
out3 = arch3.make_graph(input3, edge_features=data.img[0])
return out3
