def __call__(self, i, reuse=False):
act_fn = get_activation(self.config.get('activation', 'relu'),
self.config.get('activation_params', {}))
norm_fn, norm_params = get_normalization(
self.config.get('batch_norm', 'batch_norm'),
self.config.get('batch_norm_params', self.normalizer_params))
winit_fn = get_weightsinit(
self.config.get('weightsinit', 'normal'),
self.config.get('weightsinit_params', '0.00 0.02'))
filters = self.config.get('nb_filters', 32)
# fully connected parameters
including_top = self.config.get("including top", True)
nb_fc_nodes = self.config.get('nb_fc_nodes', [1024, 1024])
# output stage parameters
output_dims = self.config.get("output dims",
0) # zero for no output layer
output_act_fn = get_activation(
self.config.get('output_activation', 'none'),
self.config.get('output_activation_params', ''))
with tf.variable_scope(self.name):
if reuse:
tf.get_variable_scope().reuse_variables()
else:
assert tf.get_variable_scope().reuse is False
end_points = {}
# x : 299 * 299 * 3
x = tcl.conv2d(i,
filters,
3,
stride=2,
activation_fn=act_fn,
normalizer_fn=norm_fn,
normalizer_params=norm_params,
padding='VALID',
weights_initializer=winit_fn,
scope='conv1')
end_points['conv1'] = x
# x : 149 * 149 * 32
x = tcl.conv2d(x,
filters,
3,
stride=1,
activation_fn=act_fn,
normalizer_fn=norm_fn,
normalizer_params=norm_params,
padding='VALID',
weights_initializer=winit_fn,
scope='conv2')
end_points['conv2'] = x
# x : 147 * 147 * 32
x = tcl.conv2d(x,
2 * filters,
3,
stride=1,
activation_fn=act_fn,
normalizer_fn=norm_fn,
normalizer_params=norm_params,
padding='SAME',
weights_initializer=winit_fn,
scope='conv3')
end_points['conv3'] = x
# x : 147 * 147 * 64
x = tcl.max_pool2d(x, 3, stride=2, padding='VALID', scope='pool1')
end_points['pool1'] = x
# x : 73 * 73 * 64
x = tcl.conv2d(x,
80,
3,
stride=1,
activation_fn=act_fn,
normalizer_fn=norm_fn,
normalizer_params=norm_params,
padding='VALID',
weights_initializer=winit_fn,
scope='conv4')
end_points['conv4'] = x
# x : 71 * 71 * 80
x = tcl.conv2d(x,
192,
3,
stride=2,
activation_fn=act_fn,
normalizer_fn=norm_fn,
normalizer_params=norm_params,
padding='VALID',
weights_initializer=winit_fn,
scope='conv5')
end_points['conv5'] = x
# x : 35 * 35 * 192
x = tcl.conv2d(x,
288,
3,
stride=1,
activation_fn=act_fn,
normalizer_fn=norm_fn,
normalizer_params=norm_params,
padding='SAME',
weights_initializer=winit_fn,
scope='conv6')
end_points['conv6'] = x
# x : 35 * 35 * 288
x, end_points = inception_v3_figure5('inception1a',
x,
end_points,
act_fn=act_fn,
norm_fn=norm_fn,
norm_params=norm_params,
winit_fn=winit_fn)
x, end_points = inception_v3_figure5('inception1b',
x,
end_points,
act_fn=act_fn,
norm_fn=norm_fn,
norm_params=norm_params,
winit_fn=winit_fn)
x, end_points = inception_v3_figure5_downsample(
'inception1c',
x,
end_points,
act_fn=act_fn,
norm_fn=norm_fn,
norm_params=norm_params,
winit_fn=winit_fn)
# x : 17 * 17 * 768
x, end_points = inception_v3_figure6('inception2a',
x,
end_points,
n=7,
act_fn=act_fn,
norm_fn=norm_fn,
norm_params=norm_params,
winit_fn=winit_fn)
x, end_points = inception_v3_figure6('inception2b',
x,
end_points,
n=7,
act_fn=act_fn,
norm_fn=norm_fn,
norm_params=norm_params,
winit_fn=winit_fn)
x, end_points = inception_v3_figure6('inception2c',
x,
end_points,
n=7,
act_fn=act_fn,
norm_fn=norm_fn,
norm_params=norm_params,
winit_fn=winit_fn)
x, end_points = inception_v3_figure6('inception2d',
x,
end_points,
n=7,
act_fn=act_fn,
norm_fn=norm_fn,
norm_params=norm_params,
winit_fn=winit_fn)
x, end_points = inception_v3_figure6_downsample(
'inception2e',
x,
end_points,
n=7,
act_fn=act_fn,
norm_fn=norm_fn,
norm_params=norm_params,
winit_fn=winit_fn)
# x : 8 * 8 * 1280
x, end_points = inception_v3_figure7('inception3a',
x,
end_points,
act_fn=act_fn,
norm_fn=norm_fn,
norm_params=norm_params,
winit_fn=winit_fn)
x, end_points = inception_v3_figure7('inception3b',
x,
end_points,
act_fn=act_fn,
norm_fn=norm_fn,
norm_params=norm_params,
winit_fn=winit_fn)
# construct top fully connected layer
if including_top:
with tf.variable_scope("logits"):
x = tcl.avg_pool2d(x,
kernel_size=[8, 8],
padding="VALID",
scope="avgpool_1a_8x8")
x = tcl.dropout(x, keep_prob=0.5, scope="dropout_1b")
end_points["global_avg_pooling"] = x
x = tcl.flatten(x)
if output_dims != 0:
x = tcl.fully_connected(x,
output_dims,
activation_fn=output_act_fn,
normalizer_fn=None,
weights_initializer=winit_fn,
scope='fc_out')
end_points['fc_out'] = x
return x, end_points
def deconv2d(self,
name,
x,
nb_filters,
ksize,
stride,
*,
norm_fn='none',
norm_params=None,
act_fn='relu',
winit_fn='xavier',
binit_fn='zeros',
padding='SAME',
has_bias=True,
disp=True,
collect_end_points=True):
if callable(act_fn):
act_fn_str = 'func'
act_fn = act_fn
else:
act_fn_str = self.config.get(name + ' activation', act_fn)
act_fn = get_activation(act_fn_str)
if callable(norm_fn):
norm_fn_str = 'func'
norm_fn = norm_fn
else:
norm_fn_str = self.config.get(name + ' normalization', norm_fn)
norm_fn = get_normalization(norm_fn_str)
winit_fn_str = self.config.get(name + ' weightsinit', winit_fn)
if 'special' in winit_fn_str:
split = winit_fn_str.split()
winit_name = split[0]
if winit_name == 'he_uniform':
input_nb_filters = int(x.get_shape()[-1])
fan_in = input_nb_filters * (ksize**2) / (stride**2)
fan_out = nb_filters * (ksize**2)
filters_stdev = np.sqrt(4.0 / (fan_in + fan_out))
winit_fn = self.uniform_initializer(filters_stdev)
else:
raise Exception('Error weights initializer function name : ' +
winit_fn_str)
else:
winit_fn = get_weightsinit(winit_fn_str)
binit_fn_str = self.config.get(name + ' biasesinit', binit_fn)
binit_fn = get_weightsinit(binit_fn_str)
_padding = self.config.get(name + ' padding', padding)
if self.using_tcl_library:
x = tcl.conv2d_transpose(x,
nb_filters,
ksize,
stride=stride,
use_bias=True,
activation_fn=act_fn,
normalizer_fn=norm_fn,
normalizer_params=norm_params,
weights_initializer=winit_fn,
padding=_padding,
scope=name)
else:
x = tl.conv2d_transpose(x,
nb_filters,
ksize,
strides=stride,
padding=_padding,
use_bias=has_bias,
kernel_initializer=winit_fn,
bias_initializer=binit_fn,
trainable=True,
name=name)
with tf.variable_scope(name):
if norm_fn is not None:
norm_params = norm_params or {}
x = norm_fn(x, **norm_params)
if act_fn is not None:
x = act_fn(x)
if disp:
print('\t\tDeonv2D(' + str(name) + ') --> ', x.get_shape(), ' ',
(act_fn_str, norm_fn_str, winit_fn_str, _padding))
if collect_end_points:
self.end_points[name] = x
return x
def fc(self,
name,
x,
nb_nodes,
*,
norm_fn='none',
norm_params=None,
act_fn='none',
winit_fn='xavier',
binit_fn='zeros',
has_bias=True,
disp=True,
collect_end_points=True):
if callable(act_fn):
act_fn_str = 'func'
act_fn = act_fn
else:
act_fn_str = self.config.get(name + ' activation', act_fn)
act_fn = get_activation(act_fn_str)
if callable(norm_fn):
norm_fn_str = 'func'
norm_fn = norm_fn
else:
norm_fn_str = self.config.get(name + ' normalization', norm_fn)
norm_fn = get_normalization(norm_fn_str)
winit_fn_str = self.config.get(name + ' weightsinit', winit_fn)
if 'special' in winit_fn_str:
split = winit_fn_str.split()
winit_name = split[0]
if winit_name == 'glorot_uniform':
input_nb_nodes = int(x.get_shape()[-1])
filters_stdev = np.sqrt(2.0 / (input_nb_nodes + nb_nodes))
winit_fn = self.uniform_initializer(filters_stdev)
else:
raise Exception('Error weights initializer function name : ' +
winit_fn_str)
else:
winit_fn = get_weightsinit(winit_fn_str)
binit_fn_str = self.config.get(name + ' biasesinit', binit_fn)
binit_fn = get_weightsinit(binit_fn_str)
if self.using_tcl_library:
x = tcl.fully_connected(x,
nb_nodes,
activation_fn=act_fn,
normalizer_fn=norm_fn,
normalizer_params=norm_params,
weights_initializer=winit_fn,
scope=name)
else:
x = tl.dense(x,
nb_nodes,
use_bias=has_bias,
kernel_initializer=winit_fn,
bias_initializer=binit_fn,
trainable=True,
name=name)
with tf.variable_scope(name):
if norm_fn is not None:
norm_params = norm_params or {}
x = norm_fn(x, **norm_params)
if act_fn is not None:
x = act_fn(x)
if disp:
print('\t\tFC(' + str(name) + ') --> ', x.get_shape(), ' ',
(act_fn_str, norm_fn_str, winit_fn_str))
if collect_end_points:
self.end_points[name] = x
return x
def __call__(self, i, condition=None):
act_fn = get_activation(self.config.get('activation', 'relu'))
norm_fn, norm_params = get_normalization(
self.config.get('batch_norm', 'batch_norm'),
self.config.get('batch_norm_params', self.normalizer_params))
winit_fn = get_weightsinit(
self.config.get('weightsinit', 'normal 0.00 0.02'))
nb_fc_nodes = self.config.get('nb_fc_nodes', [1024, 1024])
output_dims = self.config.get("output dims", 3)
output_act_fn = get_activation(
self.config.get('output_activation', 'none'))
x, end_points = self.network(i)
x = tcl.flatten(x)
if condition is not None:
x = tf.concatenate([x, condition], axis=-1)
with tf.variable_scope(self.name):
if self.reuse:
tf.get_variable_scope().reuse_variables()
else:
assert tf.get_variable_scope().reuse is False
self.reuse = True
for ind, nb_nodes in enumerate(nb_fc_nodes):
x = tcl.fully_connected(x,
nb_nodes,
activation_fn=act_fn,
normalizer_fn=norm_fn,
normalizer_params=norm_params,
weights_initializer=winit_fn,
scope='fc%d' % ind)
if self.output_distribution == 'gaussian':
mean = tcl.fully_connected(x,
output_dims,
activation_fn=output_act_fn,
weights_initializer=winit_fn,
scope='fc_out_mean')
log_var = tcl.fully_connected(x,
output_dims,
activation_fn=output_act_fn,
weights_initializer=winit_fn,
scope='fc_out_log_var')
return mean, log_var
elif self.output_distribution == 'mean':
mean = tcl.fully_connected(x,
output_dims,
activation_fn=output_act_fn,
weights_initializer=winit_fn,
scope='fc_out_mean')
return mean
elif self.output_distribution == 'none':
out = tcl.fully_connected(x,
output_dims,
activation_fn=output_act_fn,
weights_initializer=winit_fn,
scope='fc_out_mean')
return out
else:
raise Exception("None output distribution named " +
self.output_distribution)