def build_model(data_tensor, reuse, training, output_shape):
"""Create the hgru from Learning long-range..."""
if isinstance(output_shape, list):
output_shape = output_shape[0]
with tf.variable_scope('cnn', reuse=reuse):
# Add input
so_filters = np.load(
os.path.join('weights', 'so_filters.npy')).squeeze().reshape(
11, 11, 3, 8 * 4)[..., 0:32:4]
so_filter_tensor = tf.get_variable(
name='so_filters',
initializer=so_filters,
trainable=training)
so_bias = tf.get_variable(
name='so_bias',
initializer=tf.zeros(so_filters.shape[-1]),
trainable=training)
in_emb = tf.nn.conv2d(
input=data_tensor,
filter=so_filter_tensor,
strides=[1, 1, 1, 1],
padding='SAME',
name='so')
in_emb = tf.nn.bias_add(in_emb, so_bias)
in_emb = in_emb ** 2
layer_hgru = hgru.hGRU(
'hgru_1',
x_shape=in_emb.get_shape().as_list(),
timesteps=12,
h_ext=1,
crf=15,
strides=[1, 1, 1, 1],
padding='SAME',
aux={
'reuse': False,
'constrain': False,
'nonnegative': True,
'while_loop': False,
'hidden_states': 'gru',
'horizontal_dilations': [1, 1, 1, 1]
},
train=training)
h2 = layer_hgru.build(in_emb)
h2 = normalization.batch(
bottom=h2,
renorm=False,
name='hgru_bn',
training=training)
activity = conv.readout_layer(
activity=h2,
reuse=reuse,
training=training,
pool_type='max', # 'select',
output_shape=output_shape,
features=so_filters.shape[-1])
extra_activities = {
'activity': h2
}
return activity, extra_activities
def build_model(data_tensor, reuse, training, output_shape):
"""Create the hgru from Learning long-range..."""
if isinstance(output_shape, list):
output_shape = output_shape[0]
with tf.variable_scope('cnn', reuse=reuse):
# Add input
in_emb = conv.input_layer(
X=data_tensor,
reuse=reuse,
training=training,
features=20,
conv_activation=[tf.nn.relu, None],
conv_kernel_size=7,
pool_kernel_size=[1, 2, 2, 1],
pool_kernel_strides=[1, 2, 2, 1],
pool=True,
name='l0')
layer_hgru = hgru.hGRU(
'fgru',
x_shape=in_emb.get_shape().as_list(),
timesteps=1,
h_ext=[{'h1': [15, 15]}, {'h2': [1, 1]}, {'fb1': [1, 1]}],
strides=[1, 1, 1, 1],
hgru_ids=[{'h1': 20}, {'h2': 128}, {'fb1': 20}],
hgru_idx=[{'h1': 0}, {'h2': 1}, {'fb1': 2}],
padding='SAME',
aux={
'readout': 'fb',
'intermediate_ff': [32, 128],
'intermediate_ks': [[3, 3], [3, 3]],
'intermediate_repeats': [3, 3],
'while_loop': False,
'skip': False,
'force_horizontal': True,
'symmetric_weights': True,
'include_pooling': True
},
pool_strides=[2, 2],
pooling_kernel=[2, 2],
train=training)
h2 = layer_hgru.build(in_emb)
h2 = normalization.batch(
bottom=h2,
renorm=False,
name='hgru_bn',
training=training)
activity = conv.readout_layer(
activity=h2,
reuse=reuse,
training=training,
output_shape=output_shape)
extra_activities = {
'activity': h2
}
return activity, extra_activities
def build_model(data_tensor, reuse, training, output_shape):
"""Create the hgru from Learning long-range..."""
if isinstance(output_shape, list):
output_shape = output_shape[0]
with tf.variable_scope('cnn', reuse=reuse):
# Add input
so_filters = np.load(
'/media/data_cifs/clicktionary/pretrained_weights/alexnet.npy'
).item()
so_filter_tensor = tf.get_variable(name='so_filters',
initializer=so_filters['conv1'][0],
trainable=training)
so_bias = tf.get_variable(name='so_bias',
initializer=so_filters['conv1'][1],
trainable=training)
in_emb = tf.nn.conv2d(input=data_tensor,
filter=so_filter_tensor,
strides=[1, 1, 1, 1],
padding='SAME',
name='so')
in_emb = tf.nn.bias_add(in_emb, so_bias)
# in_emb = in_emb ** 2
layer_hgru = hgru.hGRU('hgru_1',
x_shape=in_emb.get_shape().as_list(),
timesteps=8,
h_ext=11,
strides=[1, 1, 1, 1],
padding='SAME',
aux={
'reuse': False,
'constrain': False,
'nonnegative': True,
'while_loop': False,
'horizontal_dilations': [1, 2, 2, 1]
},
train=training)
h2 = layer_hgru.build(in_emb)
h2 = normalization.batch(bottom=h2,
renorm=False,
name='hgru_bn',
training=training)
activity = conv.readout_layer(
activity=h2,
reuse=reuse,
training=training,
pool_type='max', # 'select',
output_shape=output_shape,
features=output_shape)
extra_activities = {'activity': h2}
return activity, extra_activities
def build_model(data_tensor, reuse, training, output_shape):
"""Create the hgru from Learning long-range..."""
if isinstance(output_shape, list):
output_shape = output_shape[0]
alpha = 5.
beta = 1.
bs = data_tensor.get_shape().as_list()[0]
dist = tf.distributions.Beta(alpha, beta)
rand = dist.sample([])
sel = tf.cast(tf.round(tf.random_uniform([], minval=1, maxval=bs - 1)),
tf.int32)
with tf.variable_scope('cnn', reuse=reuse):
with tf.variable_scope('input', reuse=reuse):
o_x = tf.layers.conv2d(inputs=data_tensor,
filters=24,
kernel_size=11,
name='l0',
strides=(1, 1),
padding='same',
activation=tf.nn.relu,
trainable=training,
use_bias=True)
x = rand * o_x + ((1 - rand) * tf.roll(o_x, sel, 0))
layer_hgru = hgru.hGRU('hgru_1',
x_shape=x.get_shape().as_list(),
timesteps=8,
h_ext=15,
strides=[1, 1, 1, 1],
padding='SAME',
aux={
'reuse': False,
'constrain': False
},
train=training)
h2, inh = layer_hgru.build(x)
with tf.variable_scope('ro', reuse=reuse):
h2 = normalization.batch(bottom=h2,
renorm=False,
name='hgru_bn',
training=training)
activity = conv.readout_layer(activity=h2,
reuse=reuse,
training=training,
output_shape=output_shape)
extra_activities = {'activity': h2}
return activity, extra_activities
def build_model(data_tensor, reuse, training, output_shape):
"""Create the hgru from Learning long-range..."""
if isinstance(output_shape, list):
output_shape = output_shape[0]
reduction_filters = 16
with tf.variable_scope('cnn', reuse=reuse):
# Add input
so_filters = np.load(os.path.join('weights',
'so_filters.npy')).squeeze().reshape(
11, 11, 3, 8 * 4)
in_emb = tf.nn.conv2d(input=data_tensor,
filter=so_filters,
strides=[1, 1, 1, 1],
padding='SAME',
name='so')
in_emb = tf.layers.conv2d(inputs=in_emb,
filters=reduction_filters,
kernel_size=(1, 1))
in_emb = in_emb**2
layer_hgru = hgru.hGRU('hgru_1',
x_shape=in_emb.get_shape().as_list(),
timesteps=8,
h_ext=11,
strides=[1, 1, 1, 1],
padding='SAME',
aux={
'reuse': False,
'constrain': False,
'nonnegative': True,
'while_loop': False
},
train=training)
h2 = layer_hgru.build(in_emb)
h2 = normalization.batch(bottom=h2,
renorm=False,
name='hgru_bn',
training=training)
activity = conv.readout_layer(
activity=h2,
reuse=reuse,
training=training,
pool_type='max', # 'select',
output_shape=output_shape,
features=reduction_filters)
extra_activities = {'activity': h2}
return activity, extra_activities
def build_model(data_tensor, reuse, training, output_shape):
"""Create the hgru from Learning long-range..."""
if isinstance(output_shape, list):
output_shape = output_shape[0]
with tf.variable_scope('cnn', reuse=reuse):
# Add input
in_emb = conv.input_layer(
X=data_tensor,
reuse=reuse,
training=training,
features=20,
conv_activation=[tf.nn.relu, None],
conv_kernel_size=7,
pool_kernel_size=[1, 2, 2, 1],
pool_kernel_strides=[1, 2, 2, 1],
pool=True,
name='l0')
layer_hgru = hgru.hGRU(
'hgru_1',
x_shape=in_emb.get_shape().as_list(),
timesteps=8,
h_ext=15,
strides=[1, 1, 1, 1],
padding='SAME',
aux={'reuse': False, 'constrain': False},
train=training)
h2 = layer_hgru.build(in_emb)
h2 = normalization.batch(
bottom=h2,
renorm=False,
name='hgru_bn',
training=training)
activity = conv.readout_layer(
activity=h2,
reuse=reuse,
training=training,
output_shape=output_shape)
extra_activities = {
'activity': h2
}
return activity, extra_activities
def build_model(data_tensor, reuse, training, output_shape):
"""Create the hgru from Learning long-range..."""
if isinstance(output_shape, list):
output_shape = output_shape[0]
with tf.variable_scope('cnn', reuse=reuse):
# Add input
conv_aux = {
'pretrained': os.path.join('weights',
'gabors_for_contours_11.npy'),
'pretrained_key': 's1',
'nonlinearity': 'square'
}
activity = conv.conv_layer(bottom=data_tensor,
name='gabor_input',
stride=[1, 1, 1, 1],
padding='SAME',
trainable=training,
use_bias=True,
aux=conv_aux)
layer_hgru = hgru.hGRU('hgru_1',
x_shape=activity.get_shape().as_list(),
timesteps=8,
h_ext=15,
strides=[1, 1, 1, 1],
padding='SAME',
aux={
'reuse': False,
'constrain': False
},
train=training)
h2 = layer_hgru.build(activity)
h2 = normalization.batch(bottom=h2,
renorm=False,
name='hgru_bn',
training=training)
activity = conv.readout_layer(activity=h2,
reuse=reuse,
training=training,
output_shape=output_shape)
extra_activities = {'activity': h2}
return activity, extra_activities
def build_model(data_tensor, reuse, training, output_shape):
"""Create the hgru from Learning long-range..."""
if isinstance(output_shape, list):
output_shape = output_shape[0]
elif isinstance(output_shape, dict):
output_shape = output_shape['output']
with tf.variable_scope('cnn', reuse=reuse):
# # Concatenate standard deviation
# _, var = tf.nn.moments(data_tensor, axes=[3])
# std = tf.expand_dims(tf.sqrt(var), axis=-1)
# data_tensor = tf.concat([data_tensor, std], axis=-1)
# Add input
in_emb = conv.input_layer(X=data_tensor,
reuse=reuse,
training=training,
features=[12, 12],
conv_activation=[tf.nn.relu, tf.nn.relu],
conv_kernel_size=7,
pool_kernel_size=[1, 4, 4, 1],
pool_kernel_strides=[1, 2, 2, 1],
pool=True,
name='l0')
# Run fGRU
hgru_kernels = OrderedDict()
hgru_kernels['h1'] = [15, 15] # height/width
hgru_kernels['h2'] = [1, 1]
hgru_kernels['fb1'] = [1, 1]
hgru_features = OrderedDict()
hgru_features['h1'] = [12, 12] # Fan-in/fan-out, I and E (match fb1)
hgru_features['h2'] = [48, 48]
hgru_features['fb1'] = [12, 12] # (match h1)
# hgru_features['fb1'] = [24, 12] # (match h1 unless squeeze_fb)
intermediate_ff = [24, 48] # Last feature must match h2
intermediate_ks = [[3, 3], [3, 3]]
intermediate_repeats = [1, 1] # Repeat each interm this many times
layer_hgru = hgru.hGRU(
'fgru',
x_shape=in_emb.get_shape().as_list(),
timesteps=8,
strides=[1, 1, 1, 1],
hgru_features=hgru_features,
hgru_kernels=hgru_kernels,
intermediate_ff=intermediate_ff,
intermediate_ks=intermediate_ks,
intermediate_repeats=intermediate_repeats,
padding='SAME',
aux={
'readout': 'fb',
'squeeze_fb': False, # Compress Inh-hat with a 1x1 conv
'td_gate': True, # Add top-down activity to the in-gate
'attention': 'gala',
'attention_layers': 2,
'td_cell_state': False,
'excite_se': False, # Add S/E in the excitation stage
'residual': True, # intermediate resid connections
'while_loop': False,
'skip': True,
'time_skips': False,
'symmetric_weights': False,
'timestep_output': False,
'include_pooling': True
},
pool_strides=[2, 2],
pooling_kernel=[2, 2],
train=training)
h2 = layer_hgru.build(in_emb)
if isinstance(h2, list):
all_activities = []
for idx, it_h2 in enumerate(h2):
it_h2 = normalization.batch(bottom=it_h2,
renorm=False,
name='hgru_bn_%s' % idx,
training=training)
activity = conv.readout_layer(activity=it_h2,
reuse=reuse,
training=training,
var_scope='readout_%s' % idx,
output_shape=output_shape)
all_activities += [activity]
h2 = all_activities[-1]
activity = all_activities
else:
h2 = normalization.batch(bottom=h2,
renorm=False,
name='hgru_bn',
training=training)
activity = conv.readout_layer(activity=h2,
reuse=reuse,
training=training,
output_shape=output_shape)
extra_activities = {'activity': h2}
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']
output_shape = 1
# 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 = tf.concat((data_tensor, data_tensor, data_tensor), -1) # Expand to 3D
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=False, # 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, vgg.fgru_1, vgg.fgru_2]
activities = []
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=False,
# trainable=False,
# reuse=reuse)
activities.append(aux['image_resize'](
h, vgg.fgru_0.get_shape().as_list()[1:3], align_corners=True))
activity = tf.concat(activities, -1)
# Then read out
# activity = normalization.apply_normalization(
# activity=res,
# name='output_norm1',
# normalization_type=output_normalization_type,
# data_format=data_format,
# training=training,
# trainable=training,
# reuse=reuse)
with tf.variable_scope('readout', reuse=reuse):
activity = normalization.batch(bottom=activity,
renorm=False,
name='hgru_bn',
training=training)
activity = conv.readout_layer(activity=activity,
reuse=reuse,
training=training,
output_shape=output_shape)
if long_data_format is 'channels_first':
activity = tf.transpose(activity, (0, 2, 3, 1))
# extra_activities = {'orientations': orientations, 'model_output_y': vgg.fgru_0, 'model_output_x': vgg.fgru_0_init}
extra_activities = {}
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):
"""Create the hgru from Learning long-range..."""
if isinstance(output_shape, list):
output_shape = output_shape[0]
with tf.variable_scope('cnn', reuse=reuse):
# Unclear if we should include l0 in the down/upsample cascade
with tf.variable_scope('g1', reuse=reuse):
# Downsample
act11 = conv_block(x=data_tensor,
name='l1_1',
filters=64,
training=training,
reuse=reuse,
pool=False)
act12 = conv_block(x=act11,
name='l1_2',
filters=64,
training=training,
reuse=reuse,
pool=False)
poolact12 = pooling.max_pool(bottom=act12, name='l1_2_pool')
with tf.variable_scope('g2', reuse=reuse):
# Downsample
act21 = conv_block(x=poolact12,
name='l2_1',
filters=128,
training=training,
reuse=reuse,
pool=False)
act22 = conv_block(x=act21,
filters=128,
name='l2_2',
training=training,
reuse=reuse,
pool=False)
poolact22 = pooling.max_pool(bottom=act22, name='l2_2_pool')
with tf.variable_scope('g3', reuse=reuse):
# Downsample
act31 = conv_block(x=poolact22,
name='l3_1',
filters=256,
training=training,
reuse=reuse,
pool=False)
act32 = conv_block(x=act31,
filters=256,
name='l3_2',
training=training,
reuse=reuse,
pool=False)
act33 = conv_block(x=act32,
filters=256,
name='l3_3',
training=training,
reuse=reuse,
pool=False)
poolact33 = pooling.max_pool(bottom=act33, name='l3_3_pool')
with tf.variable_scope('g4', reuse=reuse):
# Downsample
act41 = conv_block(x=poolact33,
name='l4_1',
filters=512,
training=training,
reuse=reuse,
pool=False)
act42 = conv_block(x=act41,
filters=512,
name='l4_2',
training=training,
reuse=reuse,
pool=False)
act43 = conv_block(x=act42,
filters=512,
name='l4_3',
training=training,
reuse=reuse,
pool=False)
poolact43 = pooling.max_pool(bottom=act43, name='l4_3_pool')
with tf.variable_scope('g5', reuse=reuse):
# Downsample
act51 = conv_block(x=poolact43,
name='l5_1',
filters=512,
training=training,
reuse=reuse,
pool=False)
act52 = conv_block(x=act51,
filters=512,
name='l5_2',
training=training,
reuse=reuse,
pool=False)
act53 = conv_block(x=act52,
filters=512,
name='l5_3',
training=training,
reuse=reuse,
pool=False)
poolact53 = pooling.max_pool(bottom=act53, name='l5_3_pool')
with tf.variable_scope('g5_skip', reuse=reuse):
upact5 = up_block(inputs=poolact53,
skip=act53,
up_filters=512,
name='ul5',
training=training,
reuse=reuse)
with tf.variable_scope('g4_skip', reuse=reuse):
upact4 = up_block(inputs=upact5,
skip=act43,
up_filters=512,
name='ul4',
training=training,
reuse=reuse)
with tf.variable_scope('g3_skip', reuse=reuse):
upact3 = up_block(inputs=upact4,
skip=act33,
up_filters=256,
name='ul3',
training=training,
reuse=reuse)
with tf.variable_scope('g2_skip', reuse=reuse):
upact2 = up_block(inputs=upact3,
skip=act22,
up_filters=128,
name='ul2',
training=training,
reuse=reuse)
with tf.variable_scope('g1_skip', reuse=reuse):
upact1 = up_block(inputs=upact2,
skip=act12,
up_filters=64,
name='ul1',
training=training,
reuse=reuse)
activity = conv.readout_layer(activity=upact1,
reuse=reuse,
training=training,
output_shape=output_shape)
extra_activities = {'activity': activity}
return activity, extra_activities
