def model(is_training, reuse, num_classes=5, dropout_keep_prob=0.5):
common_args = common_layer_args(is_training, reuse)
conv_args = make_args(batch_norm=True,
activation=prelu,
w_init=initz.he_normal(scale=1),
untie_biases=False,
**common_args)
conv_args_fm = make_args(w_init=initz.he_normal(scale=1),
untie_biases=False,
**common_args)
pool_args = make_args(padding='SAME', **common_args)
inputs = input((None, crop_size[1], crop_size[0], 3), **common_args)
with tf.variable_scope('squeezenet', values=[inputs]):
net = conv2d(inputs, 96, stride=(2, 2), name='conv1', **conv_args)
net = max_pool(net, name='maxpool1', **pool_args)
net = fire_module(net, 16, 64, name='fire2', **conv_args_fm)
net = bottleneck_simple(net, 16, 64, name='fire3', **conv_args_fm)
net = batch_norm(net,
activation_fn=tf.nn.relu,
name='fire3_bn',
is_training=is_training,
reuse=reuse)
net = fire_module(net, 32, 128, name='fire4', **conv_args_fm)
net = max_pool(net, name='maxpool4', **pool_args)
net = bottleneck_simple(net, 32, 128, name='fire5', **conv_args_fm)
net = batch_norm(net,
activation_fn=tf.nn.relu,
name='fire5_bn',
is_training=is_training,
reuse=reuse)
net = fire_module(net, 48, 192, name='fire6', **conv_args_fm)
net = bottleneck_simple(net, 48, 192, name='fire7', **conv_args_fm)
net = batch_norm(net,
activation_fn=tf.nn.relu,
name='fire7_bn',
is_training=is_training,
reuse=reuse)
net = fire_module(net, 64, 256, name='fire8', **conv_args_fm)
net = max_pool(net, name='maxpool8', **pool_args)
net = bottleneck_simple(net, 64, 256, name='fire9', **conv_args_fm)
net = batch_norm(net,
activation_fn=tf.nn.relu,
name='fire9_bn',
is_training=is_training,
reuse=reuse)
# Reversed avg and conv layers per 'Network in Network'
net = dropout(net,
drop_p=1 - dropout_keep_prob,
name='dropout6',
**common_args)
net = conv2d(net,
num_classes,
filter_size=(1, 1),
name='conv10',
**conv_args_fm)
logits = global_avg_pool(net, name='logits', **pool_args)
predictions = softmax(logits, name='predictions', **common_args)
return end_points(is_training)
def generator(z_shape, is_training, reuse, batch_size=32):
common_args = common_layer_args(is_training, reuse)
conv_args = make_args(batch_norm=True,
activation=lrelu,
w_init=initz.he_normal(scale=1),
untie_biases=False,
**common_args)
conv_args_1st = make_args(batch_norm=None,
activation=lrelu,
w_init=initz.he_normal(scale=1),
untie_biases=False,
**common_args)
fc_args = make_args(activation=lrelu,
w_init=initz.he_normal(scale=1),
**common_args)
pool_args = make_args(padding='SAME', **common_args)
# project `z` and reshape
# TODO think about phase again
end_points = {}
z = get_z(z_shape, reuse)
end_points['z'] = z
z_fc = fully_connected(z, 4 * 4 * 512, name="g_fc", **fc_args)
end_points['g_fc'] = z_fc
x = tf.reshape(z_fc, [batch_size, 4, 4, 512])
end_points['g_reshaped'] = x
x = upsample2d(x, [batch_size, 8, 8, 256],
name="g_deconv2d_1",
**conv_args)
end_points['g_deconv2d_1'] = x
x = upsample2d(x, [batch_size, 16, 16, 128],
name="g_deconv2d_2",
**conv_args)
end_points['g_deconv2d_2'] = x
# x = upsample2d(x, [batch_size, 32, 32, 16 * 3],
# name="g_deconv2d_3", **conv_args)
# end_points['g_deconv2d_3'] = x
# for now lets examine cifar
# x = subpixel2d(x, 4, name='z_subpixel1')
# x shape[batch_size, 128, 128, 3]
# end_points['subpixel1'] = x
x = upsample2d(x, [batch_size, 32, 32, 64],
name="g_deconv2d_3",
**conv_args)
end_points['g_deconv2d_3'] = x
x = upsample2d(x, [batch_size, 64, 64, 32],
name="g_deconv2d_4",
**conv_args)
end_points['g_deconv2d_4'] = x
x = upsample2d(x, [batch_size, 128, 128, 3],
name="g_deconv2d_5",
**conv_args_1st)
end_points['g_deconv2d_5'] = x
end_points['softmax'] = tf.nn.tanh(x)
return end_points
def encoder(inputs, is_training, reuse, z_dim=512):
common_args = common_layer_args(is_training, reuse)
conv_args = make_args(batch_norm=True,
activation=lrelu,
w_init=initz.he_normal(scale=1),
untie_biases=False,
**common_args)
conv_args_1st = make_args(batch_norm=None,
activation=lrelu,
w_init=initz.he_normal(scale=1),
untie_biases=False,
**common_args)
logits_args = make_args(activation=None,
w_init=initz.he_normal(scale=1),
**common_args)
pool_args = make_args(padding='SAME', **common_args)
end_points = {}
x = inputs
end_points['inputs'] = x
x = dropout(x, drop_p=0.2, name="input_dropout1", **common_args)
x = conv2d(x,
96,
filter_size=(5, 5),
stride=(2, 2),
name="e_conv1_1",
**conv_args_1st)
end_points['e_conv1_1'] = x
x = conv2d(x, 96, name="e_conv1_2", **conv_args)
end_points['e_conv1_2'] = x
x = conv2d(x, 96, stride=(2, 2), name="e_conv1_3", **conv_args)
end_points['e_conv1_3'] = x
x = dropout(x, drop_p=0.2, name="dropout1", **common_args)
x = conv2d(x, 192, name="e_conv2_1", **conv_args)
end_points['e_conv2_1'] = x
x = conv2d(x, 192, name="e_conv2_2", **conv_args)
end_points['e_conv2_2'] = x
# x = conv2d(x, 192, stride=(2, 2), name="e_conv2_3", **conv_args)
# end_points['e_conv2_3'] = x
x = dropout(x, drop_p=0.2, name="dropout2", **common_args)
# x = conv2d(x, 192, stride=(2, 2), name="e_conv3_1", **conv_args)
# end_points['e_conv3_1'] = x
x = conv2d(x, 192, filter_size=(1, 1), name="e_conv4_1", **conv_args)
end_points['e_conv4_1'] = x
x = conv2d(x, 192, filter_size=(1, 1), name="e_conv4_2", **conv_args)
end_points['e_conv4_2'] = x
x = global_avg_pool(x, name="global_pool")
end_points['global_pool'] = x
logits1 = fully_connected(x, z_dim, name="e_logits1", **logits_args)
logits2 = fully_connected(x, z_dim, name="e_logits2", **logits_args)
logits2 = tf.tanh(logits2, name='e_logits2_tanh')
end_points['e_logits1'] = logits1
end_points['e_logits2'] = logits2
return end_points
def discriminator(inputs, is_training, reuse, num_classes=1):
common_args = common_layer_args(is_training, reuse)
conv_args = make_args(batch_norm=True,
activation=lrelu,
w_init=initz.he_normal(scale=1),
untie_biases=False,
**common_args)
conv_args_1st = make_args(batch_norm=None,
activation=lrelu,
w_init=initz.he_normal(scale=1),
untie_biases=False,
**common_args)
logits_args = make_args(activation=None,
w_init=initz.he_normal(scale=1),
**common_args)
pool_args = make_args(padding='SAME', **common_args)
end_points = {}
x = inputs
end_points['inputs'] = x
x = dropout(x, drop_p=0.2, name="input_dropout1", **common_args)
x = conv2d(x,
96,
filter_size=(5, 5),
stride=(2, 2),
name="d_conv1_1",
**conv_args_1st)
end_points['d_conv1_1'] = x
x = conv2d(x, 96, name="d_conv1_2", **conv_args)
end_points['d_conv1_2'] = x
x = conv2d(x, 96, stride=(2, 2), name="d_conv1_3", **conv_args)
end_points['d_conv1_3'] = x
x = dropout(x, drop_p=0.2, name="dropout1", **common_args)
x = conv2d(x, 192, name="d_conv2_1", **conv_args)
end_points['d_conv2_1'] = x
x = conv2d(x, 192, name="d_conv2_2", **conv_args)
end_points['d_conv2_2'] = x
# x = conv2d(x, 192, stride=(2, 2), name="d_conv2_3", **conv_args)
# end_points['d_conv2_3'] = x
x = dropout(x, drop_p=0.2, name="dropout2", **common_args)
# x = conv2d(x, 192, stride=(2, 2), name="d_conv3_1", **conv_args)
# end_points['d_conv3_1'] = x
x = conv2d(x, 192, filter_size=(1, 1), name="d_conv4_1", **conv_args)
end_points['d_conv4_1'] = x
x = conv2d(x, 192, filter_size=(1, 1), name="d_conv4_2", **conv_args)
end_points['d_conv4_2'] = x
x = global_avg_pool(x, name="global_pool")
end_points['global_pool'] = x
logits = fully_connected(x, num_classes, name="d_logits", **logits_args)
end_points['logits'] = logits
end_points['predictions'] = softmax(logits,
name='predictions',
**common_args)
return end_points
def model(inputs,
is_training,
reuse,
input_size=image_size[0],
drop_p_conv=0.0,
drop_p_trans=0.0,
n_filters=64,
n_layers=[1, 2, 2, 3],
num_classes=5, **kwargs):
common_args = common_layer_args(is_training, reuse)
conv_args = make_args(
batch_norm=True,
activation=prelu,
w_init=initz.he_normal(scale=1),
untie_biases=True,
**common_args)
fc_args = make_args(activation=prelu, w_init=initz.he_normal(scale=1), **common_args)
logit_args = make_args(activation=None, w_init=initz.he_normal(scale=1), **common_args)
pred_args = make_args(activation=prelu, w_init=initz.he_normal(scale=1), **common_args)
pool_args = make_args(padding='SAME', filter_size=(2, 2), stride=(2, 2), **common_args)
x = conv2d(inputs, 48, filter_size=(7, 7), name="conv1", **conv_args)
x = max_pool(x, name='pool1', **pool_args)
x = conv2d(x, 64, name="conv2_1", **conv_args)
x = conv2d(x, 64, name="conv2_2", **conv_args)
x = max_pool(x, name='pool2', **pool_args)
# 112
for block_idx in range(3):
x, n_filters = dense_block(
x,
n_filters,
num_layers=n_layers[block_idx],
drop_p=drop_p_conv,
block_name='dense_' + str(block_idx),
**conv_args)
x = trans_block(
x, n_filters, drop_p=drop_p_trans, block_name='trans_' + str(block_idx), **conv_args)
x, n_filters = dense_block(
x, n_filters, num_layers=n_layers[3], drop_p=drop_p_trans, block_name='dense_3', **conv_args)
# 8
x = global_avg_pool(x, name='avgpool_1a_8x8')
logits = fully_connected(x, n_output=num_classes, name="logits", **logit_args)
predictions = softmax(logits, name='predictions', **common_args)
return end_points(is_training)
def fully_connected(x,
n_output,
is_training,
reuse,
activation=None,
batch_norm=None,
batch_norm_args=None,
w_init=initz.he_normal(),
use_bias=True,
b_init=0.0,
w_regularizer=tf.nn.l2_loss,
outputs_collections=None,
trainable=True,
name='fc'):
input_shape = helper.get_input_shape(x)
assert len(input_shape) > 1, "Input Tensor shape must be > 1-D"
if len(x.get_shape()) != 2:
x = _flatten(x)
n_input = helper.get_input_shape(x)[1]
with tf.variable_scope(name, reuse=reuse) as curr_scope:
shape = [n_input, n_output] if hasattr(w_init, '__call__') else None
W = tf.get_variable(name='weights',
shape=shape,
initializer=w_init,
regularizer=w_regularizer,
trainable=trainable)
output = tf.matmul(x, W)
if use_bias:
b = tf.get_variable(name='biases',
shape=[n_output],
initializer=tf.constant_initializer(b_init),
trainable=trainable)
output = tf.nn.bias_add(value=output, bias=b)
if batch_norm is not None and batch_norm is not False:
if isinstance(batch_norm, bool):
batch_norm = batch_norm_tf
batch_norm_args = batch_norm_args or {}
output = batch_norm(output,
is_training=is_training,
reuse=reuse,
trainable=trainable,
**batch_norm_args)
if activation:
output = activation(output,
is_training=is_training,
reuse=reuse,
trainable=trainable)
return _collect_named_outputs(outputs_collections,
curr_scope.original_name_scope, name,
output)
def conv2d_same(inputs, n_output_channels, is_training, reuse, filter_size=(3, 3), stride=(1, 1), dilation_rate=1,
activation=None, batch_norm=None, batch_norm_args=None, w_init=initz.he_normal(),
use_bias=True, untie_biases=False, b_init=0.0, w_regularizer=tf.nn.l2_loss,
outputs_collections=None, trainable=True, name='conv2d_same'):
"""Strided 2-D convolution with 'SAME' padding.
When stride > 1, then we do explicit zero-padding, followed by conv2d with
'VALID' padding.
Note that
net = conv2d_same(inputs, num_outputs, 3, stride=stride)
is equivalent to
net = slim.conv2d(inputs, num_outputs, 3, stride=1, padding='SAME')
net = subsample(net, factor=stride)
whereas
net = slim.conv2d(inputs, num_outputs, 3, stride=stride, padding='SAME')
is different when the input's height or width is even, which is why we add the
current function. For more details, see ResnetUtilsTest.testConv2DSameEven().
"""
if stride[0] == 1:
return conv2d(inputs, n_output_channels, is_training, reuse, filter_size=filter_size, stride=stride,
dilation_rate=dilation_rate, padding='SAME', activation=activation, batch_norm=batch_norm,
batch_norm_args=batch_norm_args, w_init=w_init, use_bias=use_bias, untie_biases=untie_biases,
b_init=b_init, w_regularizer=w_regularizer, outputs_collections=outputs_collections,
trainable=trainable, name=name)
else:
kernel_size = filter_size[0]
kernel_size_effective = kernel_size + (kernel_size - 1) * (dilation_rate - 1)
pad_total = kernel_size_effective - 1
pad_beg = pad_total // 2
pad_end = pad_total - pad_beg
inputs = tf.pad(inputs,
[[0, 0], [pad_beg, pad_end], [pad_beg, pad_end], [0, 0]])
return conv2d(inputs, n_output_channels, is_training, reuse, filter_size=filter_size, stride=stride,
dilation_rate=dilation_rate, padding='VALID', activation=activation, batch_norm=batch_norm,
batch_norm_args=batch_norm_args, w_init=w_init, use_bias=use_bias, untie_biases=untie_biases,
b_init=b_init, w_regularizer=w_regularizer, outputs_collections=outputs_collections,
trainable=trainable, name=name)
def conv2d(x,
n_output_channels,
is_training,
reuse,
filter_size=(3, 3),
stride=(1, 1),
dilation_rate=1,
padding='SAME',
activation=None,
batch_norm=None,
batch_norm_args=None,
w_init=initz.he_normal(),
use_bias=True,
untie_biases=False,
b_init=0.0,
w_regularizer=tf.nn.l2_loss,
outputs_collections=None,
trainable=True,
name='conv2d'):
input_shape = helper.get_input_shape(x)
assert len(input_shape) == 4, "Input Tensor shape must be 4-D"
with tf.variable_scope(name, reuse=reuse) as curr_scope:
shape = [
filter_size[0], filter_size[1],
x.get_shape()[-1], n_output_channels
] if hasattr(w_init, '__call__') else None
W = tf.get_variable(name='weights',
shape=shape,
initializer=w_init,
regularizer=w_regularizer,
trainable=trainable)
if dilation_rate == 1:
output = tf.nn.conv2d(input=x,
filter=W,
strides=[1, stride[0], stride[1], 1],
padding=padding)
else:
if len([s for s in stride if s > 1]) > 0:
raise ValueError(
"Stride (%s) cannot be more than 1 if rate (%d) is not 1" %
(stride, dilation_rate))
output = tf.nn.atrous_conv2d(value=x,
filters=W,
rate=dilation_rate,
padding=padding)
if use_bias:
if untie_biases:
b = tf.get_variable(
name='biases',
shape=output.get_shape()[1:],
initializer=tf.constant_initializer(b_init),
trainable=trainable)
output = tf.add(output, b)
else:
b = tf.get_variable(
name='biases',
shape=[n_output_channels],
initializer=tf.constant_initializer(b_init),
trainable=trainable)
output = tf.nn.bias_add(value=output, bias=b)
if batch_norm is not None:
if isinstance(batch_norm, bool):
batch_norm = batch_norm_tf
batch_norm_args = batch_norm_args or {}
output = batch_norm(output,
is_training=is_training,
reuse=reuse,
trainable=trainable,
**batch_norm_args)
if activation:
output = activation(output,
is_training=is_training,
reuse=reuse,
trainable=trainable)
return _collect_named_outputs(outputs_collections,
curr_scope.original_name_scope, output)
def model(inputs, is_training, reuse, num_classes=21, batch_size=1):
common_args = common_layer_args(is_training, reuse)
conv_args = make_args(batch_norm=True, activation=lrelu, w_init=initz.he_normal(
scale=1), untie_biases=False, **common_args)
upsample_args = make_args(
batch_norm=False, activation=lrelu, use_bias=False, **common_args)
logits_args = make_args(
activation=None, **common_args)
pool_args = make_args(padding='SAME', **common_args)
conv1_1 = conv2d(inputs, 64, name="vgg_19/conv1/conv1_1", **conv_args)
conv1_2 = conv2d(conv1_1, 64, name="vgg_19/conv1/conv1_2", **conv_args)
pool1 = max_pool(conv1_2, stride=2, name='pool1', **pool_args)
conv2_1 = conv2d(pool1, 128, name="vgg_19/conv2/conv2_1", **conv_args)
conv2_2 = conv2d(conv2_1, 128, name="vgg_19/conv2/conv2_2", **conv_args)
pool2 = max_pool(conv2_2, stride=2, name='pool2', **pool_args)
conv3_1 = conv2d(pool2, 256, name="vgg_19/conv3/conv3_1", **conv_args)
conv3_2 = conv2d(conv3_1, 256, name="vgg_19/conv3/conv3_2", **conv_args)
conv3_3 = conv2d(conv3_2, 256, name="vgg_19/conv3/conv3_3", **conv_args)
conv3_4 = conv2d(conv3_3, 256, name="vgg_19/conv3/conv3_4", **conv_args)
pool3 = max_pool(conv3_4, stride=2, name='pool3', **pool_args)
conv4_1 = conv2d(pool3, 512, name="vgg_19/conv4/conv4_1", **conv_args)
conv4_2 = conv2d(conv4_1, 512, name="vgg_19/conv4/conv4_2", **conv_args)
conv4_3 = conv2d(conv4_2, 512, name="vgg_19/conv4/conv4_3", **conv_args)
conv4_4 = conv2d(conv4_3, 512, name="vgg_19/conv4/conv4_4", **conv_args)
pool4 = max_pool(conv4_4, stride=2, name='pool4', **pool_args)
conv5_1 = conv2d(pool4, 512, name="vgg_19/conv5/conv5_1", **conv_args)
conv5_2 = conv2d(conv5_1, 512, name="vgg_19/conv5/conv5_2", **conv_args)
conv5_3 = conv2d(conv5_2, 512, name="vgg_19/conv5/conv5_3", **conv_args)
conv5_4 = conv2d(conv5_3, 512, name="vgg_19/conv5/conv5_4", **conv_args)
pool5 = max_pool(conv5_4, stride=2, name='pool5', **pool_args)
fc6 = conv2d(pool5, 4096, filter_size=(7, 7),
name="vgg_19/fc6", **conv_args)
fc6 = dropout(fc6, **common_args)
fc7 = conv2d(fc6, 4096, filter_size=(1, 1), name="vgg_19/fc7", **conv_args)
fc7 = dropout(fc7, **common_args)
score_fr = conv2d(fc7, num_classes, filter_size=(1, 1),
name="score_fr", **conv_args)
pred = tf.argmax(score_fr, axis=3)
pool4_shape = pool4.get_shape().as_list()
upscore2 = upsample2d(score_fr, [batch_size, pool4_shape[1], pool4_shape[2], num_classes], filter_size=(4, 4), stride=(2, 2),
name="deconv2d_1", w_init=initz.bilinear((4, 4, num_classes, num_classes)), **upsample_args)
score_pool4 = conv2d(pool4, num_classes, filter_size=(1, 1),
name="score_pool4", **conv_args)
fuse_pool4 = tf.add(upscore2, score_pool4)
pool3_shape = pool3.get_shape().as_list()
upscore4 = upsample2d(fuse_pool4, [batch_size, pool3_shape[1], pool3_shape[2], num_classes], filter_size=(4, 4), stride=(2, 2),
name="deconv2d_2", w_init=initz.bilinear((4, 4, num_classes, num_classes)), **upsample_args)
score_pool3 = conv2d(pool3, num_classes, filter_size=(1, 1),
name="score_pool3", **conv_args)
fuse_pool3 = tf.add(upscore4, score_pool3)
input_shape = inputs.get_shape().as_list()
upscore32 = upsample2d(fuse_pool3, [batch_size, input_shape[1], input_shape[2], num_classes], filter_size=(16, 16), stride=(8, 8),
name="deconv2d_3", w_init=initz.bilinear((16, 16, num_classes, num_classes)), **logits_args)
logits = register_to_collections(tf.reshape(
upscore32, shape=(-1, num_classes)), name='logits', **common_args)
pred_up = tf.argmax(upscore32, axis=3)
pred_up = register_to_collections(
pred_up, name='final_prediction_map', **common_args)
predictions = softmax(logits, name='predictions', **common_args)
return end_points(is_training)
def discriminator(inputs, is_training, reuse, num_classes=11, batch_size=32):
common_args = common_layer_args(is_training, reuse)
conv_args = make_args(batch_norm=True,
activation=lrelu,
w_init=initz.he_normal(scale=1),
untie_biases=False,
**common_args)
conv_args_1st = make_args(batch_norm=None,
activation=lrelu,
w_init=initz.he_normal(scale=1),
untie_biases=False,
**common_args)
logits_args = make_args(activation=None,
w_init=initz.he_normal(scale=1),
**common_args)
pool_args = make_args(padding='SAME', **common_args)
end_points = {}
x = inputs
end_points['inputs'] = x
x = dropout(x, drop_p=0.2, name="input_dropout1", **common_args)
x = conv2d(x,
96,
filter_size=(5, 5),
stride=(2, 2),
name="d_conv1_1",
**conv_args_1st)
end_points['d_conv1_1'] = x
x = conv2d(x, 96, name="d_conv1_2", **conv_args)
end_points['d_conv1_2'] = x
x = conv2d(x, 96, stride=(2, 2), name="d_conv1_3", **conv_args)
end_points['d_conv1_3'] = x
x = dropout(x, drop_p=0.2, name="dropout1", **common_args)
x = conv2d(x, 192, name="d_conv2_1", **conv_args)
end_points['d_conv2_1'] = x
x = conv2d(x, 192, name="d_conv2_2", **conv_args)
end_points['d_conv2_2'] = x
x = conv2d(x, 192, stride=(2, 2), name="d_conv2_3", **conv_args)
end_points['d_conv2_3'] = x
x = dropout(x, drop_p=0.2, name="dropout2", **common_args)
x = conv2d(x, 192, stride=(2, 2), name="d_conv3_1", **conv_args)
end_points['d_conv3_1'] = x
x = conv2d(x, 192, filter_size=(1, 1), name="d_conv4_1", **conv_args)
end_points['d_conv4_1'] = x
x = conv2d(x, 192, filter_size=(1, 1), name="d_conv4_2", **conv_args)
end_points['d_conv4_2'] = x
x = global_avg_pool(x, name="global_pool")
end_points['global_pool'] = x
logits = fully_connected(x, num_classes, name="d_logits", **logits_args)
end_points['logits'] = logits
end_points['predictions'] = softmax(logits,
name='predictions',
**common_args)
if is_training:
batch_size = 2 * batch_size
generated_class_logits = tf.squeeze(
tf.slice(logits, [0, num_classes - 1], [batch_size, 1]))
end_points['generated_class_logits'] = generated_class_logits
positive_class_logits = tf.slice(logits, [0, 0],
[batch_size, num_classes - 1])
end_points['positive_class_logits'] = positive_class_logits
max_ = tf.reduce_max(positive_class_logits, 1, keep_dims=True)
safe_pos_class_logits = positive_class_logits - max_
end_points['safe_pos_class_logits'] = safe_pos_class_logits
gan_logits = tf.log(
tf.reduce_sum(tf.exp(safe_pos_class_logits),
1)) + tf.squeeze(max_) - generated_class_logits
end_points['gan_logits'] = gan_logits
assert len(gan_logits.get_shape()) == 1
probs = tf.nn.sigmoid(gan_logits)
end_points['probs'] = probs
class_logits = tf.slice(logits, [0, 0], [batch_size / 2, num_classes])
end_points['class_logits'] = class_logits
D_on_data = tf.slice(probs, [0], [batch_size / 2])
end_points['D_on_data'] = D_on_data
D_on_data_logits = tf.slice(gan_logits, [0], [batch_size / 2])
end_points['D_on_data_logits'] = D_on_data_logits
D_on_G = tf.slice(probs, [batch_size / 2], [batch_size / 2])
end_points['D_on_G'] = D_on_G
D_on_G_logits = tf.slice(gan_logits, [batch_size / 2],
[batch_size / 2])
end_points['D_on_G_logits'] = D_on_G_logits
return end_points
else:
return end_points
def vgg_16(is_training, reuse,
num_classes=1000,
dropout_keep_prob=0.5,
spatial_squeeze=True,
name='vgg_16'):
"""Oxford Net VGG 16-Layers version D Example.
Note: All the fully_connected layers have been transformed to conv2d layers.
To use in classification mode, resize input to 224x224.
Args:
inputs: a tensor of size [batch_size, height, width, channels].
num_classes: number of predicted classes.
is_training: whether or not the model is being trained.
dropout_keep_prob: the probability that activations are kept in the dropout
layers during training.
spatial_squeeze: whether or not should squeeze the spatial dimensions of the
outputs. Useful to remove unnecessary dimensions for classification.
name: Optional name for the variables.
Returns:
the last op containing the log predictions and end_points dict.
"""
common_args = common_layer_args(is_training, reuse)
conv_args = make_args(batch_norm=True, activation=prelu, w_init=initz.he_normal(
scale=1), untie_biases=False, **common_args)
logit_args = make_args(
activation=None, w_init=initz.he_normal(scale=1), **common_args)
pred_args = make_args(
activation=prelu, w_init=initz.he_normal(scale=1), **common_args)
pool_args = make_args(padding='SAME', **common_args)
inputs = input((None, crop_size[1], crop_size[0], 3), **common_args)
with tf.variable_scope(name, 'vgg_16', [inputs]):
net = repeat(inputs, 2, conv2d,
64, filter_size=(3, 3), name='conv1', **conv_args)
net = max_pool(net, name='pool1', **pool_args)
net = repeat(net, 2, conv2d, 128, filter_size=(
3, 3), name='conv2', **conv_args)
net = max_pool(net, name='pool2', **pool_args)
net = repeat(net, 3, conv2d, 256, filter_size=(
3, 3), name='conv3', **conv_args)
net = max_pool(net, name='pool3', **pool_args)
net = repeat(net, 3, conv2d, 512, filter_size=(
3, 3), name='conv4', **conv_args)
net = max_pool(net, name='pool4', **pool_args)
net = repeat(net, 3, conv2d, 512, filter_size=(
3, 3), name='conv5', **conv_args)
net = max_pool(net, name='pool5', **pool_args)
# Use conv2d instead of fully_connected layers.
net = conv2d(net, 4096, filter_size=(7, 7), name='fc6', **conv_args)
net = dropout(net, drop_p=1 - dropout_keep_prob, is_training=is_training,
name='dropout6', **common_args)
net = conv2d(net, 4096, filter_size=(1, 1), name='fc7', **conv_args)
net = dropout(net, drop_p=1 - dropout_keep_prob, is_training=is_training,
name='dropout7', **common_args)
logits = conv2d(net, num_classes, filter_size=(1, 1),
activation=None,
name='logits', **logit_args)
# Convert end_points_collection into a end_point dict.
if spatial_squeeze:
logits = tf.squeeze(logits, [1, 2], name='logits/squeezed')
predictions = softmax(logits, name='predictions', **pred_args)
return end_points(is_training)
def model(inputs,
is_training,
reuse,
num_classes=5,
dropout_keep_prob=0.5,
spatial_squeeze=True,
name='alexnet_v2',
**kwargs):
"""AlexNet version 2.
Described in: http://arxiv.org/pdf/1404.5997v2.pdf
Parameters from:
github.com/akrizhevsky/cuda-convnet2/blob/master/layers/
layers-imagenet-1gpu.cfg
Note: All the fully_connected layers have been transformed to conv2d layers.
To use in classification mode, resize input to 224x224. To use in fully
convolutional mode, set spatial_squeeze to false.
The LRN layers have been removed and change the initializers from
random_normal_initializer to xavier_initializer.
Args:
inputs: a tensor of size [batch_size, height, width, channels].
num_classes: number of predicted classes.
is_training: whether or not the model is being trained.
dropout_keep_prob: the probability that activations are kept in the dropout
layers during training.
spatial_squeeze: whether or not should squeeze the spatial dimensions of the
outputs. Useful to remove unnecessary dimensions for classification.
name: Optional name for the variables.
Returns:
the last op containing the log predictions and end_points dict.
"""
common_args = common_layer_args(is_training, reuse)
conv_args = make_args(batch_norm=True,
activation=prelu,
w_init=initz.he_normal(scale=1),
untie_biases=False,
**common_args)
logit_args = make_args(activation=None,
w_init=initz.he_normal(scale=1),
**common_args)
pred_args = make_args(activation=prelu,
w_init=initz.he_normal(scale=1),
**common_args)
pool_args = make_args(padding='SAME', **common_args)
# inputs = input((None, crop_size[1], crop_size[0], 3), **common_args)
with tf.variable_scope(name, 'alexnet_v2', [inputs]):
net = conv2d(inputs,
64,
filter_size=(11, 11),
stride=(4, 4),
name='conv1',
**conv_args)
net = max_pool(net, stride=(2, 2), name='pool1', **pool_args)
net = conv2d(net, 192, filter_size=(5, 5), name='conv2', **conv_args)
net = max_pool(net, stride=(2, 2), name='pool2', **pool_args)
net = conv2d(net, 384, name='conv3', **conv_args)
net = conv2d(net, 384, name='conv4', **conv_args)
net = conv2d(net, 256, name='conv5', **conv_args)
net = max_pool(net, stride=(2, 2), name='pool5', **pool_args)
# Use conv2d instead of fully_connected layers.
net = conv2d(net, 4096, filter_size=(5, 5), name='fc6', **conv_args)
net = dropout(net,
drop_p=1 - dropout_keep_prob,
name='dropout6',
**common_args)
net = conv2d(net, 4096, filter_size=(1, 1), name='fc7', **conv_args)
net = dropout(net,
drop_p=1 - dropout_keep_prob,
name='dropout7',
**common_args)
net = global_avg_pool(net)
logits = fc(net, num_classes, name='logits', **logit_args)
predictions = softmax(logits, name='predictions', **common_args)
return end_points(is_training)
def resnet_v1(inputs,
is_training,
reuse,
blocks,
num_classes=None,
global_pool=True,
output_stride=None,
include_root_block=True,
name=None):
"""Generator for v2 (preactivation) ResNet models.
This function generates a family of ResNet v2 models. See the resnet_v2_*()
methods for specific model instantiations, obtained by selecting different
block instantiations that produce ResNets of various depths.
Training for image classification on Imagenet is usually done with [224, 224]
inputs, resulting in [7, 7] feature maps at the output of the last ResNet
block for the ResNets defined in [1] that have nominal stride equal to 32.
However, for dense prediction tasks we advise that one uses inputs with
spatial dimensions that are multiples of 32 plus 1, e.g., [321, 321]. In
this case the feature maps at the ResNet output will have spatial shape
[(height - 1) / output_stride + 1, (width - 1) / output_stride + 1]
and corners exactly aligned with the input image corners, which greatly
facilitates alignment of the features to the image. Using as input [225, 225]
images results in [8, 8] feature maps at the output of the last ResNet block.
For dense prediction tasks, the ResNet needs to run in fully-convolutional
(FCN) mode and global_pool needs to be set to False. The ResNets in [1, 2] all
have nominal stride equal to 32 and a good choice in FCN mode is to use
output_stride=16 in order to increase the density of the computed features at
small computational and memory overhead, cf. http://arxiv.org/abs/1606.00915.
Args:
inputs: A tensor of size [batch, height_in, width_in, channels].
blocks: A list of length equal to the number of ResNet blocks. Each element
is a resnet_utils.Block object describing the units in the block.
num_classes: Number of predicted classes for classification tasks. If None
we return the features before the logit layer.
is_training: whether is training or not.
global_pool: If True, we perform global average pooling before computing the
logits. Set to True for image classification, False for dense prediction.
output_stride: If None, then the output will be computed at the nominal
network stride. If output_stride is not None, it specifies the requested
ratio of input to output spatial resolution.
include_root_block: If True, include the initial convolution followed by
max-pooling, if False excludes it. If excluded, `inputs` should be the
results of an activation-less convolution.
reuse: whether or not the network and its variables should be reused. To be
able to reuse 'scope' must be given.
name: Optional variable_scope.
Returns:
net: A rank-4 tensor of size [batch, height_out, width_out, channels_out].
If global_pool is False, then height_out and width_out are reduced by a
factor of output_stride compared to the respective height_in and width_in,
else both height_out and width_out equal one. If num_classes is None, then
net is the output of the last ResNet block, potentially after global
average pooling. If num_classes is not None, net contains the pre-softmax
activations.
end_points: A dictionary from components of the network to the corresponding
activation.
Raises:
ValueError: If the target output_stride is not valid.
"""
common_args = common_layer_args(is_training, reuse)
conv_args = make_args(batch_norm=True,
activation=prelu,
w_init=initz.he_normal(scale=1),
untie_biases=False,
**common_args)
logits_args = make_args(activation=None,
w_init=initz.he_normal(scale=1),
**common_args)
pred_args = make_args(activation=prelu,
w_init=initz.he_normal(scale=1),
**common_args)
pool_args = make_args(padding='SAME', **common_args)
with tf.variable_scope(name, 'resnet_v2', [inputs], reuse=reuse):
net = inputs
if include_root_block:
if output_stride is not None:
if output_stride % 4 != 0:
raise ValueError(
'The output_stride needs to be a multiple of 4.')
output_stride /= 4
# We do not include batch normalization or activation functions in
# conv1 because the first ResNet unit will perform these. Cf.
# Appendix of [2].
net = resnet_utils.conv2d_same(net,
64,
7,
stride=2,
name='conv1',
**common_args)
net = max_pool(net, name='pool1', **pool_args)
net = resnet_utils.stack_blocks_dense(net, blocks, output_stride,
**conv_args)
# This is needed because the pre-activation variant does not have batch
# normalization or activation functions in the residual unit output. See
# Appendix of [2].
net = batch_norm(net,
activation=tf.nn.relu,
name='postnorm',
is_training=is_training,
reuse=reuse)
if global_pool:
# Global average pooling.
net = tf.reduce_mean(net, [1, 2], name='pool5', keep_dims=True)
if num_classes is not None:
net = conv2d(net,
num_classes,
filter_size=(1, 1),
name='logits',
**logits_args)
if num_classes is not None:
predictions = softmax(net, name='predictions', **pred_args)
return end_points(is_training)
def _linear(x,
n_output,
reuse,
trainable=True,
w_init=initz.he_normal(),
b_init=0.0,
w_regularizer=tf.nn.l2_loss,
name='fc',
layer_norm=None,
layer_norm_args=None,
activation=None,
outputs_collections=None,
use_bias=True):
"""Adds a fully connected layer.
`fully_connected` creates a variable called `weights`, representing a fully
connected weight matrix, which is multiplied by the `x` to produce a
`Tensor` of hidden units. If a `layer_norm` is provided (such as
`layer_norm`), it is then applied. Otherwise, if `layer_norm` is
None and a `b_init` and `use_bias` is provided then a `biases` variable would be
created and added the hidden units. Finally, if `activation` is not `None`,
it is applied to the hidden units as well.
Note: that if `x` have a rank greater than 2, then `x` is flattened
prior to the initial matrix multiply by `weights`.
Args:
x: A `Tensor` of with at least rank 2 and value for the last dimension,
i.e. `[batch_size, depth]`, `[None, None, None, channels]`.
n_output: Integer or long, the number of output units in the layer.
reuse: whether or not the layer and its variables should be reused. To be
able to reuse the layer scope must be given.
activation: activation function, set to None to skip it and maintain
a linear activation.
layer_norm: normalization function to use. If
`batch_norm` is `True` then google original implementation is used and
if another function is provided then it is applied.
default set to None for no normalizer function
layer_norm_args: normalization function parameters.
w_init: An initializer for the weights.
w_regularizer: Optional regularizer for the weights.
b_init: An initializer for the biases. If None skip biases.
outputs_collections: The collections to which the outputs are added.
trainable: If `True` also add variables to the graph collection
`GraphKeys.TRAINABLE_VARIABLES` (see tf.Variable).
name: Optional name or scope for variable_scope/name_scope.
use_bias: Whether to add bias or not
Returns:
The 2-D `Tensor` variable representing the result of the series of operations.
e.g: 2-D `Tensor` [batch, n_output].
Raises:
ValueError: if `x` has rank less than 2 or if its last dimension is not set.
ValueError: linear is expecting 2D arguments
ValueError: linear expects shape[1] to be provided for shape
"""
if not (isinstance(n_output, six.integer_types)):
raise ValueError('n_output should be int or long, got %s.', n_output)
if not helper.is_sequence(x):
x = [x]
n_input = 0
shapes = [_x.get_shape() for _x in x]
for shape in shapes:
if shape.ndims != 2:
raise ValueError("linear is expecting 2D arguments: %s" % shapes)
if shape[1].value is None:
raise ValueError(
"linear expects shape[1] to be provided for shape %s, but saw %s"
% (shape, shape[1]))
else:
n_input += shape[1].value
with tf.variable_scope(name, reuse=reuse):
shape = [n_input, n_output] if hasattr(w_init, '__call__') else None
W = tf.get_variable(name='W',
shape=shape,
dtype=tf.float32,
initializer=w_init,
regularizer=w_regularizer,
trainable=trainable)
if len(x) == 1:
output = tf.matmul(x[0], W)
else:
output = tf.matmul(tf.concat_v2(x, 1), W)
if use_bias:
b = tf.get_variable(
name='b',
shape=[n_output],
dtype=tf.float32,
initializer=tf.constant_initializer(b_init),
trainable=trainable,
)
output = tf.nn.bias_add(value=output, bias=b)
if layer_norm is not None:
layer_norm_args = layer_norm_args or {}
output = layer_norm(output,
reuse=reuse,
trainable=trainable,
**layer_norm_args)
if activation:
output = activation(output, reuse=reuse, trainable=trainable)
return _collect_named_outputs(outputs_collections, name, output)
def model(height,
width,
num_actions,
is_training=False,
reuse=None,
name=None):
common_args = common_layer_args(is_training, reuse)
conv_args = make_args(batch_norm=True,
activation=prelu,
w_init=initz.he_normal(scale=1),
untie_biases=False,
**common_args)
logits_args = make_args(activation=None,
w_init=initz.he_normal(scale=1),
**common_args)
fc_args = make_args(activation=prelu,
w_init=initz.he_normal(scale=1),
**common_args)
pool_args = make_args(padding='SAME', **common_args)
with tf.variable_scope(name):
state = register_to_collections(tf.placeholder(
shape=[None, 4, height, width], dtype=tf.float32, name='state'),
name='state',
**common_args)
state_perm = tf.transpose(state, perm=[0, 2, 3, 1])
summary_ops = [
tf.summary.image("states",
state[:, 0, :, :][..., tf.newaxis],
max_outputs=10,
collections='train')
]
conv1_0 = conv2d(state_perm,
32,
filter_size=8,
stride=(1, 1),
name="conv1_0",
**conv_args)
conv1_1 = conv2d(conv1_0,
64,
filter_size=8,
stride=(2, 2),
name="conv1_1",
**conv_args)
pool = max_pool(conv1_1, filter_size=2, name="maxpool", **pool_args)
conv2_0 = conv2d(pool,
128,
filter_size=4,
stride=2,
name="conv2_0",
**conv_args)
conv2_1 = conv2d(conv2_0,
256,
filter_size=3,
stride=(2, 2),
name="conv2_1",
**conv_args)
conv3_0 = conv2d(conv2_1,
256,
filter_size=4,
stride=1,
name="conv3_0",
**conv_args)
conv3_1 = conv2d(conv3_0,
512,
filter_size=4,
stride=2,
name="conv3_1",
**conv_args)
# Dueling
value_hid = fc(conv3_1, 512, name="value_hid", **fc_args)
adv_hid = fc(conv3_1, 512, name="adv_hid", **fc_args)
value = fc(value_hid, 1, name="value", **logits_args)
advantage = fc(adv_hid, num_actions, name="advantage", **logits_args)
# Average Dueling
Qs = value + (advantage -
tf.reduce_mean(advantage, axis=1, keep_dims=True))
# action with highest Q values
a = register_to_collections(tf.argmax(Qs, 1), name='a', **common_args)
# Q value belonging to selected action
Q = register_to_collections(tf.reduce_max(Qs, 1),
name='Q',
**common_args)
summary_ops.append(tf.summary.histogram("Q", Q, collections='train'))
# For training
Q_target = register_to_collections(tf.placeholder(shape=[None],
dtype=tf.float32),
name='Q_target',
**common_args)
actions = register_to_collections(tf.placeholder(shape=[None],
dtype=tf.int32),
name='actions',
**common_args)
actions_onehot = tf.one_hot(actions,
num_actions,
on_value=1.,
off_value=0.,
axis=1,
dtype=tf.float32)
Q_tmp = tf.reduce_sum(tf.multiply(Qs, actions_onehot), axis=1)
loss = register_to_collections(tf.reduce_mean(
tf.square(Q_target - Q_tmp)),
name='loss',
**common_args)
summary_ops.append(tf.summary.scalar("loss", loss,
collections='train'))
register_to_collections(summary_ops, name='summary_ops', **common_args)
return end_points(is_training)
