def upsample2x(name, x):
"""
Nearest neighbor up-sampling
"""
return FixedUnPooling(
name, x, 2, unpool_mat=np.ones((2, 2), dtype='float32'),
data_format='channels_first')
def upsample2x(name, x):
dtype_str = 'float16' if fp16 else 'float32'
return FixedUnPooling(name,
x,
2,
unpool_mat=np.ones((2, 2), dtype=dtype_str),
data_format='channels_first')
def upsample2x(name, x):
return FixedUnPooling(
name,
x,
2,
unpool_mat=np.ones((2, 2), dtype='float32'),
data_format='channels_first')
def upsample2x(name, x):
try:
resize = tf.compat.v2.image.resize_images
with tf.name_scope(name):
shp2d = tf.shape(x)[2:]
x = tf.transpose(x, [0, 2, 3, 1])
x = resize(x, shp2d * 2, 'nearest')
x = tf.transpose(x, [0, 3, 1, 2])
return x
except AttributeError:
return FixedUnPooling(
name, x, 2, unpool_mat=np.ones((2, 2), dtype='float32'),
data_format='channels_first')
d4 = tf.stop_gradient(d4) if freeze else d4
d4 = Conv2D('conv_bot', d4, 1024, 1, padding='same')
return [d1, d2, d3, d4]
def atrous_spatial_pyramid_pooling(x, filters=64):
"""
ASPP layer: Dilated convolutions with rate tau = (6,12,18)
"""
pad = 'valid'
with tf.variable_scope('ASSP_layers'):
image_level_features = tf.reduce_mean(x, [2, 3], keep_dims=True)
image_level_features = Conv2D('image_level_features', image_level_features, filters, 1, strides=1, padding=pad, activation=BNReLU)
image_level_features = FixedUnPooling('unp_image_level_features', image_level_features, shape=(x.shape[2].value, x.shape[3].value), data_format='channels_first')
at_pool1x1 = Conv2D('at_pool1x1', x, filters, 1, strides=1, padding=pad, activation=BNReLU)
at_pool3x3_1 = Conv2D('at_pool3x3_1', x, filters, 3, strides=1, padding='same', activation=BNReLU, dilation_rate=6)
at_pool3x3_2 = Conv2D('at_pool3x3_2', x, filters, 3, strides=1, padding='same', activation=BNReLU, dilation_rate=12)
at_pool3x3_3 = Conv2D('at_pool3x3_3', x, filters, 3, strides=1, padding='same', activation=BNReLU, dilation_rate=18)
net = tf.concat((image_level_features, at_pool1x1, at_pool3x3_1, at_pool3x3_2, at_pool3x3_3), axis=1)
net = Conv2D('at_out', net, filters, 1, strides=1, padding=pad, activation=BNReLU)
return net
def decoder(name, i, use_assp=True):
pad = 'same' #Change to valid if non local
def upsample2x(x):
# TODO may not be optimal in speed or math
return FixedUnPooling(x, 2, data_format='channels_first')