def crnn_conv_layer(self, inputs, widths, is_training):
batch_norm_params = {
'is_training': is_training,
'decay': 0.9,
'updates_collections': None
}
with slim.arg_scope([slim.conv2d],
kernel_size=[3, 3],
weights_regularizer=slim.l2_regularizer(1e-4)):
with slim.arg_scope([slim.max_pool2d],
kernel_size=[2, 1],
stride=[2, 1],
padding='SAME'):
conv1 = slim.conv2d(inputs, 64, scope='conv1')
poo1 = slim.max_pool2d(conv1, scope='pool1')
conv2 = slim.conv2d(poo1, 128, scope='conv2')
pool2 = slim.max_pool2d(conv2, scope='pool2')
conv3 = slim.conv2d(pool2, 256, scope='conv3')
conv4 = slim.conv2d(conv3, 256, scope='conv4')
pool3 = slim.max_pool2d(conv1,
kernel_size=[2, 2],
stride=[2, 2],
scope='pool3')
conv5 = slim.conv2d(pool3,
512,
scope='conv5',
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params)
conv6 = slim.conv2d(conv5,
512,
scope='conv6',
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params)
pool4 = slim.max_pool2d(conv1,
kernel_size=[2, 2],
stride=[2, 2],
scope='pool4')
conv7 = slim.conv2d(pool4, 512, padding='SAME', scope='conv7')
features = tf.squeeze(conv7, axis=1, name='features')
conv1_trim = tf.constant(2 * (3 // 2),
dtype=tf.int32,
name='conv1_trim')
after_pool1 = tf.floor_div(widths, 2)
after_pool2 = tf.floor_div(after_pool1, 2)
after_pool3 = after_pool2
after_pool4 = after_pool3
after_conv7 = after_pool4 - conv1_trim
sequence_length = tf.reshape(after_conv7, [-1], name='seq_len')
sequence_length = tf.maximum(sequence_length, 1)
return features, sequence_length
def max_pooling(sigma_g, argmax, num_filters, new_size, image_size ):
argmax1= tf.transpose(argmax, [0, 3, 1, 2])
argmax2 = tf.reshape(argmax1,[1, num_filters, -1])#shape=[1, num_filters, new_size*new_size]
new_sigma_g = tf.reshape(sigma_g,[ num_filters*image_size*image_size,-1])
x_index = tf.mod(tf.floor_div(argmax2,tf.constant(num_filters ,shape=[1,num_filters, new_size*new_size], dtype='int64')),tf.constant(image_size ,shape=[1,num_filters, new_size*new_size], dtype='int64'))
aux = tf.floor_div(tf.floor_div(argmax2,tf.constant(num_filters,shape=[1,num_filters, new_size*new_size], dtype='int64')),tf.constant(image_size,shape=[1,num_filters, new_size*new_size], dtype='int64'))
y_index = tf.mod(aux,tf.constant(image_size,shape=[1,num_filters,new_size*new_size], dtype='int64'))
index = tf.multiply(y_index,image_size) + x_index
index = tf.squeeze(index) # shape=[num_filters,new_size*new_size]
for i in range(num_filters):
if(i==0):
ind1 = tf.gather(index, tf.constant(i))
new_ind = ind1
else:
ind1 = (image_size*image_size*i)+ tf.gather(index, tf.constant(i))
new_ind = tf.concat([new_ind,ind1],0) # shape=[num_filters*new_size*new_size]
column1 = tf.gather(new_sigma_g,new_ind)
column2 = tf.reshape(column1, [num_filters, new_size*new_size, -1])
column3 = tf.transpose(column2, [0, 2, 1])
column4 = tf.reshape(column3, [num_filters*image_size*image_size, -1])
final = tf.gather(column4,new_ind)
sigma_p = tf.reshape(final,[num_filters,new_size*new_size,new_size*new_size]) #shape=[num_filters,new_size*new_size, new_size*new_size]
return sigma_p
def pad_to_size(self, param_dict, tf_image, tf_label, *args):
# Pad image with zeros on all four sides to produce an output image of size [target_height, target_width]
# The input image will be approximately centered in the output image.
#
# input:
# param_dict dict with the following key-value pairs:
# 'target_height': height in pixels of output image (int)
# 'target_width': width in pixels of output image (int)
# 'constant': padding value (float)
#
target_height = param_dict['height']
target_width = param_dict['width']
constant_value = param_dict['constant']
with tf.name_scope('pad_to_size_' + '{:d}'.format(target_height) + '_' + '{:d}'.format(target_width)):
tf_image_shape = tf.shape(tf_image)
image_height = tf.cond(tf.equal(tf.rank(tf_image),4),lambda: tf_image_shape[1], lambda: tf_image_shape[0])
image_width = tf.cond(tf.equal(tf.rank(tf_image),4),lambda: tf_image_shape[2],lambda: tf_image_shape[1])
height_diff = target_height - image_height
width_diff = target_width - image_width
paddings = [[tf.floor_div(height_diff,2), height_diff - tf.floor_div(height_diff,2)],
[tf.floor_div(width_diff,2), width_diff - tf.floor_div(width_diff,2)],
[0,0]]
tf_image = tf.pad(tf_image, paddings, mode='CONSTANT', name=None, constant_values=constant_value)
return (tf_image, tf_label, *args)
def crop_to_size(self, param_dict, tf_image, tf_label, *args):
# Crop or pad image with zeros on all four sides to produce an output image of size [height, width]
#
# input:
# param_dict dict with the following key-value pairs:
# 'height': height in pixels of output image (int)
# 'width': width in pixels of output image (int)
#
target_height = param_dict['height']
target_width = param_dict['width']
with tf.name_scope('crop_to_size_' + '{:d}'.format(target_height) + '_' + '{:d}'.format(target_width)):
tf_image_shape = tf.shape(tf_image)
image_height = tf.cond(tf.equal(tf.rank(tf_image),4),lambda: tf_image_shape[1], lambda: tf_image_shape[0])
image_width = tf.cond(tf.equal(tf.rank(tf_image),4),lambda: tf_image_shape[2],lambda: tf_image_shape[1])
height_diff = image_height - target_height
width_diff = image_width - target_width
tf_image = tf.image.crop_to_bounding_box(tf_image,
tf.floor_div(height_diff,2),
tf.floor_div(width_diff,2),
target_height,
target_width)
return (tf_image, tf_label, *args)
def _parser(example):
zero = tf.zeros([1], dtype=tf.int64)
features = {
'image/encoded':
tf.FixedLenFeature((), tf.string, default_value=''),
'image/height':
tf.FixedLenFeature([1], tf.int64, default_value=zero),
'image/width':
tf.FixedLenFeature([1], tf.int64, default_value=zero),
'image/class':
tf.VarLenFeature(tf.int64),
}
res = tf.parse_single_example(example, features)
img = tf.image.decode_png(res['image/encoded'], channels=3)
original_w = tf.cast(res['image/width'][0], tf.int32)
original_h = tf.cast(res['image/height'][0], tf.int32)
img = tf.reshape(img, [original_h, original_w, 3])
w = tf.maximum(tf.cast(original_w, tf.float32), 1.0)
h = tf.maximum(tf.cast(original_h, tf.float32), 1.0)
ratio_w = tf.maximum(w / max_width, 1.0)
ratio_h = tf.maximum(h / 32.0, 1.0)
ratio = tf.maximum(ratio_w, ratio_h)
nw = tf.cast(tf.maximum(tf.floor_div(w, ratio), 1.0), tf.int32)
nh = tf.cast(tf.maximum(tf.floor_div(h, ratio), 1.0), tf.int32)
img = tf.image.resize_images(img, [nh, nw])
padw = tf.maximum(0, int(max_width) - nw)
padh = tf.maximum(0, 32 - nh)
img = tf.image.pad_to_bounding_box(img, 0, 0, nh + padh, nw + padw)
img = tf.cast(img, tf.float32) / 127.5 - 1
label = tf.sparse_tensor_to_dense(res['image/class'])
logging.info("Label: {}".format(label))
label = tf.reshape(label, [-1])
label = tf.cast(label, tf.int32)
logging.info("Label: {}".format(label))
return img, label
def meanDistance(y_true, y_pred):
in_shape = tf.shape(y_true)
# Flatten height/width dims
flat_true = tf.reshape(y_true, [in_shape[0], -1, in_shape[-1]])
flat_pred = tf.reshape(y_pred, [in_shape[0], -1, in_shape[-1]])
# Find peaks in linear indices
idx_true = tf.argmax(flat_true, axis=1)
idx_pred = tf.argmax(flat_pred, axis=1)
# Convert linear indices to subscripts
rows_true = tf.floor_div(tf.cast(idx_true, tf.int32), in_shape[2])
cols_true = tf.floormod(tf.cast(idx_true, tf.int32), in_shape[2])
rows_pred = tf.floor_div(tf.cast(idx_pred, tf.int32), in_shape[2])
cols_pred = tf.floormod(tf.cast(idx_pred, tf.int32), in_shape[2])
row_diff = tf.square(
tf.subtract(tf.cast(rows_true, tf.float32),
tf.cast(rows_pred, tf.float32)))
col_diff = tf.square(
tf.subtract(tf.cast(cols_true, tf.float32),
tf.cast(cols_pred, tf.float32)))
distances = tf.sqrt(tf.add(row_diff, col_diff))
return tf.reduce_mean(distances)
def convnet_layers(inputs, widths, mode, args):
"""Build convolutional network layers attached to the given input tensor"""
training = (mode == learn.ModeKeys.TRAIN)
# inputs should have shape [ ?, 32, ?, 1 ]
with tf.variable_scope("convnet"): # h,w
conv1 = conv_layer(inputs, layer_params[0], training) # 30,30
conv2 = conv_layer(conv1, layer_params[1], training) # 30,30
pool2 = pool_layer(conv2, 2, 'valid', 'pool2') # 15,15
conv3 = conv_layer(pool2, layer_params[2], training) # 15,15
conv4 = conv_layer(conv3, layer_params[3], training) # 15,15
pool4 = pool_layer(conv4, 2, 'valid', 'pool4') # 7,14
conv5 = conv_layer(pool4, layer_params[4], training) # 7,14
# layer_params[4][3]+='2'
# conv5 = conv_layer(conv5, layer_params[4], training)
conv6 = conv_layer(conv5, layer_params[5], training) # 7,14
pool6 = pool_layer(conv6, 1, 'valid', 'pool6') # 3,13
conv7 = conv_layer(pool6, layer_params[6], training) # 3,13
# layer_params[6][3] += '2'
# conv7 = conv_layer(conv7, layer_params[6], training)
conv8 = conv_layer(conv7, layer_params[7], training) # 3,13
# pool8 = tf.layers.max_pooling2d(conv8, [3, 1], [3, 1],
# padding='valid', name='pool8') # 1,13
# features = tf.squeeze(pool8, axis=1, name='features') # squeeze row dim
print(conv8.shape)
features = tf.reshape(tf.transpose(conv8, [0, 2, 1, 3]), [args.batch_size, -1, 512 * 7])
Wcnn = tf.get_variable(
'cnn_weights',
[512*7, 1024],
initializer=tf.truncated_normal_initializer()
)
bcnn = tf.get_variable(
'cnn_bias',
1024,
initializer=tf.constant_initializer(),
)
features = tf.reshape(tf.matmul(tf.reshape(features,[-1,512*7]), Wcnn)+bcnn, [args.batch_size, -1, 1024])
features = tf.layers.dropout(features, rate=dropout_rate )
kernel_sizes = [params[1] for params in layer_params]
# Calculate resulting sequence length from original image widths
conv1_trim = tf.constant(2 * (kernel_sizes[0] // 2),
dtype=tf.int32,
name='conv1_trim')
one = tf.constant(1, dtype=tf.int32, name='one')
two = tf.constant(2, dtype=tf.int32, name='two')
after_conv1 = tf.subtract(widths, conv1_trim)
after_pool2 = tf.floor_div(after_conv1, two)
# after_pool4 = tf.subtract(after_pool2, one)
after_pool4 = tf.floor_div(after_pool2, two)
after_pool6 = tf.subtract(after_pool4, one)
after_pool8 = after_pool6
sequence_length = tf.reshape(after_pool8, [-1], name='seq_len') # Vectorize
print(sequence_length)
return features, sequence_length
def convnet_layers(inputs, widths, mode):
"""Build convolutional network layers attached to the given input tensor"""
training = (mode == learn.ModeKeys.TRAIN)
# inputs should have shape [ ?, 32, ?, 1 ]
with tf.variable_scope("convnet"): # h,w
conv0 = conv_layer(inputs, layer_params[-2], training) # 63,63
conv02 = conv_layer(conv0, layer_params[-1], training) # 63,63
pool0 = pool_layer(conv02, 2, 'valid', 'pool0')
conv1 = conv_layer(pool0, layer_params[0], training) # 30,30
conv2 = conv_layer(conv1, layer_params[1], training) # 30,30
pool2 = pool_layer(conv2, 2, 'valid', 'pool2') # 15,15
conv3 = conv_layer(pool2, layer_params[2], training) # 15,15
conv4 = conv_layer(conv3, layer_params[3], training) # 15,15
pool4 = pool_layer(conv4, 2, 'valid', 'pool4') # 7,14
conv5 = conv_layer(pool4, layer_params[4], training) # 7,14
conv6 = conv_layer(conv5, layer_params[5], training) # 7,14
pool6 = pool_layer(conv6, 1, 'valid', 'pool6') # 3,13
conv7 = conv_layer(pool6, layer_params[6], training) # 3,13
conv8 = conv_layer(conv7, layer_params[7], training) # 3,13
print(conv8.shape)
pool8 = tf.layers.max_pooling2d(conv8, [3, 1], [3, 1],
padding='valid',
name='pool8') # 1,13
pool8 = tf.reshape(tf.transpose(conv8, [0, 2, 1, 3]),
[32, -1, 512 * 3])
print(pool8.shape)
features = tf.squeeze(pool8, axis=1,
name='features') # squeeze row dim
kernel_sizes = [params[1] for params in layer_params]
# Calculate resulting sequence length from original image widths
conv1_trim = tf.constant(2 * (kernel_sizes[0] // 2),
dtype=tf.int32,
name='conv1_trim')
one = tf.constant(1, dtype=tf.int32, name='one')
two = tf.constant(2, dtype=tf.int32, name='two')
after_conv0 = tf.subtract(widths, conv1_trim)
after_pool0 = tf.floor_div(after_conv0, two)
after_conv1 = tf.subtract(after_pool0, conv1_trim)
after_pool2 = tf.floor_div(after_conv1, two)
#after_pool2 = tf.subtract( after_conv1, one )
#after_pool4 = tf.subtract(after_pool2, one)
after_pool4 = tf.floor_div(after_pool2, two)
after_pool6 = tf.subtract(after_pool4, one)
#after_pool6 = tf.floor_div(after_pool4, two)
after_pool8 = after_pool6
print('fuckkkkkkk')
sequence_length = tf.reshape(after_pool8, [-1],
name='seq_len') # Vectorize
return features, sequence_length
def __makeMeshgrids(self, nx, ny, width, height):
x_num_between = tf.floor_div(width, nx) - 1
y_num_between = tf.floor_div(height, ny) - 1
x_step = 1 / tf.floor_div(width, nx)
y_step = 1 / tf.floor_div(height, ny)
x_range = tf.range(0., nx + x_step * x_num_between, x_step)[:width]
x_range = tf.clip_by_value(x_range, 0., nx - 1)
y_range = tf.range(0., ny + y_step * y_num_between, y_step)[:height]
y_range = tf.clip_by_value(y_range, 0., ny - 1)
xx, yy = tf.meshgrid(x_range, y_range)
return xx, yy
def _loadImageFunction(self, filename, color, weight):
if self.randomize_size:
max_upper_dev = int((self.target_size * 0.3) // self.size_factor)
max_lower_dev = -int((self.target_size * 0.3) // self.size_factor)
delta_size = tf.random_uniform(shape=(), minval=max_lower_dev, maxval=max_upper_dev,
dtype=tf.int32) * self.size_factor
else:
delta_size = 0
image_size = self.target_size + delta_size
image_string = tf.read_file(filename)
image_decoded = tf.image.decode_png(image_string, 3)
image_decoded.set_shape([None, None, 3])
image_decoded = tf.cast(image_decoded, tf.float32)
image_decoded = image_decoded / 255.0
if self.random_crop:
target_size = tf.cast(tf.multiply(tf.cast(image_size, dtype=tf.float32), 1.3), dtype=tf.int32)
image_decoded = aspect_preserving_resize(image_decoded, target_size=target_size, resize_mode="crop")
image_decoded = tf.random_crop(image_decoded, [image_size, image_size, 3])
else:
image_decoded = aspect_preserving_resize(image_decoded, target_size=image_size, resize_mode="pad")
if self.crop_to_size_factor:
resized_size = tf.shape(image_decoded)[:2]
target_crop_size = (resized_size // self.size_factor) * self.size_factor
image_decoded = image_decoded[:target_crop_size[0], :target_crop_size[1], :]
if self.random_brightness:
image_decoded = tf.image.random_brightness(image_decoded, max_delta=0.2)
if self.random_contrast:
image_decoded = tf.image.random_contrast(image_decoded, lower=.9, upper=1.1)
if self.random_saturation:
image_decoded = tf.image.random_saturation(image_decoded, lower=.9, upper=1.1)
# image_decoded = tf.py_func(self._py_read_image, [filename], tf.uint8)
image_shape = tf.shape(image_decoded)
if self.square_pad:
new_image_shape = tf.stack([tf.reduce_max(image_shape), tf.reduce_max(image_shape)])
image_decoded = tf.image.resize_image_with_crop_or_pad(image_decoded, new_image_shape[0],
new_image_shape[1])
elif self.crop_to_size_factor:
new_image_shape = image_shape
else:
new_image_shape = (tf.floor_div(image_shape, self.size_factor) +
tf.cast(tf.greater(tf.floormod(image_shape, self.size_factor), 0),
dtype=tf.int32)) * self.size_factor
image_decoded = tf.image.resize_image_with_crop_or_pad(image_decoded, new_image_shape[0],
new_image_shape[1])
offset_y = (new_image_shape[0] - image_shape[0]) // 2
offset_x = (new_image_shape[1] - image_shape[1]) // 2
if self.random_flip:
image_decoded = tf.image.random_flip_left_right(image_decoded)
image_decoded = (image_decoded * 2.0) - 1.0
return image_decoded, color, weight, filename, [offset_y, offset_x], image_shape
def tf_find_peaks(x):
""" Finds the maximum value in each channel and returns the location and value.
Args:
x: rank-4 tensor (samples, height, width, channels)
Returns:
peaks: rank-3 tensor (samples, [x, y, val], channels)
"""
# Store input shape
in_shape = tf.shape(x)
# Flatten height/width dims
flattened = tf.reshape(x, [in_shape[0], -1, in_shape[-1]])
# Find peaks in linear indices
idx = tf.argmax(flattened, axis=1)
# Convert linear indices to subscripts
rows = tf.floor_div(tf.cast(idx,tf.int32), in_shape[1])
cols = tf.floormod(tf.cast(idx,tf.int32), in_shape[1])
# Dumb way to get actual values without indexing
vals = tf.reduce_max(flattened, axis=1)
# Return N x 3 x C tensor
return tf.stack([
tf.cast(cols, tf.float32),
tf.cast(rows, tf.float32),
vals
], axis=1)
def upsampel_impl(now_count, need_count):
# sample with replacement
left_count = need_count - now_count
select_indices = tf.random_shuffle(tf.range(now_count))[:tf.floormod(left_count, now_count)]
select_indices = tf.concat([tf.tile(tf.range(now_count), [tf.floor_div(left_count, now_count) + 1]), select_indices], axis = 0)
return select_indices
def Convnet_8(inputs, widths, is_training):
with tf.variable_scope('convnet') as scope:
conv1 = conv_layer(inputs, layer_params[0], is_training) # 30,30
conv2 = conv_layer(conv1, layer_params[1], is_training) # 30,30
pool2 = pool_layer(conv2, 2, 'valid', 'pool2') # 15,15
conv3 = conv_layer(pool2, layer_params[2], is_training) # 15,15
conv4 = conv_layer(conv3, layer_params[3], is_training) # 15,15
pool4 = pool_layer(conv4, 1, 'valid', 'pool4') # 7,14
conv5 = conv_layer(pool4, layer_params[4], is_training) # 7,14
conv6 = conv_layer(conv5, layer_params[5], is_training) # 7,14
pool6 = pool_layer(conv6, 1, 'valid', 'pool6') # 3,13
conv7 = conv_layer(pool6, layer_params[6], is_training) # 3,13
conv8 = conv_layer(conv7, layer_params[7], is_training) # 3,13
pool8 = tf.layers.max_pooling2d(conv8, [3, 1], [3, 1],
padding='valid',
name='pool8') # 1,13
features = tf.squeeze(pool8, axis=1, name='features') #squeeze row dim
kernel_sizes = [params[1] for params in layer_params]
# Calculate resulting sequence length from original image widths
conv1_trim = tf.constant(2 * (kernel_sizes[0] // 2),
dtype=tf.int32,
name='conv1_trim')
one = tf.constant(1, dtype=tf.int32, name='one')
two = tf.constant(2, dtype=tf.int32, name='two')
after_conv1 = tf.subtract(widths, conv1_trim)
after_pool2 = tf.floor_div(after_conv1, two)
after_pool4 = tf.subtract(after_pool2, one)
after_pool6 = tf.subtract(after_pool4, one)
after_pool8 = after_pool6
sequence_length = \
tf.reshape(after_pool8, [-1], name='seq_len') # Vectorize
return features, sequence_length, rnn_size
def minibatch():
scope.reuse_variables()
remainder = tf.mod(leng, maxbatch, name="remainder")
splits = tf.identity(tf.floor_div(leng - remainder, maxbatch),
"splits")
remainder_inp = tf.slice(inp, [
leng - remainder if i == 0 else 0
for i in range(len(inp.shape))
], [-1 for i in range(len(inp.shape))])
majority_inp = tf.slice(inp,
[0 for i in range(len(inp.shape))], [
leng - remainder if i == 0 else -1
for i in range(len(inp.shape))
])
split_inp = tf.reshape(
majority_inp,
tf.concat([[splits, maxbatch],
tf.shape(inp)[1:]], 0))
majority_out = tf.map_fn(fn, split_inp)
scope.reuse_variables()
remainder_out = fn(remainder_inp)
out = tf.concat([
tf.reshape(
majority_out,
tf.concat([[leng - remainder],
tf.shape(majority_out)[2:]], 0)),
remainder_out
], 0)
if inp.shape[0].value is not None:
out = tf.reshape(
out,
tf.concat([[int(inp.shape[0])],
tf.shape(out)[1:]], 0))
return out
def shard_tensor(tensor, shard_limit, exact_sharding=False):
'''
Reshape a tensor s.t. it's first axis represents manageable
shards over which computations can be run seperately.
Unless indicated otherwise, the tensor will be broken up into a
set of shards and a remainder.
'''
shape = tf.shape(tensor)
num_slices = shape[0] #num. slices along the first axis
shard_size = tf.minimum(num_slices, shard_limit)
# Split up X into mapped shards and a remainder
if exact_sharding:
num_shards = tf.floor_div(num_slices, shard_size)
num_mapped = num_slices
mapped_part, remainder = tensor, None
else:
num_shards = tf.cast(tf.ceil(num_slices / shard_size) - 1,
dtype=shard_size.dtype)
num_mapped = num_shards * shard_size #num. slices computed via map_fn
partitions = [num_mapped, num_slices - num_mapped]
mapped_part, remainder = tf.split(tensor, partitions)
# Reshape mapped part of tensor into (n+1)-dimensional shards
sharding = tf.concat([num_shards[None], shard_size[None], shape[1:]], 0)
shards = tf.reshape(mapped_part, sharding)
return shards, remainder
def matrix_other():
isess = tf.InteractiveSession()
X = tf.Variable(tf.eye(3))
W = tf.Variable(tf.random_normal(shape=(3, 3)))
X.initializer.run()
W.initializer.run()
logger.info("X\n%s" % X.eval())
logger.info("W\n%s" % W.eval())
logger.info("tf.div(X,W)\n%s" % tf.div(X, W).eval())
logger.info("tf.truediv(X,W)\n%s" % tf.truediv(X, W).eval())
logger.info("tf.floordiv(X,W)\n%s" % tf.floordiv(X, W).eval())
logger.info("tf.realdiv(X,W)\n%s" % tf.realdiv(X, W).eval())
# logger.info("tf.truncatediv(X,W)\n%s" % tf.truncatediv(X, W).eval())
logger.info("tf.floor_div(X,W)\n%s" % tf.floor_div(X, W).eval())
logger.info("tf.truncatemod(X,W)\n%s" % tf.truncatemod(X, W).eval())
logger.info("tf.floormod(X,W)\n%s" % tf.floormod(X, W).eval())
logger.info("tf.cross(X,W)\n%s" % tf.cross(X, W).eval())
logger.info("tf.add_n(X,W)\n%s" % tf.add_n([X, W]).eval())
logger.info("tf.squared_difference(X,W)\n%s" %
tf.squared_difference(X, W).eval())
isess.close()
def __init__(self,
learning_rate_d=0.0004,
learning_rate_g=0.0001,
p_lambda=10.0,
p_gamma=1.0,
epsilon=0.001,
z_length=512,
n_imgs=800000,
lipschitz_penalty=True,
args=None
):
self.channels = [512, 512, 512, 256, 128, 64, 32, 16]
self.batch_size = [256, 256, 256, 128, 64, 32, 16, 8]
self.learning_rate_d = learning_rate_d
self.learning_rate_g = learning_rate_g
self.p_lambda = p_lambda
self.p_gamma = p_gamma
self.epsilon = epsilon
self.z_length = z_length
self.n_imgs = n_imgs
self.lipschitz_penalty = lipschitz_penalty
self.z = tf.placeholder(tf.float32, [None, self.z_length])
self.channel_num = 6
self.class_num = 128
self.input_length = 512
self.sampling = args.sample
self.record = args.record
with tf.variable_scope('image_count'):
self.total_imgs = tf.Variable(0.0, name='image_step', trainable=False)
self.img_count_placeholder = tf.placeholder(tf.float32)
self.img_step_op = tf.assign(self.total_imgs,
tf.add(self.total_imgs, self.img_count_placeholder))
self.img_step = tf.mod(tf.add(self.total_imgs, self.n_imgs), self.n_imgs * 2)
self.alpha = tf.minimum(1.0, tf.div(self.img_step, self.n_imgs))
self.layer = tf.floor_div(tf.add(self.total_imgs, self.n_imgs), self.n_imgs * 2)
self.get_dataset = self.make_dataset()
self.x, self.label = self.next_batch()
self.g_optimizer = tf.train.AdamOptimizer(learning_rate=self.learning_rate_g, beta1=0, beta2=0.9)
self.d_optimizer = tf.train.AdamOptimizer(learning_rate=self.learning_rate_d, beta1=0, beta2=0.9)
self.n_layers = 7
self.global_step = tf.train.get_or_create_global_step()
self.networks = [self._create_network(i + 1) for i in range(self.n_layers)]
print('Networks set.')
self.GPU_options = tf.GPUOptions(allow_growth=True, allocator_type='BFC')
self.config = tf.ConfigProto(allow_soft_placement=True, gpu_options=self.GPU_options)
self.sess = tf.Session(config=self.config)
self.sess.run(tf.global_variables_initializer())
self.writer = tf.summary.FileWriter(logdir='Logs_v2', graph=self.sess.graph)
self.saver = tf.train.Saver()
if tf.train.latest_checkpoint('Checkpoints_v2') is not None:
print('Restoring...')
self.saver.restore(self.sess, tf.train.latest_checkpoint('Checkpoints_v2'))
print('Completely restored.')
print('Initialization complete.')
def tf_find_peaks(x):
""" Finds the maximum value in each channel and returns the location and value.
Args:
x: rank-4 tensor (samples, height, width, channels)
Returns:
peaks: rank-3 tensor (samples, [x, y, val], channels)
"""
# Store input shape
in_shape = tf.shape(x)
# Flatten height/width dims
flattened = tf.reshape(x, [in_shape[0], -1, in_shape[-1]])
# Find peaks in linear indices
idx = tf.argmax(flattened, axis=1)
# Convert linear indices to subscripts
rows = tf.floor_div(tf.cast(idx, tf.int32), in_shape[1])
cols = tf.floormod(tf.cast(idx, tf.int32), in_shape[1])
# Dumb way to get actual values without indexing
vals = tf.reduce_max(flattened, axis=1)
# Return N x 3 x C tensor
return tf.stack(
[tf.cast(cols, tf.float32),
tf.cast(rows, tf.float32), vals], axis=1)
def add_coordinates(self, input):
shape = tf.shape(input)
positions = tf.range(shape[0] * shape[1] * shape[2])
position_embedding = tf.reshape(positions,
shape=[shape[0], shape[1] * shape[2]])
position_embedding = position_embedding % (shape[1] * shape[2])
position_embedding = tf.cast(position_embedding, tf.float32)
#
x_embedding = tf.cast(
tf.floor_div(position_embedding, tf.cast(shape[2], tf.float32)),
tf.int32)
y_embedding = tf.cast(
tf.cast(position_embedding, tf.int32) % shape[2], tf.int32)
pe_x = tf.nn.embedding_lookup(self.position_embedding_x, x_embedding)
pe_y = tf.nn.embedding_lookup(self.position_embedding_y, y_embedding)
pe_x = tf.reshape(pe_x,
shape=[shape[0], shape[1], shape[2], self.pe_dim])
pe_y = tf.reshape(pe_y,
shape=[shape[0], shape[1], shape[2], self.pe_dim])
final_pe = tf.concat([pe_x, pe_y], axis=-1)
return input + final_pe
def _padding(tensor, out_size):
t_width = tensor.get_shape()[1]
delta = tf.subtract(out_size, t_width)
pad_left = tf.floor_div(delta, 2)
pad_right = delta - pad_left
return tf.pad(
tensor, [[0, 0], [pad_left, pad_right], [pad_left, pad_right], [0, 0]],
'CONSTANT')
def test_floor_div(self):
shape = [3, 4, 5]
graph = tf.Graph()
with graph.as_default() as g:
a = tf.placeholder(tf.float32, shape=shape, name='a')
b = tf.placeholder(tf.float32, shape=shape, name='b')
out = tf.floor_div(a, b)
self._test_tf_model_constant(graph, {'a': shape, 'b': shape}, [out.op.name])
def gray2color(gray, spectrum=Spectrum.Color):
indices = tf.floor_div(gray, 64)
t = tf.expand_dims((gray - indices * 64) / 64, axis=-1)
indices = tf.cast(indices, dtype=tf.int32)
return tf.add(tf.multiply(tf.gather(spectrum, indices), 1 - t),
tf.multiply(tf.gather(spectrum, indices + 1), t))
def do_meshing_scores(start_scores, end_scores):
""" start_scores: [B, N, TP]
end_scores: [B, N, TP]
"""
TP = tf.shape(start_scores)[2]
#
# mesh
s_tiled = tf.tile(tf.expand_dims(start_scores, 3), [1, 1, 1, TP])
e_tiled = tf.tile(tf.expand_dims(end_scores, 2), [1, 1, TP, 1])
span_probs = s_tiled * e_tiled
#
# 取右上三角
# span_probs = tf.linalg.band_part(span_probs, 0, -1) # [B, NP, TP, TP]
mask = tf.linalg.band_part(tf.ones_like(span_probs), 0, -1)
span_probs = span_probs + 1e30 * (mask - 1.0)
#
# reshape & norm
shape_probs = tf.shape(span_probs)
B = shape_probs[0]
NP = shape_probs[1]
# TP = shape_probs[2]
#
span_probs_reshaped = tf.reshape(span_probs, [B, -1]) # [B, N*T*T]
span_probs_reshaped = tf.nn.softmax(span_probs_reshaped, -1)
#
span_probs_normed = tf.reshape(span_probs_reshaped,
[B, NP, TP, TP]) # [B, N, T, T]
#
#
# 找出最大位置
posi_1d = tf.argmax(span_probs_reshaped, -1, output_type=tf.int32)
#
# parse
TP2 = TP * TP
idx_passage = tf.floor_div(posi_1d, TP2)
#
posi_text = posi_1d - idx_passage * TP2
idx_start = tf.floor_div(posi_text, TP)
#
idx_end = posi_text - idx_start * TP
#
#
return span_probs_normed, idx_passage, idx_start, idx_end
def psf2otf(psf, img_shape):
# shape and type of the point spread function(s)
psf_shape = tf.shape(psf)
psf_type = psf.dtype
# coordinates for 'cutting up' the psf tensor
midH = tf.floor_div(psf_shape[0], 2)
midW = tf.floor_div(psf_shape[1], 2)
# slice the psf tensor into four parts
top_left = psf[:midH, :midW, :, :]
top_right = psf[:midH, midW:, :, :]
bottom_left = psf[midH:, :midW, :, :]
bottom_right = psf[midH:, midW:, :, :]
# prepare zeros for filler
zeros_bottom = tf.zeros([
psf_shape[0] - midH, img_shape[1] - psf_shape[1], psf_shape[2],
psf_shape[3]
],
dtype=psf_type)
zeros_top = tf.zeros(
[midH, img_shape[1] - psf_shape[1], psf_shape[2], psf_shape[3]],
dtype=psf_type)
# construct top and bottom row of new tensor
top = tf.concat([bottom_right, zeros_bottom, bottom_left], 1)
bottom = tf.concat([top_right, zeros_top, top_left], 1)
# prepare additional filler zeros and put everything together
zeros_mid = tf.zeros([
img_shape[0] - psf_shape[0], img_shape[1], psf_shape[2], psf_shape[3]
],
dtype=psf_type)
pre_otf = tf.concat([top, zeros_mid, bottom], 0)
# output shape: [img_shape[0], img_shape[1], channels_in, channels_out]
# fast fourier transform, transposed because tensor must have shape [..., height, width] for this
otf = tf.fft2d(
tf.cast(tf.transpose(pre_otf, perm=[2, 3, 0, 1]), tf.complex64))
# output shape: [channels_in, channels_out, img_shape[0], img_shape[1]]
return otf
def call(self, inputs, mask=None):
padded_inputs, adjustments, observations, blur_kernels, lambdas = inputs
imagesize = tf.shape(padded_inputs)[1:3]
kernelsize = tf.shape(blur_kernels)[1:3]
padding = tf.floor_div(kernelsize, 2)
mask_int = tf.ones(
(imagesize[0] - 2 * padding[0], imagesize[1] - 2 * padding[1]),
dtype=tf.float32)
mask_int = tf.pad(mask_int,
[[padding[0], padding[0]], [padding[1], padding[1]]],
mode='CONSTANT')
mask_int = tf.expand_dims(mask_int, 0)
filters = tf.matmul(self.B, self.filter_weights)
filters = tf.reshape(
filters,
[self.filter_size[0], self.filter_size[1], 1, self.nb_filters])
filter_otfs = psf2otf(filters, imagesize)
otf_term = tf.reduce_sum(tf.square(tf.abs(filter_otfs)), axis=1)
k = tf.expand_dims(tf.transpose(blur_kernels, [1, 2, 0]), -1)
k_otf = psf2otf(k, imagesize)[:, 0, :, :]
if self.stage > 1:
# boundary adjustment
Kx_fft = tf.fft2d(tf.cast(padded_inputs[:, :, :, 0],
tf.complex64)) * k_otf
Kx = tf.to_float(tf.ifft2d(Kx_fft))
Kx_outer = (1.0 - mask_int) * Kx
y_inner = mask_int * observations[:, :, :, 0]
y_adjusted = y_inner + Kx_outer
dataterm_fft = tf.fft2d(tf.cast(y_adjusted,
tf.complex64)) * tf.conj(k_otf)
else:
# standard data term
observations_fft = tf.fft2d(
tf.cast(observations[:, :, :, 0], tf.complex64))
dataterm_fft = observations_fft * tf.conj(k_otf)
lambdas = tf.expand_dims(lambdas, -1)
adjustment_fft = tf.fft2d(
tf.cast(adjustments[:, :, :, 0], tf.complex64))
numerator_fft = tf.cast(lambdas,
tf.complex64) * dataterm_fft + adjustment_fft
KtK = tf.square(tf.abs(k_otf))
denominator_fft = lambdas * KtK + otf_term
denominator_fft = tf.cast(denominator_fft, tf.complex64)
frac_fft = numerator_fft / denominator_fft
return tf.expand_dims(tf.to_float(tf.ifft2d(frac_fft)), -1)
def _modify_state(self):
# shuffle all states
self.states = tf.random_shuffle(self.states)
# slice off whatever won't fit in the the reshape
self.num_slices = tf.floor_div(self.k, self._part_sz)
if self._dim == 3:
self.states = tf.slice(
self.states, [0, 0, 0, 0, 0], [])
self.states = tf.transpose(
self.states, [0, 5, 2, 3, 4, 1])
def GET_dataset(dataset_name, dataset_dir, batch_size, preprocessing_name, split):
if split == 'train':
sff = True
threads = 4
is_training = True
else:
sff = False
threads = 1
is_training = False
with tf.variable_scope('dataset_%s'%split):
dataset = dataset_factory.get_dataset(dataset_name, split, dataset_dir)
with tf.device('/device:CPU:0'):
if split == 'train':
global_step = slim.create_global_step()
p = tf.floor_div(tf.cast(global_step, tf.float32), tf.cast(int(dataset.num_samples / float(batch_size)), tf.float32))
else:
global_step = None
p = None
provider = slim.dataset_data_provider.DatasetDataProvider(dataset,
shuffle=sff,
num_readers = threads,
common_queue_capacity=dataset.num_samples,
common_queue_min=0)
images, labels = provider.get(['image', 'label'])
image_preprocessing_fn = preprocessing_factory.get_preprocessing(preprocessing_name, is_training)
images = image_preprocessing_fn(images)
if split == 'train':
batch_images, batch_labels = tf.train.shuffle_batch([images, labels],
batch_size = batch_size,
num_threads = threads,
capacity = dataset.num_samples,
min_after_dequeue = 0)
with tf.variable_scope('1-hot_encoding'):
batch_labels = slim.one_hot_encoding(batch_labels, dataset.num_classes,on_value=1.0)
batch_queue = slim.prefetch_queue.prefetch_queue([batch_images, batch_labels], capacity=40*batch_size)
image, label = batch_queue.dequeue()
else:
batch_images, batch_labels = tf.train.batch([images, labels],
batch_size = batch_size,
num_threads = threads,
capacity = dataset.num_samples)
with tf.variable_scope('1-hot_encoding'):
batch_labels = slim.one_hot_encoding(batch_labels, dataset.num_classes,on_value=1.0)
batch_queue = slim.prefetch_queue.prefetch_queue([batch_images, batch_labels], capacity=8*batch_size)
image, label = batch_queue.dequeue()
return p, global_step, dataset, image, label
def evaluate(self, data_loader):
with tf.device('/cpu:0'):
input_images, input_labels, input_widths = data_loader.read_with_bucket_queue(
batch_size=cfg.TEST.BATCH_SIZE,
num_threads=cfg.TEST.THREADS,
num_epochs=1,
shuffle=False)
with tf.device('/gpu:0'):
logits = get_models(cfg.MODEL.BACKBONE)(cfg.MODEL.NUM_CLASSES).build(input_images, False)
seqlen = tf.cast(tf.floor_div(input_widths, 2), tf.int32, name='sequence_length')
softmax = tf.nn.softmax(logits, dim=-1, name='softmax')
decoded, log_prob = tf.nn.ctc_greedy_decoder(softmax, seqlen)
distance = tf.reduce_mean(tf.edit_distance(tf.cast(decoded[0], tf.int32), input_labels))
saver = tf.train.Saver(tf.global_variables())
gpu_options = tf.GPUOptions(allow_growth=True)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options, allow_soft_placement=True)) as sess:
saver.restore(sess, tf.train.latest_checkpoint(self.output_dir))
sess.run(tf.local_variables_initializer())
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
try:
cnt = 0
dm = 1e-5
while not coord.should_stop():
dt = sess.run([distance, ])[0]
cnt += 1
dm = (dm + dt) / cnt
if cfg.TEST.VIS:
dd, il, ii = sess.run([decoded, input_labels, input_images])
gts = self.decoder.sparse_to_strlist(il.indices, il.values, cfg.TEST.BATCH_SIZE)
pts = self.decoder.sparse_to_strlist(dd[0].indices, dd[0].values, cfg.TEST.BATCH_SIZE)
tb = PrettyTable()
tb.field_names = ['Index', 'GroundTruth', 'Predict', '{:.3f}/{:.3f}'.format(dt, dm)]
for i in range(len(gts)):
tb.add_row([i, gts[i], pts[i], ''])
print(tb)
else:
print('EditDistance: {:.3f}/{:.3f}'.format(dt, dm))
except tf.errors.OutOfRangeError:
print('Epochs Complete!')
finally:
coord.request_stop()
coord.join(threads)
def base_conv_layer(self, inputs, widths, is_training):
batch_norm_params = {'is_training': is_training, 'decay': 0.9
, 'updates_collections': None}
with slim.arg_scope([slim.conv2d], kernel_size=[3, 3], weights_regularizer=slim.l2_regularizer(1e-4),
normalizer_fn=slim.batch_norm, normalizer_params=batch_norm_params):
with slim.arg_scope([slim.max_pool2d], kernel_size=[2, 2], stride=[2, 1], padding='VALID'):
conv1 = slim.conv2d(inputs, 64, padding='VALID', scope='conv1')
conv2 = slim.conv2d(conv1, 64, scope='conv2')
poo1 = slim.max_pool2d(conv2, kernel_size=[2, 2], stride=[2, 2], scope='pool1')
conv3 = slim.conv2d(poo1, 128, scope='conv3')
conv4 = slim.conv2d(conv3, 128, scope='conv4')
pool2 = slim.max_pool2d(conv4, scope='pool2')
conv5 = slim.conv2d(pool2, 256, scope='conv5')
conv6 = slim.conv2d(conv5, 256, scope='conv6')
pool3 = slim.max_pool2d(conv6, scope='pool3')
conv7 = slim.conv2d(pool3, 512, scope='conv7')
conv8 = slim.conv2d(conv7, 512, scope='conv8')
# pool4 = slim.max_pool2d(conv8, kernel_size=[3, 1], stride=[3, 1], scope='pool4')
conv9 = slim.conv2d(conv8,512,kernel_size = [2,2],stride=1,padding='VALID',scope='conv9')
features = tf.transpose(conv9,[0,2,1,3])
conv1_trim = tf.constant(2 * (3 // 2),
dtype=tf.int32,
name='conv1_trim')
after_conv1 = widths - conv1_trim
after_pool1 = tf.floor_div(after_conv1, 2)
after_pool2 = after_pool1 - 1
after_pool3 = after_pool2 - 1
after_conv9 = after_pool3 -1
after_conv9 = after_conv9 * 2
num = tf.argmax(after_conv9)
max_lenght = after_conv9[num]
features = tf.reshape(features,[-1,max_lenght,512],name='features')
# features = tf.squeeze(features, axis=1, name='features')
sequence_length = tf.reshape(after_conv9, [-1], name='seq_len')
sequence_length = tf.maximum(sequence_length, 1)
return features, sequence_length
def dwise_with_add(
self,
inputs,
kernel_size,
filter_num,
seq_lens,
strides,
padding,
is_add=True,
kernel_initializer=tf.contrib.layers.xavier_initializer(0.1),
dilation_rate=(1, 1),
activation_fn=None,
name=None):
if padding.lower() == 'valid':
k = (kernel_size[0] - 1) * dilation_rate[0] + 1
seq_lens = seq_lens - k + 1
new_seq_len = 1 + tf.floor_div((seq_lens - 1), strides[0])
bottlen_channel = inputs.get_shape().as_list()[-1]
output1 = tf.layers.conv2d(inputs,
filter_num,
kernel_size=[1, 1],
padding='same',
activation=activation_fn,
kernel_initializer=kernel_initializer)
with tf.variable_scope(name):
dwise_filter = tf.get_variable(
name='w_f',
shape=[kernel_size[0], kernel_size[1], filter_num, 1],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer(0.1))
output2 = tf.nn.depthwise_conv2d(
input=output1,
filter=dwise_filter,
strides=[1, strides[0], strides[1], 1],
padding=padding)
output2 = activation_fn(output2)
output3 = tf.layers.conv2d(output2,
bottlen_channel,
kernel_size=[1, 1],
padding='same',
activation=None,
kernel_initializer=kernel_initializer)
if is_add:
return tf.add(inputs, output3), new_seq_len
else:
return output3, new_seq_len
def jsma(model, x, target, nb_epoch=None, delta=1., clip_min=0., clip_max=1.):
if nb_epoch is None:
nb_epoch = tf.floor_div(tf.size(x), 20)
def _cond(x_adv, epoch):
ybar = tf.reshape(model(x_adv), [-1])
return tf.logical_and(tf.less(ybar[target], 0.9),
tf.less(epoch, nb_epoch))
def _body(x_adv, epoch):
y = model(x_adv)
nb_input = tf.size(x_adv)
nb_output = tf.size(y)
mask = tf.one_hot(target, nb_output, on_value=True, off_value=False)
mask = tf.expand_dims(mask, axis=0)
yt = tf.boolean_mask(y, mask)
yo = tf.boolean_mask(y, tf.logical_not(mask))
dt_dx, = tf.gradients(yt, x_adv)
do_dx, = tf.gradients(yo, x_adv)
score = -dt_dx * do_dx
cond1 = tf.cond(delta > tf.constant(0.), lambda: x_adv < clip_max,
lambda: x_adv > clip_min)
cond2 = tf.logical_and(dt_dx > 0, do_dx < 0)
ind = tf.where(tf.logical_and(cond1, cond2))
score = tf.gather_nd(score, ind)
p = tf.argmax(score, axis=0)
p = tf.gather(ind, p)
p = tf.expand_dims(p, axis=0)
p = tf.to_int32(p)
dx = tf.scatter_nd(p, [delta], tf.shape(x_adv))
x_adv = tf.stop_gradient(x_adv + dx)
if (clip_min is not None) and (clip_max is not None):
x_adv = tf.clip_by_value(x_adv, clip_min, clip_max)
epoch += 1
return x_adv, epoch
epoch = tf.Variable(0, tf.int32)
x_adv, epoch = tf.while_loop(_cond, _body, (x, epoch))
return x_adv
def convnet_layers(inputs, widths, mode):
"""Build convolutional network layers attached to the given input tensor"""
training = (mode == learn.ModeKeys.TRAIN)
# inputs should have shape [ ?, 32, ?, 1 ]
with tf.variable_scope("convnet"): # h,w
conv1 = conv_layer(inputs, layer_params[0], training ) # 30,30
conv2 = conv_layer( conv1, layer_params[1], training ) # 30,30
pool2 = pool_layer( conv2, 2, 'valid', 'pool2') # 15,15
conv3 = conv_layer( pool2, layer_params[2], training ) # 15,15
conv4 = conv_layer( conv3, layer_params[3], training ) # 15,15
pool4 = pool_layer( conv4, 1, 'valid', 'pool4' ) # 7,14
conv5 = conv_layer( pool4, layer_params[4], training ) # 7,14
conv6 = conv_layer( conv5, layer_params[5], training ) # 7,14
pool6 = pool_layer( conv6, 1, 'valid', 'pool6') # 3,13
conv7 = conv_layer( pool6, layer_params[6], training ) # 3,13
conv8 = conv_layer( conv7, layer_params[7], training ) # 3,13
pool8 = tf.layers.max_pooling2d( conv8, [3,1], [3,1],
padding='valid', name='pool8') # 1,13
features = tf.squeeze(pool8, axis=1, name='features') # squeeze row dim
kernel_sizes = [ params[1] for params in layer_params]
# Calculate resulting sequence length from original image widths
conv1_trim = tf.constant( 2 * (kernel_sizes[0] // 2),
dtype=tf.int32,
name='conv1_trim')
one = tf.constant(1, dtype=tf.int32, name='one')
two = tf.constant(2, dtype=tf.int32, name='two')
after_conv1 = tf.subtract( widths, conv1_trim)
after_pool2 = tf.floor_div( after_conv1, two )
after_pool4 = tf.subtract(after_pool2, one)
sequence_length = tf.reshape(after_pool4,[-1], name='seq_len') # Vectorize
return features,sequence_length
def sample_halton_sequence(dim,
num_results=None,
sequence_indices=None,
dtype=tf.float32,
randomized=True,
seed=None,
name=None):
r"""Returns a sample from the `dim` dimensional Halton sequence.
Warning: The sequence elements take values only between 0 and 1. Care must be
taken to appropriately transform the domain of a function if it differs from
the unit cube before evaluating integrals using Halton samples. It is also
important to remember that quasi-random numbers without randomization are not
a replacement for pseudo-random numbers in every context. Quasi random numbers
are completely deterministic and typically have significant negative
autocorrelation unless randomization is used.
Computes the members of the low discrepancy Halton sequence in dimension
`dim`. The `dim`-dimensional sequence takes values in the unit hypercube in
`dim` dimensions. Currently, only dimensions up to 1000 are supported. The
prime base for the k-th axes is the k-th prime starting from 2. For example,
if `dim` = 3, then the bases will be [2, 3, 5] respectively and the first
element of the non-randomized sequence will be: [0.5, 0.333, 0.2]. For a more
complete description of the Halton sequences see
[here](https://en.wikipedia.org/wiki/Halton_sequence). For low discrepancy
sequences and their applications see
[here](https://en.wikipedia.org/wiki/Low-discrepancy_sequence).
If `randomized` is true, this function produces a scrambled version of the
Halton sequence introduced by [Owen (2017)][1]. For the advantages of
randomization of low discrepancy sequences see [here](
https://en.wikipedia.org/wiki/Quasi-Monte_Carlo_method#Randomization_of_quasi-Monte_Carlo).
The number of samples produced is controlled by the `num_results` and
`sequence_indices` parameters. The user must supply either `num_results` or
`sequence_indices` but not both.
The former is the number of samples to produce starting from the first
element. If `sequence_indices` is given instead, the specified elements of
the sequence are generated. For example, sequence_indices=tf.range(10) is
equivalent to specifying n=10.
#### Examples
```python
import tensorflow as tf
import tensorflow_probability as tfp
# Produce the first 1000 members of the Halton sequence in 3 dimensions.
num_results = 1000
dim = 3
sample = tfp.mcmc.sample_halton_sequence(
dim,
num_results=num_results,
seed=127)
# Evaluate the integral of x_1 * x_2^2 * x_3^3 over the three dimensional
# hypercube.
powers = tf.range(1.0, limit=dim + 1)
integral = tf.reduce_mean(tf.reduce_prod(sample ** powers, axis=-1))
true_value = 1.0 / tf.reduce_prod(powers + 1.0)
with tf.Session() as session:
values = session.run((integral, true_value))
# Produces a relative absolute error of 1.7%.
print ("Estimated: %f, True Value: %f" % values)
# Now skip the first 1000 samples and recompute the integral with the next
# thousand samples. The sequence_indices argument can be used to do this.
sequence_indices = tf.range(start=1000, limit=1000 + num_results,
dtype=tf.int32)
sample_leaped = tfp.mcmc.sample_halton_sequence(
dim,
sequence_indices=sequence_indices,
seed=111217)
integral_leaped = tf.reduce_mean(tf.reduce_prod(sample_leaped ** powers,
axis=-1))
with tf.Session() as session:
values = session.run((integral_leaped, true_value))
# Now produces a relative absolute error of 0.05%.
print ("Leaped Estimated: %f, True Value: %f" % values)
```
Args:
dim: Positive Python `int` representing each sample's `event_size.` Must
not be greater than 1000.
num_results: (Optional) positive Python `int`. The number of samples to
generate. Either this parameter or sequence_indices must be specified but
not both. If this parameter is None, then the behaviour is determined by
the `sequence_indices`.
Default value: `None`.
sequence_indices: (Optional) `Tensor` of dtype int32 and rank 1. The
elements of the sequence to compute specified by their position in the
sequence. The entries index into the Halton sequence starting with 0 and
hence, must be whole numbers. For example, sequence_indices=[0, 5, 6] will
produce the first, sixth and seventh elements of the sequence. If this
parameter is None, then the `num_results` parameter must be specified
which gives the number of desired samples starting from the first sample.
Default value: `None`.
dtype: (Optional) The dtype of the sample. One of: `float16`, `float32` or
`float64`.
Default value: `tf.float32`.
randomized: (Optional) bool indicating whether to produce a randomized
Halton sequence. If True, applies the randomization described in
[Owen (2017)][1].
Default value: `True`.
seed: (Optional) Python integer to seed the random number generator. Only
used if `randomized` is True. If not supplied and `randomized` is True,
no seed is set.
Default value: `None`.
name: (Optional) Python `str` describing ops managed by this function. If
not supplied the name of this function is used.
Default value: "sample_halton_sequence".
Returns:
halton_elements: Elements of the Halton sequence. `Tensor` of supplied dtype
and `shape` `[num_results, dim]` if `num_results` was specified or shape
`[s, dim]` where s is the size of `sequence_indices` if `sequence_indices`
were specified.
Raises:
ValueError: if both `sequence_indices` and `num_results` were specified or
if dimension `dim` is less than 1 or greater than 1000.
#### References
[1]: Art B. Owen. A randomized Halton algorithm in R. _arXiv preprint
arXiv:1706.02808_, 2017. https://arxiv.org/abs/1706.02808
"""
if dim < 1 or dim > _MAX_DIMENSION:
raise ValueError(
'Dimension must be between 1 and {}. Supplied {}'.format(_MAX_DIMENSION,
dim))
if (num_results is None) == (sequence_indices is None):
raise ValueError('Either `num_results` or `sequence_indices` must be'
' specified but not both.')
if not dtype.is_floating:
raise ValueError('dtype must be of `float`-type')
with tf.name_scope(name, 'sample', values=[sequence_indices]):
# Here and in the following, the shape layout is as follows:
# [sample dimension, event dimension, coefficient dimension].
# The coefficient dimension is an intermediate axes which will hold the
# weights of the starting integer when expressed in the (prime) base for
# an event dimension.
indices = _get_indices(num_results, sequence_indices, dtype)
radixes = tf.constant(_PRIMES[0:dim], dtype=dtype, shape=[dim, 1])
max_sizes_by_axes = _base_expansion_size(tf.reduce_max(indices),
radixes)
max_size = tf.reduce_max(max_sizes_by_axes)
# The powers of the radixes that we will need. Note that there is a bit
# of an excess here. Suppose we need the place value coefficients of 7
# in base 2 and 3. For 2, we will have 3 digits but we only need 2 digits
# for base 3. However, we can only create rectangular tensors so we
# store both expansions in a [2, 3] tensor. This leads to the problem that
# we might end up attempting to raise large numbers to large powers. For
# example, base 2 expansion of 1024 has 10 digits. If we were in 10
# dimensions, then the 10th prime (29) we will end up computing 29^10 even
# though we don't need it. We avoid this by setting the exponents for each
# axes to 0 beyond the maximum value needed for that dimension.
exponents_by_axes = tf.tile([tf.range(max_size)], [dim, 1])
# The mask is true for those coefficients that are irrelevant.
weight_mask = exponents_by_axes >= max_sizes_by_axes
capped_exponents = tf.where(
weight_mask,
tf.zeros_like(exponents_by_axes),
exponents_by_axes)
weights = radixes ** capped_exponents
# The following computes the base b expansion of the indices. Suppose,
# x = a0 + a1*b + a2*b^2 + ... Then, performing a floor div of x with
# the vector (1, b, b^2, b^3, ...) will produce
# (a0 + s1 * b, a1 + s2 * b, ...) where s_i are coefficients we don't care
# about. Noting that all a_i < b by definition of place value expansion,
# we see that taking the elements mod b of the above vector produces the
# place value expansion coefficients.
coeffs = tf.floor_div(indices, weights)
coeffs *= 1. - tf.cast(weight_mask, dtype)
coeffs %= radixes
if not randomized:
coeffs /= radixes
return tf.reduce_sum(coeffs / weights, axis=-1)
seed = distributions_util.gen_new_seed(
seed, salt='mcmc_sample_halton_sequence_1')
coeffs = _randomize(coeffs, radixes, seed=seed)
# Remove the contribution from randomizing the trailing zero for the
# axes where max_size_by_axes < max_size. This will be accounted
# for separately below (using zero_correction).
coeffs *= 1. - tf.cast(weight_mask, dtype)
coeffs /= radixes
base_values = tf.reduce_sum(coeffs / weights, axis=-1)
# The randomization used in Owen (2017) does not leave 0 invariant. While
# we have accounted for the randomization of the first `max_size_by_axes`
# coefficients, we still need to correct for the trailing zeros. Luckily,
# this is equivalent to adding a uniform random value scaled so the first
# `max_size_by_axes` coefficients are zero. The following statements perform
# this correction.
seed = distributions_util.gen_new_seed(
seed, salt='mcmc_sample_halton_sequence_2')
zero_correction = tf.random_uniform([dim, 1], seed=seed, dtype=dtype)
zero_correction /= radixes ** max_sizes_by_axes
return base_values + tf.reshape(zero_correction, [-1])
def get_control_flag(control, field):
return tf.equal(tf.mod(tf.floor_div(control, field), 2), 1)
# writer = tf.summary.FileWriter('./graphs', sess.graph)
print(sess.run(x))
writer.close() # close the writer when you’re done using it
# Example 2: The wonderful wizard of div
a = tf.constant([2, 2], name='a')
b = tf.constant([[0, 1], [2, 3]], name='b')
with tf.Session() as sess:
print(sess.run(tf.div(b, a)))
print(sess.run(tf.divide(b, a)))
print(sess.run(tf.truediv(b, a)))
print(sess.run(tf.floordiv(b, a)))
# print(sess.run(tf.realdiv(b, a)))
print(sess.run(tf.truncatediv(b, a)))
print(sess.run(tf.floor_div(b, a)))
# Example 3: multiplying tensors
a = tf.constant([10, 20], name='a')
b = tf.constant([2, 3], name='b')
with tf.Session() as sess:
print(sess.run(tf.multiply(a, b)))
print(sess.run(tf.tensordot(a, b, 1)))
# Example 4: Python native type
t_0 = 19
x = tf.zeros_like(t_0) # ==> 0
y = tf.ones_like(t_0) # ==> 1
t_1 = ['apple', 'peach', 'banana']