def build_model(data_tensor,
reuse,
training,
output_shape,
data_format='NHWC'):
"""Create the hgru from Learning long-range..."""
if isinstance(output_shape, list):
output_shape = output_shape[-1]
elif isinstance(output_shape, dict):
output_shape = output_shape['output']
# norm_moments_training = training # Force instance norm
# normalization_type = 'no_param_batch_norm_original'
normalization_type = 'no_param_instance_norm'
# output_normalization_type = 'batch_norm_original_renorm'
output_normalization_type = 'instance_norm'
data_tensor, long_data_format = tf_fun.interpret_data_format(
data_tensor=data_tensor, data_format=data_format)
# Prepare gammanet structure
(compression, ff_kernels, ff_repeats, features, fgru_kernels,
additional_readouts) = v2_big_working()
gammanet_constructor = tf_fun.get_gammanet_constructor(
compression=compression,
ff_kernels=ff_kernels,
ff_repeats=ff_repeats,
features=features,
fgru_kernels=fgru_kernels)
aux = get_aux()
# Build model
with tf.variable_scope('vgg', reuse=reuse):
aux = get_aux()
vgg = vgg16.Vgg16(
vgg16_npy_path=
'/media/data_cifs/clicktionary/pretrained_weights/vgg16.npy',
reuse=reuse,
aux=aux,
train=training,
timesteps=8,
fgru_normalization_type=normalization_type,
ff_normalization_type=normalization_type)
vgg(rgb=data_tensor, constructor=gammanet_constructor)
# activity = vgg.fgru_0
with tf.variable_scope('fgru', reuse=reuse):
# Get side weights
h2_rem = [vgg.fgru_0]
for idx, h in enumerate(h2_rem):
res = normalization.apply_normalization(
activity=h,
name='output_norm1_%s' % idx,
normalization_type=output_normalization_type,
data_format=data_format,
training=training,
trainable=training,
reuse=reuse)
res = aux['image_resize'](res,
data_tensor.get_shape().as_list()[1:3],
align_corners=True)
activity = tf.layers.conv2d(res,
filters=res.get_shape().as_list()[-1],
kernel_size=(3, 3),
padding='same',
data_format=long_data_format,
name='out',
activation=tf.nn.relu,
trainable=training,
use_bias=True,
reuse=reuse)
activity = tf.layers.conv2d(
activity,
filters=output_shape,
kernel_size=(1, 1),
padding='same',
data_format=long_data_format,
name='out2',
# activation=tf.nn.relu,
activation=None,
trainable=training,
use_bias=True,
reuse=reuse)
if long_data_format is 'channels_first':
activity = tf.transpose(activity, (0, 2, 3, 1))
extra_activities = {} # idx: v for idx, v in enumerate(hs_0)}
if activity.dtype != tf.float32:
activity = tf.cast(activity, tf.float32)
# return [activity, h_deep], extra_activities
return activity, extra_activities
def build_model(data_tensor,
reuse,
training,
output_shape,
data_format='NHWC'):
"""Create the hgru from Learning long-range..."""
if isinstance(output_shape, list):
output_shape = output_shape[-1]
elif isinstance(output_shape, dict):
output_shape = output_shape['output']
# norm_moments_training = training # Force instance norm
# normalization_type = 'no_param_batch_norm_original'
normalization_type = 'no_param_instance_norm'
# output_normalization_type = 'batch_norm_original_renorm'
output_normalization_type = 'instance_norm'
data_tensor, long_data_format = tf_fun.interpret_data_format(
data_tensor=data_tensor, data_format=data_format)
# Prepare gammanet structure
(compression, ff_kernels, ff_repeats, features, fgru_kernels,
additional_readouts) = v2_big_working()
gammanet_constructor = tf_fun.get_gammanet_constructor(
compression=compression,
ff_kernels=ff_kernels,
ff_repeats=ff_repeats,
features=features,
fgru_kernels=fgru_kernels)
aux = get_aux()
# Build model
with tf.variable_scope('vgg', reuse=reuse):
aux = get_aux()
vgg = vgg16.Vgg16(
vgg16_npy_path=
'/media/data_cifs/clicktionary/pretrained_weights/vgg16.npy',
reuse=reuse,
aux=aux,
train=training,
timesteps=8,
fgru_normalization_type=normalization_type,
ff_normalization_type=normalization_type)
vgg(rgb=data_tensor, constructor=gammanet_constructor)
# activity = vgg.fgru_0
with tf.variable_scope('fgru', reuse=reuse):
# Get side weights
hs_0, hs_1 = [], []
h2_rem = [vgg.fgru_0, vgg.fgru_1, vgg.fgru_2, vgg.conv5_1, vgg.fgru_3]
for idx, h in enumerate(h2_rem):
res = normalization.apply_normalization(
activity=h,
name='output_norm1_%s' % idx,
normalization_type=output_normalization_type,
data_format=data_format,
training=training,
trainable=training,
reuse=reuse)
res = aux['image_resize'](res,
data_tensor.get_shape().as_list()[1:3],
align_corners=True)
cnv_0 = tf.layers.conv2d(
inputs=res,
filters=output_shape,
kernel_size=(1, 1),
padding='same',
data_format=long_data_format,
name='readout_aux_00_%s' % idx,
# activation=tf.nn.relu,
activation=None,
trainable=training,
use_bias=True,
reuse=reuse)
cnv_1 = tf.layers.conv2d(
inputs=res,
filters=output_shape,
kernel_size=(1, 1),
padding='same',
data_format=long_data_format,
name='readout_aux_10_%s' % idx,
# activation=tf.nn.relu,
activation=None,
trainable=training,
use_bias=True,
reuse=reuse)
hs_0 += [cnv_0]
hs_1 += [cnv_1]
s1, s2, s3, s4, s5 = hs_0
s11, s21, s31, s41, s51 = hs_1
o1 = tf.stop_gradient(tf.identity(s1))
o2 = tf.stop_gradient(tf.identity(s2))
o3 = tf.stop_gradient(tf.identity(s3))
o4 = tf.stop_gradient(tf.identity(s4))
o21 = tf.stop_gradient(tf.identity(s21))
o31 = tf.stop_gradient(tf.identity(s31))
o41 = tf.stop_gradient(tf.identity(s41))
o51 = tf.stop_gradient(tf.identity(s51))
p1_1 = s1
p2_1 = s2 + o1
p3_1 = s3 + o2 + o1
p4_1 = s4 + o3 + o2 + o1
p5_1 = s5 + o4 + o3 + o2 + o1
p1_2 = s11 + o21 + o31 + o41 + o51
p2_2 = s21 + o31 + o41 + o51
p3_2 = s31 + o41 + o51
p4_2 = s41 + o51
p5_2 = s51
hs = [p1_1, p2_1, p3_1, p4_1, p5_1, p1_2, p2_2, p3_2, p4_2, p5_2]
activity = tf.layers.conv2d(
tf.concat(hs, -1),
filters=output_shape,
kernel_size=(1, 1),
padding='same',
data_format=long_data_format,
name='out',
# activation=tf.nn.relu,
activation=None,
trainable=training,
use_bias=False,
reuse=reuse)
if long_data_format is 'channels_first':
activity = tf.transpose(activity, (0, 2, 3, 1))
extra_activities = {idx: v for idx, v in enumerate(hs_0)}
if activity.dtype != tf.float32:
activity = tf.cast(activity, tf.float32)
# return [activity, h_deep], extra_activities
return activity, extra_activities