def make_minibatch(self, valid_anchors):
with tf.variable_scope('rpn_minibatch'):
# in labels(shape is [N, ]): 1 is positive, 0 is negative, -1 is ignored
labels, anchor_matched_gtboxes, object_mask = \
self.rpn_find_positive_negative_samples(valid_anchors) # [num_of_valid_anchors, ]
positive_indices = tf.reshape(tf.where(tf.equal(labels, 1.0)), [-1]) # use labels is same as object_mask
num_of_positives = tf.minimum(tf.shape(positive_indices)[0],
tf.cast(self.rpn_mini_batch_size * self.rpn_positives_ratio, tf.int32))
# num of positives <= minibatch_size * 0.5
positive_indices = tf.random_shuffle(positive_indices)
positive_indices = tf.slice(positive_indices, begin=[0], size=[num_of_positives])
# positive_anchors = tf.gather(self.anchors, positive_indices)
negative_indices = tf.reshape(tf.where(tf.equal(labels, 0.0)), [-1])
num_of_negatives = tf.minimum(self.rpn_mini_batch_size - num_of_positives,
tf.shape(negative_indices)[0])
negative_indices = tf.random_shuffle(negative_indices)
negative_indices = tf.slice(negative_indices, begin=[0], size=[num_of_negatives])
# negative_anchors = tf.gather(self.anchors, negative_indices)
minibatch_indices = tf.concat([positive_indices, negative_indices], axis=0)
minibatch_indices = tf.random_shuffle(minibatch_indices)
minibatch_anchor_matched_gtboxes = tf.gather(anchor_matched_gtboxes, minibatch_indices)
object_mask = tf.gather(object_mask, minibatch_indices)
labels = tf.cast(tf.gather(labels, minibatch_indices), tf.int32)
labels_one_hot = tf.one_hot(labels, depth=2)
return minibatch_indices, minibatch_anchor_matched_gtboxes, object_mask, labels_one_hot
def __init__(self, tensors: List[tf.Tensor], cluster_indexes: tf.Tensor, n_splits, seed, train_sampling=1.0,
test_sampling=1.0):
size = tensors[0].shape[0].value
self.seed = seed
clustered_index = self.cluster_pages(cluster_indexes)
index_len = tf.shape(clustered_index)[0]
assert_op = tf.assert_equal(index_len, size, message='n_pages is not equals to size of clustered index')
with tf.control_dependencies([assert_op]):
split_nitems = int(round(size / n_splits))
split_size = [split_nitems] * n_splits
split_size[-1] = size - (n_splits - 1) * split_nitems
splits = tf.split(clustered_index, split_size)
complements = [tf.random_shuffle(tf.concat(splits[:i] + splits[i + 1:], axis=0), seed) for i in
range(n_splits)]
splits = [tf.random_shuffle(split, seed) for split in splits]
def mk_name(prefix, tensor):
return prefix + '_' + tensor.name[:-2]
def prepare_split(i):
test_size = split_size[i]
train_size = size - test_size
test_sampled_size = int(round(test_size * test_sampling))
train_sampled_size = int(round(train_size * train_sampling))
test_idx = splits[i][:test_sampled_size]
train_idx = complements[i][:train_sampled_size]
test_set = [tf.gather(tensor, test_idx, name=mk_name('test', tensor)) for tensor in tensors]
tran_set = [tf.gather(tensor, train_idx, name=mk_name('train', tensor)) for tensor in tensors]
return Split(test_set, tran_set, test_sampled_size, train_sampled_size)
self.splits = [prepare_split(i) for i in range(n_splits)]
def fast_rcnn_minibatch(self, reference_boxes):
with tf.variable_scope('fast_rcnn_minibatch'):
reference_boxes_mattached_gtboxes, object_mask, label = \
self.fast_rcnn_find_positive_negative_samples(reference_boxes)
positive_indices = tf.reshape(tf.where(tf.not_equal(object_mask, 0.)), [-1])
num_of_positives = tf.minimum(tf.shape(positive_indices)[0],
tf.cast(self.fast_rcnn_minibatch_size*self.fast_rcnn_positives_ratio, tf.int32))
positive_indices = tf.random_shuffle(positive_indices)
positive_indices = tf.slice(positive_indices, begin=[0], size=[num_of_positives])
negative_indices = tf.reshape(tf.where(tf.equal(object_mask, 0.)), [-1])
num_of_negatives = tf.minimum(tf.shape(negative_indices)[0],
self.fast_rcnn_minibatch_size - num_of_positives)
negative_indices = tf.random_shuffle(negative_indices)
negative_indices = tf.slice(negative_indices, begin=[0], size=[num_of_negatives])
minibatch_indices = tf.concat([positive_indices, negative_indices], axis=0)
minibatch_indices = tf.random_shuffle(minibatch_indices)
minibatch_reference_boxes_mattached_gtboxes = tf.gather(reference_boxes_mattached_gtboxes,
minibatch_indices)
object_mask = tf.gather(object_mask, minibatch_indices)
label = tf.gather(label, minibatch_indices)
label_one_hot = tf.one_hot(label, self.num_classes + 1)
return minibatch_indices, minibatch_reference_boxes_mattached_gtboxes, object_mask, label_one_hot
def cifar_filename_queue(filename_list):
# convert the list to a tensor
string_tensor = tf.convert_to_tensor(filename_list, dtype=tf.string)
# randomize the tensor
tf.random_shuffle(string_tensor)
# create the queue
fq = tf.FIFOQueue(capacity=10, dtypes=tf.string)
# create our enqueue_op for this q
fq_enqueue_op = fq.enqueue_many([string_tensor])
# create a QueueRunner and add to queue runner list
# we only need one thread for this simple queue
tf.train.add_queue_runner(tf.train.QueueRunner(fq, [fq_enqueue_op] * 1))
return fq
def subsample_indicator(indicator, num_samples):
"""Subsample indicator vector.
Given a boolean indicator vector with M elements set to `True`, the function
assigns all but `num_samples` of these previously `True` elements to
`False`. If `num_samples` is greater than M, the original indicator vector
is returned.
Args:
indicator: a 1-dimensional boolean tensor indicating which elements
are allowed to be sampled and which are not.
num_samples: int32 scalar tensor
Returns:
a boolean tensor with the same shape as input (indicator) tensor
"""
indices = tf.where(indicator)
indices = tf.random_shuffle(indices)
indices = tf.reshape(indices, [-1])
num_samples = tf.minimum(tf.size(indices), num_samples)
selected_indices = tf.slice(indices, [0], tf.reshape(num_samples, [1]))
selected_indicator = ops.indices_to_dense_vector(selected_indices,
tf.shape(indicator)[0])
return tf.equal(selected_indicator, 1)
def get_random_scale(min_scale_factor, max_scale_factor, step_size):
"""Gets a random scale value.
Args:
min_scale_factor: Minimum scale value.
max_scale_factor: Maximum scale value.
step_size: The step size from minimum to maximum value.
Returns:
A random scale value selected between minimum and maximum value.
Raises:
ValueError: min_scale_factor has unexpected value.
"""
if min_scale_factor < 0 or min_scale_factor > max_scale_factor:
raise ValueError('Unexpected value of min_scale_factor.')
if min_scale_factor == max_scale_factor:
return tf.cast(min_scale_factor, tf.float32)
# When step_size = 0, we sample the value uniformly from [min, max).
if step_size == 0:
return tf.random_uniform([1],
minval=min_scale_factor,
maxval=max_scale_factor)
# When step_size != 0, we randomly select one discrete value from [min, max].
num_steps = int((max_scale_factor - min_scale_factor) / step_size + 1)
scale_factors = tf.lin_space(min_scale_factor, max_scale_factor, num_steps)
shuffled_scale_factors = tf.random_shuffle(scale_factors)
return shuffled_scale_factors[0]
def get_svtcn_indices(seq_len, batch_size, num_views):
"""Gets a random window of contiguous time indices from a sequence.
Args:
seq_len: Int, number of timesteps in the image sequence.
batch_size: Int, size of the batch to construct.
num_views: Int, the number of simultaneous viewpoints at each
timestep in the dataset.
Returns:
time_indices: 1-D Int `Tensor` with size [batch_size], holding the
timestep for each batch image.
view_indices: 1-D Int `Tensor` with size [batch_size], holding the
view for each batch image. This is consistent across the batch.
"""
# Get anchor, positive time indices.
def f1():
# Choose a random contiguous range from within the sequence.
range_min = tf.random_shuffle(tf.range(seq_len-batch_size))[0]
range_max = range_min+batch_size
return tf.range(range_min, range_max)
def f2():
# Consider the full sequence.
return tf.range(seq_len)
time_indices = tf.cond(tf.greater(seq_len, batch_size), f1, f2)
# Get opposing anchor, positive view indices.
random_view = tf.random_shuffle(tf.range(num_views))[0]
view_indices = tf.tile([random_view], (batch_size,))
return time_indices, view_indices
def _build_graph(self):
"""Construct tensorflow nodes for round of clustering"""
# N.B. without tf.Variable, makes awesome glitchy clustered images
self.centroids_in = tf.Variable(tf.slice(tf.random_shuffle(self.arr),
[0, 0], [self.k, -1]), name="centroids_in")
# tiled should be shape(self.n_pixels, self.k, size_data = 2 + self.channels)
tiled_pix = tf.tile(tf.expand_dims(self.arr, 1),
multiples=[1, self.k, 1], name="tiled_pix")
# no need to take square root b/c positive reals and sqrt are isomorphic
def radical_euclidean_dist(x, y):
"""Takes in 2 tensors and returns euclidean distance radical, i.e. dist**2"""
with tf.name_scope("radical_euclidean"):
return tf.square(tf.sub(x, y))
# should be shape(self.n_pixels, self.k)
distances = tf.reduce_sum(radical_euclidean_dist(tiled_pix, self.centroids_in),
reduction_indices=2, name="distances")
# should be shape(self.n_pixels)
nearest = tf.to_int32(tf.argmin(distances, 1), name="nearest")
# should be list of len self.k with tensors of shape(size_cluster, size_data)
self.clusters = tf.dynamic_partition(self.arr, nearest, self.k)
# should be shape(self.k, size_data)
self.centroids = tf.pack([tf.reduce_mean(cluster, 0) for cluster in self.clusters],
name="centroids_out")
self.update_roids = tf.assign(self.centroids_in, self.centroids)
def scheduled_sample_count(ground_truth_x,
generated_x,
batch_size,
scheduled_sample_var):
"""Sample batch with specified mix of groundtruth and generated data points.
Args:
ground_truth_x: tensor of ground-truth data points.
generated_x: tensor of generated data points.
batch_size: batch size
scheduled_sample_var: number of ground-truth examples to include in batch.
Returns:
New batch with num_ground_truth sampled from ground_truth_x and the rest
from generated_x.
"""
num_ground_truth = scheduled_sample_var
idx = tf.random_shuffle(tf.range(batch_size))
ground_truth_idx = tf.gather(idx, tf.range(num_ground_truth))
generated_idx = tf.gather(idx, tf.range(num_ground_truth, batch_size))
ground_truth_examps = tf.gather(ground_truth_x, ground_truth_idx)
generated_examps = tf.gather(generated_x, generated_idx)
output = tf.dynamic_stitch([ground_truth_idx, generated_idx],
[ground_truth_examps, generated_examps])
# if batch size is known set it.
if isinstance(batch_size, int):
output.set_shape([batch_size] + common_layers.shape_list(output)[1:])
return output
def MiniminibatchLayer(name, n_in, dim_b, dim_c, group_size, inputs):
inputs = tf.random_shuffle(inputs)
inputs = tf.reshape(inputs, [-1, group_size, n_in])
def f(a,x):
return MinibatchLayer(name, n_in, dim_b, dim_c, x)
outputs = tf.scan(f, inputs)
return tf.reshape(outputs, [-1, n_in+dim_b])
def _add_gtboxes_as_first_stage_proposals(self, first_stage_proposals, first_stage_scores, gtboxes):
# 1. jitter gtboxes
ws = gtboxes[:, 2]
hs = gtboxes[:, 3]
thetas = gtboxes[:, 4]
hs_offset = (tf.random_normal(shape=tf.shape(hs)) - 0.5)*0.1*hs
ws_offset = (tf.random_normal(shape=tf.shape(ws)) - 0.5)*0.1*ws
thetas_offset = (tf.random_normal(shape=tf.shape(thetas)) - 0.5)*0.1*thetas
hs = hs + hs_offset
ws = ws + ws_offset
thetas = thetas + thetas_offset
new_boxes = tf.transpose(tf.stack([gtboxes[:, 0], gtboxes[:, 1], ws, hs, thetas], axis=0))
# 2. get needed added gtboxes
num_needed_add = tf.minimum(tf.cast(cfgs.FAST_RCNN_MINIBATCH_SIZE*cfgs.FAST_RCNN_POSITIVE_RATE*0.5, tf.int32),
tf.shape(gtboxes)[0])
added_boxes_indices = tf.random_shuffle(tf.range(start=0, limit=tf.shape(new_boxes)[0]))
added_boxes_indices = tf.slice(added_boxes_indices, begin=[0], size=[num_needed_add])
added_boxes = tf.gather(new_boxes, added_boxes_indices)
# 3. add them
all_boxes = tf.concat([first_stage_proposals, added_boxes], axis=0)
all_scores = tf.concat([first_stage_scores, tf.ones(shape=[tf.shape(added_boxes)[0]])*0.95], axis=0)
return all_boxes, all_scores
def disable_some_fgs():
# We want to delete a randomly-selected subset of fg_inds of
# size `fg_inds.shape[0] - max_fg`.
# We shuffle along the dimension 0 and then we get the first
# num_fg_inds - max_fg indices and we disable them.
shuffled_inds = tf.random_shuffle(fg_inds, seed=self._seed)
disable_place = (tf.shape(fg_inds)[0] - max_fg)
# This function should never run if num_fg_inds <= max_fg, so we
# add an assertion to catch the wrong behaviour if it happens.
integrity_assertion = tf.assert_positive(
disable_place,
message="disable_place in disable_some_fgs is negative."
)
with tf.control_dependencies([integrity_assertion]):
disable_inds = shuffled_inds[:disable_place]
is_disabled = tf.sparse_to_dense(
sparse_indices=disable_inds,
sparse_values=True, default_value=False,
output_shape=tf.cast(proposals_label_shape, tf.int64),
# We are shuffling the indices, so they may not be ordered.
validate_indices=False
)
return tf.where(
condition=is_disabled,
# We set it to -label for debugging purposes.
x=tf.negative(proposals_label),
y=proposals_label
)
def generate_one(d):
seed = stream()
fn = lambda _: tf.random_shuffle(tf.range(d), seed=seed)
return tf.map_fn(
fn,
sample_range,
parallel_iterations=1 if seed is not None else 10)
def __init__(self, config):
paths, meta = Input._collect(config.path)
self.dimension_count = meta['dimension_count']
self.sample_count = meta['sample_count']
self.batch_size = config.get('batch_size', 1)
if self.sample_count % self.batch_size > 0:
raise Exception(
('expected the number of samples ({}) to be ' +
'divisible by the batch size ({})').format(self.sample_count,
self.batch_size))
with tf.variable_scope('state'):
self.state = State()
with tf.variable_scope('source'):
paths = tf.Variable(paths, name='paths', dtype=tf.string,
trainable=False)
queue = tf.FIFOQueue(meta['path_count'], [tf.string])
enqueue = queue.enqueue_many([tf.random_shuffle(paths)])
tf.train.add_queue_runner(tf.train.QueueRunner(queue, [enqueue]))
_, record = tf.TFRecordReader().read(queue)
with tf.variable_scope('x'):
features = tf.parse_single_example(record, {
'data': tf.VarLenFeature(tf.float32),
})
data = tf.sparse_tensor_to_dense(features['data'])
if self.batch_size == 1:
self.x = tf.reshape(data, [1, -1, self.dimension_count])
else:
x = tf.reshape(data, [-1, self.dimension_count])
_, outputs = tf.contrib.training.bucket_by_sequence_length(
tf.shape(x)[0], [x], self.batch_size, config.buckets,
dynamic_pad=True)
self.x = outputs[0]
with tf.variable_scope('y'):
self.y = tf.pad(self.x[:, 1:, :], [[0, 0], [0, 1], [0, 0]])
def build(self, input_shape):
input_dim = input_shape[1]
#Per tree
N_DECISION = (2 ** (self.n_depth)) - 1 # Number of decision nodes
N_LEAF = 2 ** (self.n_depth + 1) # Number of leaf nodes
if self.randomize_training:
#Construct a mask that lets N trees get trained per minibatch
train_mask = np.zeros(self.n_trees, dtype=np.float32)
for i in xrange(self.randomize_training):
train_mask[i] = 1
self.random_mask = tf.random_shuffle(tf.constant(train_mask))
self.w_d_ensemble = []
self.w_l_ensemble = []
self.trainable_weights = []
for i in xrange(self.n_trees):
decision_weights = self.d_init((input_dim, N_DECISION), name=self.name+"_tree"+i+"_dW")
leaf_distributions = self.l_init((N_LEAF, self.output_classes), name=self.name+"_tree"+i+"_lW")
self.trainable_weights.append(decision_weights)
self.trainable_weights.append(leaf_distributions)
if self.randomize_training:
do_gradient = self.random_mask[i]
no_gradient = 1 - do_gradient
#This should always allow inference, but block gradient flow when do_gradient = 0
decision_weights = do_gradient * decision_weights + no_gradient * tf.stop_gradient(decision_weights)
leaf_distributions = do_gradient * leaf_distributions + no_gradient * tf.stop_gradient(leaf_distributions)
self.w_d_ensemble.append(decision_weights)
self.w_l_ensemble.append(leaf_distributions)
def PreDiscriminator(inputs):
outputs = []
for n_rows in [784]:
output = tf.reshape(inputs, [-1, n_rows, 1])
output = tf.gather(output, tf.random_shuffle(tf.range((784/n_rows)*BATCH_SIZE))[:BATCH_SIZE])
output = lib.ops.gru.GRU('Discriminator.GRU_{}'.format(1), 1, 256, output)
outputs.append(output)
return outputs
def generate_one(d):
seed[0] = distributions_util.gen_new_seed(
seed[0], salt='mcmc_sample_halton_sequence_4')
fn = lambda _: tf.random_shuffle(tf.range(d), seed=seed[0])
return tf.map_fn(
fn,
sample_range,
parallel_iterations=1 if seed[0] is not None else 10)
def sample_fg_bg(iou):
fg_mask = tf.reduce_max(iou, axis=1) >= cfg.FRCNN.FG_THRESH
fg_inds = tf.reshape(tf.where(fg_mask), [-1])
num_fg = tf.minimum(int(
cfg.FRCNN.BATCH_PER_IM * cfg.FRCNN.FG_RATIO),
tf.size(fg_inds), name='num_fg')
fg_inds = tf.random_shuffle(fg_inds)[:num_fg]
bg_inds = tf.reshape(tf.where(tf.logical_not(fg_mask)), [-1])
num_bg = tf.minimum(
cfg.FRCNN.BATCH_PER_IM - num_fg,
tf.size(bg_inds), name='num_bg')
bg_inds = tf.random_shuffle(bg_inds)[:num_bg]
add_moving_summary(num_fg, num_bg)
return fg_inds, bg_inds
def get_random_init(sess, batch): # returns (pseudo) random parameters for the algorithm
points = tf.placeholder(tf.float32, shape=[n, p, 1])
pi_init = sess.run(tf.random_uniform([1, k]), feed_dict={points: batch})
mu_init = sess.run(tf.slice(tf.random_shuffle(points), [0, 0, 0], [k, p, 1]), feed_dict={points: batch})
seed = tf.random_uniform([k, p, p])
sigma_init = sess.run(tf.batch_matmul(seed, tf.transpose(seed, [0, 2, 1])), feed_dict={points: batch})
return pi_init, mu_init, sigma_init
def get_sparse_subset_tensor(shape, subset_sizes,
initializer=tf.random_normal_initializer(
stddev=0.15)):
"""Gets the required bits and pieces for a sparse tensor bilinear product
with random subsets of the inputs (as opposed to a totally randomly sparse
tensor, this will have a kind of rectangular structure).
Args:
shape: the shape of the tensor. We can only make this work for
3-tensors so len(shape) should be 3.
subset_sizes: the number of random elements to take from each of the
inputs. Should be a sequence of length 2.
initializer: the initialiser to use for the elements of the tensor.
Returns:
(tensor, idcs_a, idcs_b): all that we need to do the bilinear product
later -- the tensor elements as a dense matrix and the indices
representing the indices into the input vectors.
"""
# first let's make sure the inputs make sense
if len(shape) != 3:
raise ValueError(
'Can do this with 3-way tensors, got shape {}'.format(shape))
if len(subset_sizes) != 2:
raise ValueError('subset_sizes needs to be of length two')
if subset_sizes[0] > shape[0]:
raise ValueError('first subset size greater than specified dimension: '
'{} > {}'.format(subset_sizes[0], shape[0]))
if subset_sizes[2] > shape[1]:
raise ValueError(
'second subset size greater than specified dimension: '
'{} > {}'.format(subset_sizes[2], shape[1]))
with tf.name_scope(name):
# now we need to make some random indices
# potentially kinda gross for a little while
a_idcs = tf.stack([tf.random_shuffle(tf.range(shape[0]))
for _ in range(shape[1])])
b_idcs = tf.stack([tf.random_shuffle(tf.range(shape[2]))
for _ in range(shape[1])])
# if we eval these they will be random every time
# so we use them as the initialiser to a new variable
a_idcs = tf.get_variable('a_idcs',
initializer=a_idcs[:, :subset_sizes[0]])
b_idcs = tf.get_variable('b_idcs',
initializer=b_idcs[:, :subset_sizes[1]])
def init_k_means(samples, num_clusters, num_samples_per_cluster):
sample_indices = tf.random_shuffle(tf.range(0,num_samples_per_cluster*num_clusters))
#Start by choosing num_cluster random points
start = [0,]
size = [num_clusters,]
size[0] = num_clusters
centroids = tf.slice(sample_indices, start, size)
chosen_centroids = tf.gather(samples, centroids)
return chosen_centroids
def prepare_split(i):
idx = tf.random_shuffle(tf.range(0, n_pages, dtype=tf.int32), seed + i)
train_tensors = [tf.gather(tensor, idx, name=mk_name('shfl', tensor)) for tensor in tensors]
if test_sampling < 1.0:
sampled_idx = idx[:n_pages]
test_tensors = [tf.gather(tensor, sampled_idx, name=mk_name('shfl_test', tensor)) for tensor in tensors]
else:
test_tensors = train_tensors
return Split(test_tensors, train_tensors, n_pages, total_pages)
def get_ensemble_idx_info(self):
if self.bayesian_config is not False:
ensemble_idxs = tf.random_shuffle(tf.range(self.transition_predictor.ensemble_size))
transition_ensemble_sample_n = self.transition_predictor.eval_sample_count
reward_ensemble_sample_n = self.reward_predictor.eval_sample_count
ensemble_idxs = ensemble_idxs[:transition_ensemble_sample_n]
return ensemble_idxs, transition_ensemble_sample_n, reward_ensemble_sample_n
else:
return None, 1, 1
def choose_random_centroids(samples, n_clusters):
n_samples = tf.shape(samples)[0]
random_indices = tf.random_shuffle(tf.range(0, n_samples))
begin = [0,]
size = [n_clusters,]
size[0] = n_clusters
centroid_indices = tf.slice(random_indices, begin, size)
initial_centroids = tf.gather(samples, centroid_indices)
return initial_centroids
def __call__(self, batch_size, **kwargs):
"""Sample a batch of context.
Args:
batch_size: Batch size.
Returns:
Two [batch_size, num_context_dims] tensors.
"""
spec = self._context_spec
context_range = self._context_range
if isinstance(context_range[0], (int, float)):
contexts = tf.random_uniform(
shape=[
batch_size,
] + spec.shape.as_list(),
minval=context_range[0],
maxval=context_range[1],
dtype=spec.dtype)
elif isinstance(context_range[0], (list, tuple, np.ndarray)):
assert len(spec.shape.as_list()) == 1
assert spec.shape.as_list()[0] == len(context_range[0])
assert spec.shape.as_list()[0] == len(context_range[1])
contexts = tf.concat(
[
tf.random_uniform(
shape=[
batch_size, 1,
] + spec.shape.as_list()[1:],
minval=context_range[0][i],
maxval=context_range[1][i],
dtype=spec.dtype) for i in range(spec.shape.as_list()[0])
],
axis=1)
else: raise NotImplementedError(context_range)
self._validate_contexts(contexts)
if 'sampler_fn' in kwargs:
other_contexts = kwargs['sampler_fn']()
else:
other_contexts = contexts
state, next_state = kwargs['state'], kwargs['next_state']
if state is not None and next_state is not None:
my_context_range = (np.array(context_range[1]) - np.array(context_range[0])) / 2 * np.ones(spec.shape.as_list())
contexts = tf.concat(
[0.1 * my_context_range[:self._k] *
tf.random_normal(tf.shape(state[:, :self._k]), dtype=state.dtype) +
tf.random_shuffle(state[:, :self._k]) - state[:, :self._k],
other_contexts[:, self._k:]], 1)
#contexts = tf.Print(contexts,
# [contexts, tf.reduce_max(contexts, 0),
# tf.reduce_min(state, 0), tf.reduce_max(state, 0)], 'contexts', summarize=15)
next_contexts = tf.concat( #LALA
[state[:, :self._k] + contexts[:, :self._k] - next_state[:, :self._k],
other_contexts[:, self._k:]], 1)
next_contexts = contexts #LALA cosine
else:
next_contexts = contexts
return tf.stop_gradient(contexts), tf.stop_gradient(next_contexts)
def tf_fastfood_transform(in_x, dd, DD, use_get=False, use_C=False):
'''Transform from d to D. Pads as necessary.
For now: assume dd and DD are known in python.'''
# Tensor d and D
#assert_D_big = tf.assert_greater_equal(DD, dd, message='d cannot be larger than D')
#with tf.control_dependencies([assert_D_big]):
# ll = tf.cast(tf.round(tf.log(tf.to_float(DD)) / np.log(2)), 'int32')
# LL = tf.pow(2, ll)
# Python d and D
assert isinstance(dd, int), 'd should be int'
assert isinstance(DD, int), 'D should be int'
assert DD >= dd, 'd cannot be larger than D'
assert dd > 0, 'd and D must be positive'
ll = int(np.ceil(np.log(DD) / np.log(2)))
LL = 2 ** ll
# Make vars
init_BB = tf.to_float(tf.random_uniform((LL,), 0, 2, dtype='int32')) * 2 - 1
init_Pi = tf.random_shuffle(tf.range(LL))
init_GG = tf.random_normal((LL,))
init_divisor = lambda GG: tf.sqrt(LL * tf.reduce_sum(tf.pow(GG.initialized_value(), 2)))
if use_get:
BB = tf.get_variable('B', initializer=init_BB, trainable=False)
Pi = tf.get_variable('Pi', initializer=init_Pi, trainable=False)
GG = tf.get_variable('G', initializer=init_GG, trainable=False)
divisor = tf.get_variable('divisor', initializer=init_divisor(GG), trainable=False)
else:
BB = tf.Variable(init_BB, name='B', trainable=False)
Pi = tf.Variable(init_Pi, name='Pi', trainable=False)
GG = tf.Variable(init_GG, name='G', trainable=False)
divisor = tf.Variable(init_divisor(GG), name='divisor', trainable=False)
fastfood_vars = [BB, Pi, GG, divisor]
# Implement transform
dd_pad = tf.pad(in_x, [[0, LL - dd]])
mul_1 = tf.multiply(BB, dd_pad)
if use_C:
mul_2 = tf_fast_walsh_hadamard(mul_1, 0, method='c', normalize=True)
else:
mul_2 = tf_fast_walsh_hadamard(mul_1, 0, method='two', normalize=False)
mul_3 = tf.gather(mul_2, Pi)
mul_4 = tf.multiply(mul_3, GG)
if use_C:
mul_5 = tf_fast_walsh_hadamard(mul_4, 0, method='c', normalize=True)
print '\nWARNING: check normalization on this next line more carefully\n'
ret = tf.divide(tf.slice(mul_5, [0], [DD]), divisor * np.sqrt(float(DD) / LL / ll))
else:
mul_5 = tf_fast_walsh_hadamard(mul_4, 0, method='two', normalize=False)
ret = tf.divide(tf.slice(mul_5, [0], [DD]), divisor * np.sqrt(float(DD) / LL))
return fastfood_vars, ret
def choose_random_centroids(samples, n_clusters, seed):
# Step 0: Initialisation: Select `n_clusters` number of random points
n_samples = tf.shape(samples)[0]
random_indices = tf.random_shuffle(tf.range(0, n_samples), seed=seed)
begin = [0,]
size = [n_clusters,]
# size[0] = n_clusters
centroid_indices = tf.slice(random_indices, begin, size)
initial_centroids = tf.gather(samples, centroid_indices)
return initial_centroids
def dequeue_batch():
with tf.name_scope("sample_n_per_batch/dequeue/"):
entries = []
for q in queues:
entries.append(q.dequeue_many(samples_per_class))
flat_batch = [tf.concat(x, 0) for x in zip(*entries)]
idx = tf.random_shuffle(tf.range(batch_size))
flat_batch = [tf.gather(f, idx, axis=0) for f in flat_batch]
return nest.pack_sequence_as(batch, flat_batch)
def random_column(columns):
"""Zeros out all except one of `columns`.
Used for rounds with global drop path.
Args:
columns: the columns of a fractal block to be selected from.
"""
num_columns = tensor_shape(columns)[0]
mask = tf.random_shuffle([True]+[False]*(num_columns-1))
return apply_mask(mask, columns)* num_columns
def preprocess_for_train(image,
output_height,
output_width,
mean_vals,
out_dim_scale=1.0):
"""Preprocesses the given image for training.
Note that the actual resizing scale is sampled from
[`resize_size_min`, `resize_size_max`].
Args:
image: A `Tensor` representing an image of arbitrary size.
output_height: The height of the image after preprocessing.
output_width: The width of the image after preprocessing.
Returns:
A preprocessed image.
"""
num_channels = image.get_shape().as_list()[-1]
image = tf.image.resize_images(image, [_RESIZE_HT, _RESIZE_WD])
# compute the crop size
base_size = float(min(_RESIZE_HT, _RESIZE_WD))
scale_ratio_h = tf.random_shuffle(tf.constant(_SCALE_RATIOS))[0]
scale_ratio_w = tf.random_shuffle(tf.constant(_SCALE_RATIOS))[0]
image = _random_crop([image],
tf.cast(output_height * scale_ratio_h, tf.int32),
tf.cast(output_width * scale_ratio_w, tf.int32))[0]
image = tf.image.resize_images(
image, [int(output_height * out_dim_scale),
int(output_width * out_dim_scale)])
image = tf.to_float(image)
image = tf.image.random_flip_left_right(image)
image.set_shape([int(output_height * out_dim_scale),
int(output_width * out_dim_scale), num_channels])
image = _mean_image_subtraction(image, mean_vals)
image = tf.expand_dims(image, 0) # 1x... image, to be consistent with eval
# Gets logged multiple times with NetVLAD, so gives an error.
# I'm anyway logging from the train code, so removing it here.
# tf.image_summary('final_distorted_image',
# tf.expand_dims(image / 128.0, 0))
return image
# -*- coding:utf-8 -*- # @Time :2018/9/23 下午3:43 # @Author :Coast Cao import tensorflow as tf # Create a tensor of shape [2, 3] consisting of random normal values, with mean # -1 and standard deviation 4. norm = tf.random_normal([2, 3], mean=-1, stddev=4) # Shuffle the first dimension of a tensor c = tf.constant([[1, 2], [3, 4], [5, 6]]) shuff = tf.random_shuffle(c) # Each time we run these ops, different results are generated sess = tf.Session() print(sess.run(norm)) print(sess.run(norm)) # Set an op-level seed to generate repeatable sequences across sessions. norm = tf.random_normal([2, 3], seed=1234) sess = tf.Session() print(sess.run(norm)) print(sess.run(norm)) sess = tf.Session() print(sess.run(norm)) print(sess.run(norm))
def train():
with tf.Graph().as_default(), tf.device('/cpu:0'):
num_gpu = len(cfgs.GPU_GROUP.strip().split(','))
global_step = slim.get_or_create_global_step()
lr = warmup_lr(cfgs.LR, global_step, cfgs.WARM_SETP, num_gpu)
tf.summary.scalar('lr', lr)
with tf.name_scope('get_batch'):
if cfgs.IMAGE_PYRAMID:
shortside_len_list = tf.constant(cfgs.IMG_SHORT_SIDE_LEN)
shortside_len = tf.random_shuffle(shortside_len_list)[0]
else:
shortside_len = cfgs.IMG_SHORT_SIDE_LEN
img_name_batch, img_batch, gtboxes_and_label_batch, num_objects_batch, img_h_batch, img_w_batch = \
next_batch(dataset_name=cfgs.DATASET_NAME,
batch_size=cfgs.BATCH_SIZE * num_gpu,
shortside_len=shortside_len,
is_training=True)
optimizer = tf.train.MomentumOptimizer(lr, momentum=cfgs.MOMENTUM)
r3det = build_whole_network_r3det_efficientnet.DetectionNetwork(
base_network_name=cfgs.NET_NAME, is_training=True)
# data processing
inputs_list = []
for i in range(num_gpu):
img = tf.expand_dims(img_batch[i], axis=0)
if cfgs.NET_NAME in [
'resnet152_v1d', 'resnet101_v1d', 'resnet50_v1d'
]:
img = img / tf.constant([cfgs.PIXEL_STD])
gtboxes_and_label_r = tf.py_func(backward_convert,
inp=[gtboxes_and_label_batch[i]],
Tout=tf.float32)
gtboxes_and_label_r = tf.reshape(gtboxes_and_label_r, [-1, 6])
gtboxes_and_label_h = get_horizen_minAreaRectangle(
gtboxes_and_label_batch[i])
gtboxes_and_label_h = tf.reshape(gtboxes_and_label_h, [-1, 5])
num_objects = num_objects_batch[i]
num_objects = tf.cast(tf.reshape(num_objects, [
-1,
]), tf.float32)
img_h = img_h_batch[i]
img_w = img_w_batch[i]
inputs_list.append([
img, gtboxes_and_label_h, gtboxes_and_label_r, num_objects,
img_h, img_w
])
tower_grads = []
biases_regularizer = tf.no_regularizer
weights_regularizer = tf.contrib.layers.l2_regularizer(
cfgs.WEIGHT_DECAY)
total_loss_dict = {
'cls_loss': tf.constant(0., tf.float32),
'reg_loss': tf.constant(0., tf.float32),
'refine_cls_loss': tf.constant(0., tf.float32),
'refine_reg_loss': tf.constant(0., tf.float32),
'refine_cls_loss_stage3': tf.constant(0., tf.float32),
'refine_reg_loss_stage3': tf.constant(0., tf.float32),
'total_losses': tf.constant(0., tf.float32),
}
if cfgs.USE_SUPERVISED_MASK:
total_loss_dict['mask_loss'] = tf.constant(0., tf.float32)
with tf.variable_scope(tf.get_variable_scope()):
for i in range(num_gpu):
with tf.device('/gpu:%d' % i):
with tf.name_scope('tower_%d' % i):
with slim.arg_scope(
[slim.model_variable, slim.variable],
device='/device:CPU:0'):
with slim.arg_scope(
[
slim.conv2d, slim.conv2d_in_plane,
slim.conv2d_transpose,
slim.separable_conv2d, slim.fully_connected
],
weights_regularizer=weights_regularizer,
biases_regularizer=biases_regularizer,
biases_initializer=tf.constant_initializer(
0.0)):
gtboxes_and_label_h, gtboxes_and_label_r = tf.py_func(
get_gtboxes_and_label,
inp=[
inputs_list[i][1], inputs_list[i][2],
inputs_list[i][3]
],
Tout=[tf.float32, tf.float32])
gtboxes_and_label_h = tf.reshape(
gtboxes_and_label_h, [-1, 5])
gtboxes_and_label_r = tf.reshape(
gtboxes_and_label_r, [-1, 6])
img = inputs_list[i][0]
img_shape = inputs_list[i][-2:]
img = tf.image.crop_to_bounding_box(
image=img,
offset_height=0,
offset_width=0,
target_height=tf.cast(
img_shape[0], tf.int32),
target_width=tf.cast(
img_shape[1], tf.int32))
outputs = r3det.build_whole_detection_network(
input_img_batch=img,
gtboxes_batch_h=gtboxes_and_label_h,
gtboxes_batch_r=gtboxes_and_label_r,
gpu_id=i)
gtboxes_in_img_h = draw_boxes_with_categories(
img_batch=img,
boxes=gtboxes_and_label_h[:, :-1],
labels=gtboxes_and_label_h[:, -1],
method=0)
gtboxes_in_img_r = draw_boxes_with_categories(
img_batch=img,
boxes=gtboxes_and_label_r[:, :-1],
labels=gtboxes_and_label_r[:, -1],
method=1)
tf.summary.image(
'Compare/gtboxes_h_gpu:%d' % i,
gtboxes_in_img_h)
tf.summary.image(
'Compare/gtboxes_r_gpu:%d' % i,
gtboxes_in_img_r)
if cfgs.ADD_BOX_IN_TENSORBOARD:
detections_in_img = draw_boxes_with_categories_and_scores(
img_batch=img,
boxes=outputs[0],
scores=outputs[1],
labels=outputs[2],
method=1)
tf.summary.image(
'Compare/final_detection_gpu:%d' % i,
detections_in_img)
loss_dict = outputs[-1]
total_losses = 0.0
for k in loss_dict.keys():
total_losses += loss_dict[k]
total_loss_dict[
k] += loss_dict[k] / num_gpu
total_losses = total_losses / num_gpu
total_loss_dict['total_losses'] += total_losses
if i == num_gpu - 1:
regularization_losses = tf.get_collection(
tf.GraphKeys.REGULARIZATION_LOSSES)
# weight_decay_loss = tf.add_n(slim.losses.get_regularization_losses())
total_losses = total_losses + tf.add_n(
regularization_losses)
tf.get_variable_scope().reuse_variables()
grads = optimizer.compute_gradients(total_losses)
if cfgs.GRADIENT_CLIPPING_BY_NORM is not None:
grads = slim.learning.clip_gradient_norms(
grads, cfgs.GRADIENT_CLIPPING_BY_NORM)
tower_grads.append(grads)
for k in total_loss_dict.keys():
tf.summary.scalar('{}/{}'.format(k.split('_')[0], k),
total_loss_dict[k])
if len(tower_grads) > 1:
grads = sum_gradients(tower_grads)
else:
grads = tower_grads[0]
if cfgs.MUTILPY_BIAS_GRADIENT is not None:
final_gvs = []
with tf.variable_scope('Gradient_Mult'):
for grad, var in grads:
scale = 1.
if '/biases:' in var.name:
scale *= cfgs.MUTILPY_BIAS_GRADIENT
if 'conv_new' in var.name:
scale *= 3.
if not np.allclose(scale, 1.0):
grad = tf.multiply(grad, scale)
final_gvs.append((grad, var))
apply_gradient_op = optimizer.apply_gradients(
final_gvs, global_step=global_step)
else:
apply_gradient_op = optimizer.apply_gradients(
grads, global_step=global_step)
variable_averages = tf.train.ExponentialMovingAverage(
0.9999, global_step)
variables_averages_op = variable_averages.apply(
tf.trainable_variables())
train_op = tf.group(apply_gradient_op, variables_averages_op)
# train_op = optimizer.apply_gradients(final_gvs, global_step=global_step)
summary_op = tf.summary.merge_all()
restorer, restore_ckpt = r3det.get_restorer()
saver = tf.train.Saver(max_to_keep=5)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
tfconfig = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
tfconfig.gpu_options.allow_growth = True
with tf.Session(config=tfconfig) as sess:
sess.run(init_op)
# sess.run(tf.initialize_all_variables())
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
summary_path = os.path.join(cfgs.SUMMARY_PATH, cfgs.VERSION)
tools.mkdir(summary_path)
summary_writer = tf.summary.FileWriter(summary_path,
graph=sess.graph)
if not restorer is None:
restorer.restore(sess, restore_ckpt)
print('restore model')
for step in range(cfgs.MAX_ITERATION // num_gpu):
training_time = time.strftime('%Y-%m-%d %H:%M:%S',
time.localtime(time.time()))
if step % cfgs.SHOW_TRAIN_INFO_INTE != 0 and step % cfgs.SMRY_ITER != 0:
_, global_stepnp = sess.run([train_op, global_step])
else:
if step % cfgs.SHOW_TRAIN_INFO_INTE == 0 and step % cfgs.SMRY_ITER != 0:
start = time.time()
_, global_stepnp, total_loss_dict_ = \
sess.run([train_op, global_step, total_loss_dict])
end = time.time()
print('***' * 20)
print("""%s: global_step:%d current_step:%d""" %
(training_time,
(global_stepnp - 1) * num_gpu, step * num_gpu))
print("""per_cost_time:%.3fs""" %
((end - start) / num_gpu))
loss_str = ''
for k in total_loss_dict_.keys():
loss_str += '%s:%.3f\n' % (k, total_loss_dict_[k])
print(loss_str)
if np.isnan(total_loss_dict_['total_losses']):
sys.exit(0)
else:
if step % cfgs.SMRY_ITER == 0:
_, global_stepnp, summary_str = sess.run(
[train_op, global_step, summary_op])
summary_writer.add_summary(
summary_str, (global_stepnp - 1) * num_gpu)
summary_writer.flush()
if (step > 0 and step % (cfgs.SAVE_WEIGHTS_INTE // num_gpu)
== 0) or (step >= cfgs.MAX_ITERATION // num_gpu - 1):
save_dir = os.path.join(cfgs.TRAINED_CKPT, cfgs.VERSION)
if not os.path.exists(save_dir):
os.mkdir(save_dir)
save_ckpt = os.path.join(
save_dir, '{}_'.format(cfgs.DATASET_NAME) + str(
(global_stepnp - 1) * num_gpu) + 'model.ckpt')
saver.save(sess, save_ckpt)
print(' weights had been saved')
coord.request_stop()
coord.join(threads)
def rpn_targets_graph(gt_boxes,
gt_cls,
anchors,
anchors_tag,
rpn_train_anchors=None):
"""
处理单个图像的rpn分类和回归目标
a)正样本为 IoU>0.7的anchor;负样本为IoU<0.3的anchor; 居中的为中性样本，丢弃
b)需要保证所有的GT都有anchor对应，即使IoU<0.3；
c)正负样本比例保持1:1
:param gt_boxes: GT 边框坐标 [MAX_GT_BOXs, (y1,x1,y2,x2,tag)] ,tag=0 为padding
:param gt_cls: GT 类别 [MAX_GT_BOXs, 1+1] ;最后一位为tag, tag=0 为padding
:param anchors: [anchor_num, (y1,x1,y2,x2)]
:param anchors_tag:[anchor_num] bool类型
:param rpn_train_anchors: 训练样本数(256)
:return:
deltas:[rpn_train_anchors,(dy,dx,dh,dw,tag)]：anchor边框回归目标,tag=1 为正样本，tag=0为padding，tag=-1为负样本
class_ids:[rpn_train_anchors,1+1]: anchor边框分类,tag=1 为正样本，tag=0为padding，tag=-1为负样本
indices:[rpn_train_anchors,(indices,tag)]: tag=1 为正样本，tag=0为padding，tag=-1为负样本
"""
# 获取真正的GT,去除标签位
gt_boxes = tf_utils.remove_pad(gt_boxes)
gt_cls = tf_utils.remove_pad(gt_cls)[:, 0] # [N,1]转[N]
# 获取有效的anchors
valid_anchor_indices = tf.where(anchors_tag)[:, 0] # [valid_anchors_num]
anchors = tf.gather(anchors, valid_anchor_indices)
# 计算IoU
iou = compute_iou(gt_boxes, anchors)
# print("iou:{}".format(iou))
# 每个GT对应的IoU最大的anchor是正样本
gt_iou_argmax = tf.argmax(iou, axis=1)
positive_gt_indices_1 = tf.range(tf.shape(gt_boxes)[0]) # 索引号就是1..n-1
positive_anchor_indices_1 = gt_iou_argmax
# 每个anchors最大iou ，且iou>0.7的为正样本
anchors_iou_max = tf.reduce_max(iou, axis=0)
# 正样本索引号（iou>0.7),
positive_anchor_indices_2 = tf.where(
anchors_iou_max > 0.7, name='rpn_target_positive_indices') # [:, 0]
# 找到正样本对应的GT boxes 索引
# anchors_iou_argmax = tf.argmax(iou, axis=0) # 每个anchor最大iou对应的GT 索引 [n]
anchors_iou_argmax = tf.cond( # 需要考虑GT个数为0的情况
tf.greater(tf.shape(gt_boxes)[0], 0),
true_fn=lambda: tf.argmax(iou, axis=0),
false_fn=lambda: tf.cast(tf.constant([]), tf.int64))
positive_gt_indices_2 = tf.gather_nd(anchors_iou_argmax,
positive_anchor_indices_2)
# 合并两部分正样本
positive_gt_indices = tf.concat(
[positive_gt_indices_1,
tf.cast(positive_gt_indices_2, tf.int32)],
axis=0,
name='rpn_gt_boxes_concat')
positive_anchor_indices = tf.concat(
[positive_anchor_indices_1, positive_anchor_indices_2[:, 0]],
axis=0,
name='rpn_positive_anchors_concat')
# 根据正负样本比1:1,确定最终的正样本
positive_num = tf.minimum(
tf.shape(positive_anchor_indices)[0], int(rpn_train_anchors * 0.9))
positive_anchor_indices, positive_gt_indices = shuffle_sample(
[positive_anchor_indices, positive_gt_indices],
tf.shape(positive_anchor_indices)[0], positive_num)
# 根据索引选择anchor和GT
positive_anchors = tf.gather(anchors, positive_anchor_indices)
positive_gt_boxes = tf.gather(gt_boxes, positive_gt_indices)
positive_gt_cls = tf.gather(gt_cls, positive_gt_indices)
# 回归目标计算
deltas = regress_target(positive_anchors, positive_gt_boxes)
# 处理负样本
negative_indices = tf.where(anchors_iou_max < 0.3,
name='rpn_target_negative_indices') # [:, 0]
# 负样本,保证负样本不超过一半
negative_num = tf.minimum(rpn_train_anchors - positive_num,
tf.shape(negative_indices)[0],
name='rpn_target_negative_num')
# negative_num = tf.minimum(int(rpn_train_anchors * 0.5), negative_num, name='rpn_target_negative_num_2')
negative_indices = tf.random_shuffle(negative_indices)[:negative_num]
negative_gt_cls = tf.zeros([negative_num]) # 负样本类别id为0
negative_deltas = tf.zeros([negative_num, 4])
# 合并正负样本
deltas = tf.concat([deltas, negative_deltas],
axis=0,
name='rpn_target_deltas')
class_ids = tf.concat([positive_gt_cls, negative_gt_cls],
axis=0,
name='rpn_target_class_ids')
indices = tf.concat([positive_anchor_indices, negative_indices[:, 0]],
axis=0,
name='rpn_train_anchor_indices')
# indices转换会原始的anchors索引
indices = tf.gather(valid_anchor_indices,
indices,
name='map_to_origin_anchor_indices')
# 计算padding
deltas, class_ids = tf_utils.pad_list_to_fixed_size(
[deltas, tf.expand_dims(class_ids, 1)], rpn_train_anchors)
# 将负样本tag标志改为-1;方便后续处理;
indices = tf_utils.pad_to_fixed_size_with_negative(
tf.expand_dims(indices, 1),
rpn_train_anchors,
negative_num=negative_num,
data_type=tf.int64)
# 其它统计指标
gt_num = tf.shape(gt_cls)[0] # GT数
miss_match_gt_num = gt_num - tf.shape(
tf.unique(positive_gt_indices)[0])[0] # 未分配anchor的GT
rpn_gt_min_max_iou = tf.reduce_min(tf.reduce_max(
iou, axis=1)) # GT匹配anchor最小的IoU
return [
deltas, class_ids, indices,
tf_utils.scalar_to_1d_tensor(gt_num),
tf_utils.scalar_to_1d_tensor(positive_num),
tf_utils.scalar_to_1d_tensor(negative_num),
tf_utils.scalar_to_1d_tensor(miss_match_gt_num),
tf_utils.scalar_to_1d_tensor(rpn_gt_min_max_iou)
]
def label_downsampler(n_labels, true_labels, n_extra_labels):
lbls = tf.random_shuffle(tf.range(n_labels))[:n_extra_labels]
lbls = tf.unique(tf.concat([lbls, true_labels], axis=0)).y
return lambda logits: tf.gather(logits, lbls, axis=-1)
points_x.append(np.random.normal(0.0, 0.6))
points_y.append(np.random.normal(0.0, 0.6))
elif np.random.random() > 0.33333:
points_x.append(np.random.normal(2.0, 0.57))
points_y.append(np.random.normal(3.0, 0.6))
else:
points_x.append(np.random.normal(2.0, 0.33))
points_y.append(np.random.normal(1.0, 0.2))
# 1000개의 포인트를 numpy array로 생성
points = np.array(np.transpose([points_x, points_y]))
# numpy array를 tensor로 변환
vectors = tf.constant(points)
centroides = tf.Variable(
tf.slice(tf.random_shuffle(vectors), [0, 0],
[num_clusters, -1])) # 중심점 초기화
# 텐서 차원 확장
expanded_vectors = tf.expand_dims(vectors, 0)
expanded_centroides = tf.expand_dims(centroides, 1)
# 거리 계산
diff = tf.subtract(expanded_vectors, expanded_centroides)
distance = tf.reduce_sum(tf.square(diff), 2)
assignments = tf.argmin(distance, 0)
# 중심점 업데이트
means = tf.concat([
tf.reduce_mean(tf.gather(
vectors, tf.reshape(tf.where(tf.equal(assignments, c)), [1, -1])),
def tensorflow_categorical(count, seed):
assert count > 0
arr = [1.] + [.0 for _ in range(count - 1)]
return tf.random_shuffle(arr, seed)
def postproc_flag(
self,
images,
NOW_SIZE1=256,
NOW_SIZE2=256,
seed_random=0,
curr_batch_size=None,
with_noise=0,
noise_level=10,
with_flip=0,
is_normal=0,
eliprep=0,
thprep=0,
sub_mean=0,
mean_path=None,
color_norm=0,
size_vary_prep=0,
with_color_noise=0,
shape_undefined=0,
size_minval=0.08,
sm_full_size=0, # sm_add
):
if curr_batch_size == None:
curr_batch_size = self.batch_size
orig_dtype = images.dtype
norm = tf.cast(images, tf.float32)
if eliprep == 1:
def prep_each(norm_):
_RESIZE_SIDE_MIN = 256
_RESIZE_SIDE_MAX = 512
if self.group == 'train':
im = preprocess_image(norm_,
self.crop_size,
self.crop_size,
is_training=True,
resize_side_min=_RESIZE_SIDE_MIN,
resize_side_max=_RESIZE_SIDE_MAX)
else:
im = preprocess_image(norm_,
self.crop_size,
self.crop_size,
is_training=False,
resize_side_min=_RESIZE_SIDE_MIN,
resize_side_max=_RESIZE_SIDE_MAX)
return im
crop_images = tf.map_fn(prep_each, norm)
elif thprep == 1:
def prep_each(norm_):
im = preprocessing_th(norm_,
self.crop_size,
self.crop_size,
is_training=self.group == 'train',
seed_random=seed_random)
return im
crop_images = prep_each(images)
crop_images = tf.expand_dims(crop_images, axis=0)
elif sm_full_size == 1:
if with_color_noise == 1 and self.group == 'train':
order_temp = tf.constant([0, 1, 2], dtype=tf.int32)
order_rand = tf.random_shuffle(order_temp, seed=seed_random)
fn_pred_fn_pairs = lambda x, image: [
(tf.equal(x, order_temp[0]), lambda: tf.image.
random_saturation(image, 0.6, 1.4, seed=seed_random)),
(tf.equal(x, order_temp[1]), lambda: tf.image.
random_brightness(image, 0.4, seed=seed_random)),
]
default_fn = lambda image: tf.image.random_contrast(
image, 0.6, 1.4, seed=seed_random)
def _color_jitter_one(_norm):
orig_shape = _norm.get_shape().as_list()
_norm = tf.case(fn_pred_fn_pairs(order_rand[0], _norm),
default=lambda: default_fn(_norm))
_norm = tf.case(fn_pred_fn_pairs(order_rand[1], _norm),
default=lambda: default_fn(_norm))
_norm = tf.case(fn_pred_fn_pairs(order_rand[2], _norm),
default=lambda: default_fn(_norm))
_norm.set_shape(orig_shape)
return _norm
norm = tf.map_fn(_color_jitter_one, norm)
if sub_mean == 1:
IMAGENET_MEAN = tf.constant(np.load(mean_path).swapaxes(
0, 1).swapaxes(1, 2)[:, :, ::-1],
dtype=tf.float32)
orig_dtype = tf.float32
norm = norm - IMAGENET_MEAN
if self.group == 'train':
if self.withflip == 1 or with_flip == 1:
def _postprocess_flip(im):
# Original way of flipping, changing to random_uniform way to be more controllable
# im = tf.image.random_flip_left_right(im, seed = seed_random)
# return im
do_flip = tf.random_uniform(shape=[1],
minval=0,
maxval=1,
dtype=tf.float32,
seed=seed_random)
def __left_right_flip(im):
flipped = tf.image.flip_left_right(im)
if is_normal == 1:
# flipped = 256 - flipped
flipped_x, flipped_y, flipped_z = tf.unstack(
flipped, axis=2)
flipped = tf.stack(
[256 - flipped_x, flipped_y, flipped_z],
axis=2)
return flipped
return tf.cond(tf.less(do_flip[0], 0.5),
fn1=lambda: __left_right_flip(im),
fn2=lambda: im)
norm = tf.map_fn(_postprocess_flip, norm, dtype=norm.dtype)
if with_noise == 1:
def _postprocess_noise(im):
do_noise = tf.random_uniform(shape=[1],
minval=0,
maxval=1,
dtype=tf.float32,
seed=None)
def __add_noise(im):
curr_level = tf.random_uniform(shape=[1],
minval=0,
maxval=noise_level,
dtype=tf.float32,
seed=None)
curr_noise = tf.random_normal(shape=tf.shape(im),
mean=0.0,
stddev=curr_level,
dtype=tf.float32)
return tf.add(im, curr_noise)
# return tf.cond(tf.less(do_noise[0], 0.5), true_fn = lambda: __add_noise(im), false_fn = lambda: im)
return tf.cond(tf.less(do_noise[0], 0.5),
fn1=lambda: __add_noise(im),
fn2=lambda: im)
norm = tf.map_fn(_postprocess_noise,
norm,
dtype=norm.dtype)
crop_images = tf.cast(norm, orig_dtype)
else:
if with_color_noise == 1 and self.group == 'train':
order_temp = tf.constant([0, 1, 2], dtype=tf.int32)
order_rand = tf.random_shuffle(order_temp, seed=seed_random)
fn_pred_fn_pairs = lambda x, image: [
(tf.equal(x, order_temp[0]), lambda: tf.image.
random_saturation(image, 0.6, 1.4, seed=seed_random)),
(tf.equal(x, order_temp[1]), lambda: tf.image.
random_brightness(image, 0.4, seed=seed_random)),
]
default_fn = lambda image: tf.image.random_contrast(
image, 0.6, 1.4, seed=seed_random)
def _color_jitter_one(_norm):
orig_shape = _norm.get_shape().as_list()
_norm = tf.case(fn_pred_fn_pairs(order_rand[0], _norm),
default=lambda: default_fn(_norm))
_norm = tf.case(fn_pred_fn_pairs(order_rand[1], _norm),
default=lambda: default_fn(_norm))
_norm = tf.case(fn_pred_fn_pairs(order_rand[2], _norm),
default=lambda: default_fn(_norm))
_norm.set_shape(orig_shape)
return _norm
norm = tf.map_fn(_color_jitter_one, norm)
if sub_mean == 1:
IMAGENET_MEAN = tf.constant(np.load(mean_path).swapaxes(
0, 1).swapaxes(1, 2)[:, :, ::-1],
dtype=tf.float32)
orig_dtype = tf.float32
norm = norm - IMAGENET_MEAN
if self.group == 'train':
if self.size_vary_prep == 0 and size_vary_prep == 0:
shape_tensor = norm.get_shape().as_list()
if self.crop_each == 0:
crop_images = tf.random_crop(norm, [
curr_batch_size, self.crop_size, self.crop_size,
shape_tensor[3]
],
seed=seed_random)
else:
# original implementation is not useful, deleted, see the end of this file
crop_images = tf.random_crop(norm, [
curr_batch_size, self.crop_size, self.crop_size,
shape_tensor[3]
],
seed=seed_random)
else: # self.size_vary_prep==1
if shape_undefined == 0:
channel_num = norm.get_shape().as_list()[-1]
else:
channel_num = 3
RandomSizedCrop_with_para = lambda image: RandomSizedCrop(
image=image,
out_height=self.crop_size,
out_width=self.crop_size,
seed_random=seed_random,
channel_num=channel_num,
fix_asp_ratio=self.fix_asp_ratio,
size_minval=size_minval,
)
if shape_undefined == 0:
crop_images = tf.map_fn(RandomSizedCrop_with_para,
norm)
curr_shape = crop_images.get_shape().as_list()
crop_images.set_shape([curr_batch_size] +
curr_shape[1:])
else:
crop_images = RandomSizedCrop_with_para(norm)
crop_images = tf.expand_dims(crop_images, axis=0)
if self.withflip == 1 or with_flip == 1:
def _postprocess_flip(im):
# Original way of flipping, changing to random_uniform way to be more controllable
# im = tf.image.random_flip_left_right(im, seed = seed_random)
# return im
do_flip = tf.random_uniform(shape=[1],
minval=0,
maxval=1,
dtype=tf.float32,
seed=seed_random)
def __left_right_flip(im):
flipped = tf.image.flip_left_right(im)
if is_normal == 1:
# flipped = 256 - flipped
flipped_x, flipped_y, flipped_z = tf.unstack(
flipped, axis=2)
flipped = tf.stack(
[256 - flipped_x, flipped_y, flipped_z],
axis=2)
return flipped
return tf.cond(tf.less(do_flip[0], 0.5),
fn1=lambda: __left_right_flip(im),
fn2=lambda: im)
crop_images = tf.map_fn(_postprocess_flip,
crop_images,
dtype=crop_images.dtype)
if with_noise == 1:
def _postprocess_noise(im):
do_noise = tf.random_uniform(shape=[1],
minval=0,
maxval=1,
dtype=tf.float32,
seed=None)
def __add_noise(im):
curr_level = tf.random_uniform(shape=[1],
minval=0,
maxval=noise_level,
dtype=tf.float32,
seed=None)
curr_noise = tf.random_normal(shape=tf.shape(im),
mean=0.0,
stddev=curr_level,
dtype=tf.float32)
return tf.add(im, curr_noise)
# return tf.cond(tf.less(do_noise[0], 0.5), true_fn = lambda: __add_noise(im), false_fn = lambda: im)
return tf.cond(tf.less(do_noise[0], 0.5),
fn1=lambda: __add_noise(im),
fn2=lambda: im)
crop_images = tf.map_fn(_postprocess_noise,
crop_images,
dtype=crop_images.dtype)
else: # not self.group=='train'
if shape_undefined == 0:
off = np.zeros(shape=[curr_batch_size, 4])
off[:, 0] = int((NOW_SIZE1 - self.crop_size) / 2)
off[:, 1] = int((NOW_SIZE2 - self.crop_size) / 2)
off[:, 2:4] = off[:, :2] + self.crop_size
off[:, 0] = off[:, 0] * 1.0 / (NOW_SIZE1 - 1)
off[:, 2] = off[:, 2] * 1.0 / (NOW_SIZE1 - 1)
off[:, 1] = off[:, 1] * 1.0 / (NOW_SIZE2 - 1)
off[:, 3] = off[:, 3] * 1.0 / (NOW_SIZE2 - 1)
box_ind = tf.constant(range(curr_batch_size))
crop_images = tf.image.crop_and_resize(
norm, off, box_ind,
tf.constant([self.crop_size, self.crop_size]))
else:
image = _aspect_preserving_resize(norm, 256)
image = _central_crop([image], self.crop_size,
self.crop_size)[0]
image.set_shape([self.crop_size, self.crop_size, 3])
crop_images = image
crop_images = tf.expand_dims(crop_images, axis=0)
crop_images = tf.cast(crop_images, orig_dtype)
if curr_batch_size == 1:
crop_images = tf.squeeze(crop_images, axis=[0])
return crop_images
def model_fn(features, labels, mode):
"""
the model_fn feeds into Estimator
"""
feature_columns = self.create_feature_columns(tf_transform_output)
input_layer = tf.feature_column.input_layer(
features=features, feature_columns=feature_columns)
# Network structure
# Batch norm after linear combination and before activation. Dropout after activation.
h1 = tf.layers.Dense(
units=MODEL_NUM_UNIT_SCALE * 4,
activation=None,
kernel_initializer=tf.glorot_normal_initializer(),
bias_initializer=tf.zeros_initializer()
)(input_layer)
h1_bn = tf.layers.batch_normalization(h1, training=(mode == tf.estimator.ModeKeys.TRAIN))
h1_act = tf.nn.relu(h1_bn)
h1_do = tf.layers.dropout(
inputs=h1_act,
rate=DROPOUT_PROB,
training=(mode == tf.estimator.ModeKeys.TRAIN))
h2 = tf.layers.Dense(
units=MODEL_NUM_UNIT_SCALE * 2,
activation=None,
kernel_initializer=tf.glorot_normal_initializer(),
bias_initializer=tf.zeros_initializer()
)(h1_do)
h2_bn = tf.layers.batch_normalization(h2, training=(mode == tf.estimator.ModeKeys.TRAIN))
h2_act = tf.nn.relu(h2_bn)
h2_do = tf.layers.dropout(
inputs=h2_act,
rate=DROPOUT_PROB,
training=(mode == tf.estimator.ModeKeys.TRAIN))
# Head for label1
h30 = tf.layers.Dense(
units=MODEL_NUM_UNIT_SCALE,
activation=None,
kernel_initializer=tf.glorot_normal_initializer(),
bias_initializer=tf.zeros_initializer()
)(h2_do)
h3_bn0 = tf.layers.batch_normalization(h30, training=(mode == tf.estimator.ModeKeys.TRAIN))
h3_act0 = tf.nn.relu(h3_bn0)
h3_do0 = tf.layers.dropout(
inputs=h3_act0,
rate=DROPOUT_PROB,
training=(mode == tf.estimator.ModeKeys.TRAIN))
logits0 = tf.layers.Dense(
units=2,
activation=None,
kernel_initializer=tf.glorot_normal_initializer(),
bias_initializer=tf.zeros_initializer()
)(h3_do0)
softmax0 = tf.contrib.layers.softmax(logits0)
q_values = tf.div(softmax0[:, 1] - tf.reduce_min(softmax0[:, 1]),
tf.reduce_max(softmax0[:, 1]) - tf.reduce_min(softmax0[:, 1]))
if mode == tf.estimator.ModeKeys.TRAIN or mode == tf.estimator.ModeKeys.EVAL:
labels0 = labels # int64 Notice: use labels but not labels[0], because we only have 1 label now.
onehot_labels0 = tf.one_hot(labels0,
depth=2) # shape(2,0) should [batch_size, num_classes] , logit should [batch_size, num_classes]
# logit(?,2)
# `ror_20_days_bool` loss definition: weighting to correct for class imbalances.
unweighted_losses0 = tf.losses.softmax_cross_entropy(
onehot_labels=onehot_labels0, logits=logits0, reduction=Reduction.NONE)
class_weights0 = tf.constant([[1., 1.]])
sample_weights0 = tf.reduce_sum(tf.multiply(onehot_labels0, class_weights0), 1)
loss0 = tf.reduce_mean(unweighted_losses0 * sample_weights0)
loss = loss0
# Metrics
auroc0 = tf.metrics.auc(labels0, softmax0[:, 1], num_thresholds=10000, curve='ROC')
prauc0 = tf.metrics.auc(labels0, softmax0[:, 1], num_thresholds=10000, curve='PR',
summation_method='careful_interpolation')
if mode == tf.estimator.ModeKeys.TRAIN:
# MSE loss, optimized with Adam
optimizer = tf.train.AdamOptimizer(FIX_LEARNING_RATE)
# This is to make sure we also update the rolling mean/var for `tf.layers.batch_normalization`
# (which is stored outside of the Estimator scope).
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(loss, global_step=tf.train.get_global_step())
# TensorBoard performance metrics.
with tf.name_scope('losses'):
tf.summary.scalar('loss_ror_20', loss0)
# TensorBoard model evolution over time.
with tf.name_scope('layer_1'):
weights = tf.get_default_graph().get_tensor_by_name(os.path.split(h1.name)[0] + '/kernel:0')
biases = tf.get_default_graph().get_tensor_by_name(os.path.split(h1.name)[0] + '/bias:0')
tf.summary.histogram('weights', weights)
tf.summary.histogram('biases', biases)
tf.summary.histogram('activations', h1_act)
with tf.name_scope('layer_2'):
weights = tf.get_default_graph().get_tensor_by_name(os.path.split(h2.name)[0] + '/kernel:0')
biases = tf.get_default_graph().get_tensor_by_name(os.path.split(h2.name)[0] + '/bias:0')
tf.summary.histogram('weights', weights)
tf.summary.histogram('biases', biases)
tf.summary.histogram('activations', h2_act)
with tf.name_scope('layer_3_ror_20'):
weights = tf.get_default_graph().get_tensor_by_name(os.path.split(h30.name)[0] + '/kernel:0')
biases = tf.get_default_graph().get_tensor_by_name(os.path.split(h30.name)[0] + '/bias:0')
tf.summary.histogram('weights', weights)
tf.summary.histogram('biases', biases)
tf.summary.histogram('activations', h3_act0)
with tf.name_scope('logits_ror_20'):
weights = tf.get_default_graph().get_tensor_by_name(
os.path.split(logits0.name)[0] + '/kernel:0')
biases = tf.get_default_graph().get_tensor_by_name(os.path.split(logits0.name)[0] + '/bias:0')
tf.summary.histogram('weights', weights)
tf.summary.histogram('biases', biases)
tf.summary.histogram('activations', h3_act0)
with tf.name_scope('q_values_ror_20'):
tf.summary.histogram('q0', softmax0[:, 0])
tf.summary.histogram('q1', softmax0[:, 1])
# Log a few predictions.label0 : ror_xxx_days_bool
# to watch the labels and softmax in training
label_and_softmax0 = tf.stack([tf.cast(labels0, tf.float32), softmax0[:, 1]], axis=1)
logging_hook = tf.train.LoggingTensorHook({
'label_and_softmax0': label_and_softmax0[0:10, :], # label_and_softmax0 size is batch size in train_config "TRAIN_BATCH_SIZE"
}, every_n_iter=LOG_FREQ_STEP)
return tf.estimator.EstimatorSpec(
mode=mode,
loss=loss,
train_op=train_op,
training_hooks=[logging_hook])
elif mode == tf.estimator.ModeKeys.EVAL:
return tf.estimator.EstimatorSpec(
mode=mode,
loss=loss,
# These metrics are computed over the complete eval dataset.
eval_metric_ops={
'metrics_ror_20_days_bool/AUC_ROC': auroc0,
'metrics_ror_20_days_bool/AUC_PR': prauc0,
}, predictions={SignatureKeys.PREDICTIONS: q_values})
elif mode == tf.estimator.ModeKeys.PREDICT:
"""
A policy derived from the Q-value network. This epsilon-greedy policy
computes the seeds with the `TOP_SEEDS_K` values and replaces them according to a
`epsilon_greedy_probability` probability with a random value in [0, 1000).
"""
# Indices of top `p.TOP_SEEDS_K` Q-values.
top_q_idx = tf.nn.top_k(q_values, k=TOP_SEEDS_K)[1]
sel_q_idx = tf.random_shuffle(top_q_idx)[0:SEEDS_K_FINAL]
# Since seeds are in [1, `p.SEEDS_K_FINAL`], we have to add 1 to the index.
predictions = sel_q_idx + 1
class_labels_ror_20 = tf.reshape(
tf.tile(tf.constant(['0', '1']), (tf.shape(softmax0)[0],)),
(tf.shape(softmax0)[0], 2))
export_outputs = {
# Default output (used in serving-infra)
# * output: Seed list. Requires using `SignatureKeys.OUTPUT` dict key, since this is
# used by the downstream SRS.
# * eps_rnd_selection: Boolean list of whether a random seed (with eps prob)
# was recommend or a predicted seed.
# * q_values: Q-values for all `SEED_LIST_LENGTH` seeds.
SignatureDefs.DEFAULT: tf.estimator.export.PredictOutput(
{SignatureKeys.OUTPUT: predictions,
"q_values": tf.transpose(q_values)}),
# Analysis output
SignatureDefs.ANALYSIS_ROR_20: tf.estimator.export.ClassificationOutput(
scores=softmax0,
classes=class_labels_ror_20),
SignatureDefs.ANALYSIS_Q: tf.estimator.export.RegressionOutput(
value=q_values)
}
return tf.estimator.EstimatorSpec(
mode=mode,
predictions={SignatureKeys.PREDICTIONS: q_values},
export_outputs=export_outputs)
def f1():
# Choose a random contiguous range from within the sequence.
range_min = tf.random_shuffle(tf.range(seq_len - batch_size))[0]
range_max = range_min + batch_size
return tf.range(range_min, range_max)
def make_data_tensor(self, train=True):
if train:
folders = self.metatrain_character_folders
# number of tasks, not number of meta-iterations. (divide by metabatch size to measure)
num_total_batches = 200000
else:
folders = self.metaval_character_folders
num_total_batches = 600
# make list of files
print('Generating filenames')
all_filenames = []
for _ in range(num_total_batches):
sampled_character_folders = random.sample(folders,
self.num_classes)
random.shuffle(sampled_character_folders)
labels_and_images = get_images(
sampled_character_folders,
range(self.num_classes),
nb_samples=self.num_samples_per_class,
shuffle=False)
# make sure the above isn't randomized order
labels = [li[0] for li in labels_and_images]
filenames = [li[1] for li in labels_and_images]
all_filenames.extend(filenames)
# make queue for tensorflow to read from
filename_queue = tf.train.string_input_producer(
tf.convert_to_tensor(all_filenames), shuffle=False)
print('Generating image processing ops')
image_reader = tf.WholeFileReader()
_, image_file = image_reader.read(filename_queue)
if FLAGS.datasource == 'miniimagenet':
image = tf.image.decode_jpeg(image_file, channels=3)
image.set_shape((self.img_size[0], self.img_size[1], 3))
image = tf.reshape(image, [self.dim_input])
image = tf.cast(image, tf.float32) / 255.0
else:
image = tf.image.decode_png(image_file)
image.set_shape((self.img_size[0], self.img_size[1], 1))
image = tf.reshape(image, [self.dim_input])
image = tf.cast(image, tf.float32) / 255.0
image = 1.0 - image # invert
num_preprocess_threads = 1 # TODO - enable this to be set to >1
min_queue_examples = 256
examples_per_batch = self.num_classes * self.num_samples_per_class
batch_image_size = self.batch_size * examples_per_batch
print('Batching images')
images = tf.train.batch(
[image],
batch_size=batch_image_size,
num_threads=num_preprocess_threads,
capacity=min_queue_examples + 3 * batch_image_size,
)
all_image_batches, all_label_batches = [], []
print('Manipulating image data to be right shape')
for i in range(self.batch_size):
image_batch = images[i * examples_per_batch:(i + 1) *
examples_per_batch]
if FLAGS.datasource == 'omniglot':
# omniglot augments the dataset by rotating digits to create new classes
# get rotation per class (e.g. 0,1,2,0,0 if there are 5 classes)
rotations = tf.multinomial(tf.log([[1., 1., 1., 1.]]),
self.num_classes)
label_batch = tf.convert_to_tensor(labels)
new_list, new_label_list = [], []
for k in range(self.num_samples_per_class):
class_idxs = tf.range(0, self.num_classes)
class_idxs = tf.random_shuffle(class_idxs)
true_idxs = class_idxs * self.num_samples_per_class + k
new_list.append(tf.gather(image_batch, true_idxs))
if FLAGS.datasource == 'omniglot': # and FLAGS.train:
new_list[-1] = tf.stack([
tf.reshape(
tf.image.rot90(tf.reshape(
new_list[-1][ind],
[self.img_size[0], self.img_size[1], 1]),
k=tf.cast(
rotations[0, class_idxs[ind]],
tf.int32)), (self.dim_input, ))
for ind in range(self.num_classes)
])
new_label_list.append(tf.gather(label_batch, true_idxs))
new_list = tf.concat(
new_list, 0
) # has shape [self.num_classes*self.num_samples_per_class, self.dim_input]
new_label_list = tf.concat(new_label_list, 0)
all_image_batches.append(new_list)
all_label_batches.append(new_label_list)
all_image_batches = tf.stack(all_image_batches)
all_label_batches = tf.stack(all_label_batches)
all_label_batches = tf.one_hot(all_label_batches, self.num_classes)
return all_image_batches, all_label_batches
def process_dataset(self, *row_parts):
row_parts = list(row_parts)
word = row_parts[0] # (, )
if not self.is_evaluating and self.config.RANDOM_CONTEXTS:
all_contexts = tf.stack(row_parts[1:])
all_contexts_padded = tf.concat([all_contexts, [self.context_pad]],
axis=-1)
index_of_blank_context = tf.where(
tf.equal(all_contexts_padded, self.context_pad))
num_contexts_per_example = tf.reduce_min(index_of_blank_context)
# if there are less than self.max_contexts valid contexts, still sample self.max_contexts
safe_limit = tf.cast(
tf.maximum(num_contexts_per_example, self.config.MAX_CONTEXTS),
tf.int32)
rand_indices = tf.random_shuffle(
tf.range(safe_limit))[:self.config.MAX_CONTEXTS]
contexts = tf.gather(all_contexts, rand_indices) # (max_contexts,)
else:
contexts = row_parts[1:(self.config.MAX_CONTEXTS +
1)] # (max_contexts,)
# contexts: (max_contexts, )
split_contexts = tf.string_split(contexts,
delimiter=",",
skip_empty=False)
sparse_split_contexts = tf.sparse.SparseTensor(
indices=split_contexts.indices,
values=split_contexts.values,
dense_shape=[self.config.MAX_CONTEXTS, 3],
)
dense_split_contexts = tf.reshape(
tf.sparse.to_dense(sp_input=sparse_split_contexts,
default_value=Common.PAD),
shape=[self.config.MAX_CONTEXTS, 3],
) # (batch, max_contexts, 3)
split_target_labels = tf.string_split(tf.expand_dims(word, -1),
delimiter="|")
target_dense_shape = [
1,
tf.maximum(
tf.cast(self.config.MAX_TARGET_PARTS, tf.int64),
# tf.to_int64(self.config.MAX_TARGET_PARTS),
split_target_labels.dense_shape[1] + 1,
),
]
sparse_target_labels = tf.sparse.SparseTensor(
indices=split_target_labels.indices,
values=split_target_labels.values,
dense_shape=target_dense_shape,
)
dense_target_label = tf.reshape(
tf.sparse.to_dense(sp_input=sparse_target_labels,
default_value=Common.PAD),
[-1],
)
index_of_blank = tf.where(tf.equal(dense_target_label, Common.PAD))
target_length = tf.reduce_min(index_of_blank)
dense_target_label = dense_target_label[:self.config.MAX_TARGET_PARTS]
clipped_target_lengths = tf.clip_by_value(
target_length,
clip_value_min=0,
clip_value_max=self.config.MAX_TARGET_PARTS)
target_word_labels = tf.concat(
[self.target_table.lookup(dense_target_label), [0]],
axis=-1) # (max_target_parts + 1) of int
path_source_strings = tf.slice(
dense_split_contexts, [0, 0],
[self.config.MAX_CONTEXTS, 1]) # (max_contexts, 1)
flat_source_strings = tf.reshape(path_source_strings,
[-1]) # (max_contexts)
split_source = tf.string_split(
flat_source_strings, delimiter="|",
skip_empty=False) # (max_contexts, max_name_parts)
sparse_split_source = tf.sparse.SparseTensor(
indices=split_source.indices,
values=split_source.values,
dense_shape=[
self.config.MAX_CONTEXTS,
tf.maximum(tf.to_int64(self.config.MAX_NAME_PARTS),
split_source.dense_shape[1]),
],
)
dense_split_source = tf.sparse.to_dense(
sp_input=sparse_split_source,
default_value=Common.PAD) # (max_contexts, max_name_parts)
dense_split_source = tf.slice(dense_split_source, [0, 0],
[-1, self.config.MAX_NAME_PARTS])
path_source_indices = self.subtoken_table.lookup(
dense_split_source) # (max_contexts, max_name_parts)
path_source_lengths = tf.reduce_sum(
tf.cast(tf.not_equal(dense_split_source, Common.PAD), tf.int32),
-1) # (max_contexts)
path_strings = tf.slice(dense_split_contexts, [0, 1],
[self.config.MAX_CONTEXTS, 1])
flat_path_strings = tf.reshape(path_strings, [-1])
split_path = tf.string_split(flat_path_strings,
delimiter="|",
skip_empty=False)
sparse_split_path = tf.sparse.SparseTensor(
indices=split_path.indices,
values=split_path.values,
dense_shape=[
self.config.MAX_CONTEXTS, self.config.MAX_PATH_LENGTH
],
)
dense_split_path = tf.sparse.to_dense(
sp_input=sparse_split_path,
default_value=Common.PAD) # (batch, max_contexts, max_path_length)
node_indices = self.node_table.lookup(
dense_split_path) # (max_contexts, max_path_length)
path_lengths = tf.reduce_sum(
tf.cast(tf.not_equal(dense_split_path, Common.PAD), tf.int32),
-1) # (max_contexts)
path_target_strings = tf.slice(
dense_split_contexts, [0, 2],
[self.config.MAX_CONTEXTS, 1]) # (max_contexts, 1)
flat_target_strings = tf.reshape(path_target_strings,
[-1]) # (max_contexts)
split_target = tf.string_split(
flat_target_strings, delimiter="|",
skip_empty=False) # (max_contexts, max_name_parts)
sparse_split_target = tf.sparse.SparseTensor(
indices=split_target.indices,
values=split_target.values,
dense_shape=[
self.config.MAX_CONTEXTS,
tf.maximum(tf.to_int64(self.config.MAX_NAME_PARTS),
split_target.dense_shape[1]),
],
)
dense_split_target = tf.sparse.to_dense(
sp_input=sparse_split_target,
default_value=Common.PAD) # (max_contexts, max_name_parts)
dense_split_target = tf.slice(dense_split_target, [0, 0],
[-1, self.config.MAX_NAME_PARTS])
path_target_indices = self.subtoken_table.lookup(
dense_split_target) # (max_contexts, max_name_parts)
path_target_lengths = tf.reduce_sum(
tf.cast(tf.not_equal(dense_split_target, Common.PAD), tf.int32),
-1) # (max_contexts)
valid_contexts_mask = tf.cast( # tf.to_float
tf.not_equal(
tf.reduce_max(path_source_indices, -1) +
tf.reduce_max(node_indices, -1) +
tf.reduce_max(path_target_indices, -1),
0,
),
dtype=tf.float32,
)
return {
TARGET_STRING_KEY: word,
TARGET_INDEX_KEY: target_word_labels,
TARGET_LENGTH_KEY: clipped_target_lengths,
PATH_SOURCE_INDICES_KEY: path_source_indices,
NODE_INDICES_KEY: node_indices,
PATH_TARGET_INDICES_KEY: path_target_indices,
VALID_CONTEXT_MASK_KEY: valid_contexts_mask,
PATH_SOURCE_LENGTHS_KEY: path_source_lengths,
PATH_LENGTHS_KEY: path_lengths,
PATH_TARGET_LENGTHS_KEY: path_target_lengths,
PATH_SOURCE_STRINGS_KEY: path_source_strings,
PATH_STRINGS_KEY: path_strings,
PATH_TARGET_STRINGS_KEY: path_target_strings,
}
y_values = []
vector_values = []
# generate random data
for i in range(num_vectors):
if np.random.random() > 0.5:
x_values.append(np.random.normal(0.4, 0.7))
y_values.append(np.random.normal(0.2, 0.8))
else:
x_values.append(np.random.normal(0.6, 0.4))
y_values.append(np.random.normal(0.8, 0.5))
vector_values = zip(x_values, y_values)
vectors = tf.constant(vector_values)
n_samples = tf.shape(vector_values)[0]
random_indices = tf.random_shuffle(tf.range(0, n_samples))
begin = [
0,
]
size = [
num_clusters,
]
size[0] = num_clusters
centroid_indices = tf.slice(random_indices, begin, size)
centroids = tf.Variable(tf.gather(vector_values, centroid_indices))
expanded_vectors = tf.expand_dims(vectors, 0)
expanded_centroids = tf.expand_dims(centroids, 1)
vectors_subtraction = tf.sub(expanded_vectors, expanded_centroids)
euclidean_distances = tf.reduce_sum(tf.square(vectors_subtraction), 2)
assignments = tf.to_int32(tf.argmin(euclidean_distances, 0))
partitions = [0, 0, 1, 1, 0]
def detection_targets_graph(config, proposals, gt_class_ids, gt_boxes,
**kwargs):
# Assertions
asserts = \
[tf.Assert(tf.greater(tf.shape(proposals)[0], 0), [proposals], name="roi_assertion"),]
with tf.control_dependencies(asserts):
proposals = tf.identity(proposals)
# Remove zero padding
proposals, _ = \
model.trim_zeros_graph(proposals, name="trim_proposals")
gt_boxes, non_zeros = \
model.trim_zeros_graph(gt_boxes, name="trim_gt_boxes")
gt_class_ids = \
model.tf.boolean_mask(gt_class_ids, non_zeros, name="trim_gt_class_ids")
# Handle COCO crowds
# A crowd box in COCO is a bounding box around several instances. Exclude
# them from training. A crowd box is given a negative class ID.
crowd_ix = tf.where(gt_class_ids < 0)[:, 0]
non_crowd_ix = tf.where(gt_class_ids > 0)[:, 0]
crowd_boxes = tf.gather(gt_boxes, crowd_ix)
gt_class_ids = tf.gather(gt_class_ids, non_crowd_ix)
gt_boxes = tf.gather(gt_boxes, non_crowd_ix)
# Compute overlaps matrix [proposals, gt_boxes]
overlaps = model.overlaps_graph(proposals, gt_boxes)
# Compute overlaps with crowd boxes [anchors, crowds]
crowd_overlaps = model.overlaps_graph(proposals, crowd_boxes)
crowd_iou_max = tf.reduce_max(crowd_overlaps, axis=1)
no_crowd_bool = (crowd_iou_max < 0.001)
# Determine postive and negative ROIs
roi_iou_max = tf.reduce_max(overlaps, axis=1)
# 1. Positive ROIs are those with >= 0.5 IoU with a GT box
positive_roi_bool = (roi_iou_max >= 0.5)
positive_indices = tf.where(positive_roi_bool)[:, 0]
# 2. Negative ROIs are those with < 0.5 with every GT box. Skip crowds.
negative_indices = tf.where(
tf.logical_and(roi_iou_max < 0.5, no_crowd_bool))[:, 0]
# Subsample ROIs. Aim for 33% positive
# Positive ROIs
positive_count = int(config.TRAIN_ROIS_PER_IMAGE *
config.ROI_POSITIVE_RATIO)
positive_indices = tf.random_shuffle(positive_indices)[:positive_count]
positive_count = tf.shape(positive_indices)[0]
# Negative ROIs. Add enough to maintain positive:negative ratio.
r = 1.0 / config.ROI_POSITIVE_RATIO
negative_count = tf.cast(r * tf.cast(positive_count, tf.float32),
tf.int32) - positive_count
negative_indices = tf.random_shuffle(negative_indices)[:negative_count]
# Gather selected ROIs
positive_rois = tf.gather(proposals, positive_indices)
negative_rois = tf.gather(proposals, negative_indices)
# Assign positive ROIs to GT boxes.
positive_overlaps = tf.gather(overlaps, positive_indices)
roi_gt_box_assignment = tf.argmax(positive_overlaps, axis=1)
roi_gt_boxes = tf.gather(gt_boxes, roi_gt_box_assignment)
roi_gt_class_ids = tf.gather(gt_class_ids, roi_gt_box_assignment)
# Compute bbox refinement for positive ROIs
deltas = utils.box_refinement_graph(positive_rois, roi_gt_boxes)
deltas /= config.BBOX_STD_DEV
# Compute mask targets
boxes = positive_rois
# Append negative ROIs and pad bbox deltas and masks that
# are not used for negative ROIs with zeros.
rois = tf.concat([positive_rois, negative_rois], axis=0)
N = tf.shape(negative_rois)[0]
P = tf.maximum(config.TRAIN_ROIS_PER_IMAGE - tf.shape(rois)[0], 0)
rois = tf.pad(rois, [(0, P), (0, 0)])
roi_gt_class_ids = tf.pad(roi_gt_class_ids, [(0, N + P)])
deltas = tf.pad(deltas, [(0, N + P), (0, 0)])
if 'gt_rboxes' in kwargs:
gt_rboxes = kwargs.get('gt_rboxes')
gt_rboxes, _ = model.trim_zeros_graph(gt_rboxes, name="trim_gt_rboxes")
gt_rboxes = tf.gather(gt_rboxes, non_crowd_ix)
roi_gt_rboxes = tf.gather(gt_rboxes, roi_gt_box_assignment)
if config.regressor == "deltas":
rbox_deltas = line_to_deltas(roi_gt_boxes, roi_gt_rboxes)
rbox_deltas /= 0.2
rbox_deltas = tf.pad(rbox_deltas, [(0, N + P), (0, 0)])
return rois, roi_gt_class_ids, deltas, rbox_deltas
elif config.regressor == "rotdim":
gt_angles = kwargs.get('gt_angles')
gt_angles = tf.boolean_mask(gt_angles,
non_zeros,
name="trim_gt_angles")
gt_angles = tf.gather(gt_angles, non_crowd_ix)
roi_gt_angles = tf.gather(gt_angles, roi_gt_box_assignment)
rbox_angles = roi_gt_angles / 0.2
# point 1
y1 = roi_gt_rboxes[:, 0]
x1 = roi_gt_rboxes[:, 1]
# point 2
y2 = roi_gt_rboxes[:, 2]
x2 = roi_gt_rboxes[:, 3]
# point 3
y3 = roi_gt_rboxes[:, 4]
x3 = roi_gt_rboxes[:, 5]
rw = tf.sqrt(tf.pow(x2 - x1, 2) + tf.pow(y2 - y1, 2))
rh = tf.sqrt(tf.pow(x3 - x2, 2) + tf.pow(y3 - y2, 2))
rbox_dim = tf.stack([rh, rw], axis=1)
rbox_dim /= 0.1
rbox_dim = tf.pad(rbox_dim, [(0, N + P), (0, 0)])
rbox_angles = tf.pad(rbox_angles, [(0, N + P), (0, 0)])
return rois, roi_gt_class_ids, deltas, rbox_angles, rbox_dim
elif config.regressor == "verts":
roi_gt_rboxes = tf.pad(roi_gt_rboxes, [(0, N + P), (0, 0)])
return rois, roi_gt_class_ids, deltas, roi_gt_rboxes
return rois, roi_gt_class_ids, deltas
import matplotlib.pyplot as plt
import pandas as pd
import seaborn as sns
df = pd.DataFrame({'x': [v[0] for v in vectors_set],
'y': [v[1] for v in vectors_set]})
sns.lmplot('x', 'y', data=df, fit_reg=False, size=6)
plt.show()
import tensorflow as tf
vectors = tf.constant(vectors_set)
k = 4
centroides = tf.Variable(tf.slice(tf.random_shuffle(vectors),[0,0],[k,-1]))
expanded_vectors = tf.expand_dims(vectors, 0)
expanded_centroides = tf.expand_dims(centroides, 0)
assignments = tf.argmin(tf.reduce_sum(tf.square(tf.sub(expanded_vectors,
expanded_centroides)),2),0)
means = tf.concat(0, [tf.reduce_sum(tf.gather(vectors, tf.reshape(tf.where(tf.equal(assignments,c)),
[1,-1])),
reduction_indices=[1]) for c in range(k)])
update_centroides = tf.assign(centroides, means)
init_op = tf.initialize_all_variables()
sess = tf.Session()
sess.run(init_op)
def build_model(self,
is_training=True,
inst_norm=False,
no_target_source=False):
real_data = tf.placeholder(tf.float32, [
self.batch_size, self.input_width, self.input_width,
self.input_filters + self.output_filters
],
name='real_A_and_B_images')
embedding_ids = tf.placeholder(tf.int64,
shape=None,
name="embedding_ids")
no_target_data = tf.placeholder(tf.float32, [
self.batch_size, self.input_width, self.input_width,
self.input_filters + self.output_filters
],
name='no_target_A_and_B_images')
no_target_ids = tf.placeholder(tf.int64,
shape=None,
name="no_target_embedding_ids")
# target images
real_A = real_data[:, :, :, :self.input_filters]
# source images
real_B = real_data[:, :, :, self.input_filters:self.input_filters +
self.output_filters]
real_A_shuffle = tf.random_shuffle(real_A)
embedding = init_embedding(self.embedding_num, self.embedding_dim)
fake_target, target_gaussian1, target_gaussian2, source_e8, layers_source = self.generator(
real_B,
real_A,
embedding_ids,
is_training=is_training,
inst_norm=inst_norm,
reuse=False)
fake_target_shuffle, target_gaussian1_shuffle, target_gaussian2_shuffle = self.generator_gaussian(
layers_source,
real_A_shuffle,
embedding_ids,
is_training=is_training,
inst_norm=inst_norm,
reuse=True)
real_A_shuffle2 = tf.random_shuffle(real_A_shuffle)
fake_target_shuffle2, target_gaussian1_shuffle2, target_gaussian2_shuffle2 = self.generator_gaussian(
layers_source,
real_A_shuffle2,
embedding_ids,
is_training=is_training,
inst_norm=inst_norm,
reuse=True)
real_A_shuffle3 = tf.random_shuffle(real_A_shuffle2)
fake_target_shuffle3, target_gaussian1_shuffle3, target_gaussian2_shuffle3 = self.generator_gaussian(
layers_source,
real_A_shuffle3,
embedding_ids,
is_training=is_training,
inst_norm=inst_norm,
reuse=True)
real_A_shuffle4 = tf.random_shuffle(real_A_shuffle3)
fake_target_shuffle4, target_gaussian1_shuffle4, target_gaussian2_shuffle4 = self.generator_gaussian(
layers_source,
real_A_shuffle4,
embedding_ids,
is_training=is_training,
inst_norm=inst_norm,
reuse=True)
real_A_shuffle5 = tf.random_shuffle(real_A_shuffle4)
fake_target_shuffle5, target_gaussian1_shuffle5, target_gaussian2_shuffle5 = self.generator_gaussian(
layers_source,
real_A_shuffle5,
embedding_ids,
is_training=is_training,
inst_norm=inst_norm,
reuse=True)
source_fake_e8, layers_source_fake = self.encoder(
fake_target_shuffle, is_training=is_training, reuse=True)
target_gaussian1_fake = self.gaussion_encoder(fake_target_shuffle,
layers_source,
is_training=is_training,
reuse=True)
#real_AB = tf.concat([real_A, real_B], 3)
#fake_AB = tf.concat([real_A, fake_B], 3)
# Note it is not possible to set reuse flag back to False
# initialize all variables before setting reuse to True
#real_D, real_D_logits, real_category_logits = self.discriminator(real_AB, is_training=is_training, reuse=False)
#fake_D, fake_D_logits, fake_category_logits = self.discriminator(fake_AB, is_training=is_training, reuse=True)
# encoding constant loss
# this loss assume that generated imaged and real image
# should reside in the same space and close to each other
#encoded_fake_B = self.encoder(fake_B, is_training, reuse=True)[0]
#const_loss = (tf.reduce_mean(tf.square(encoded_real_A - encoded_fake_B))) * self.Lconst_penalty
# category loss
#true_labels = tf.reshape(tf.one_hot(indices=embedding_ids, depth=self.embedding_num),
# shape=[self.batch_size, self.embedding_num])
#real_category_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=target_gaussian1_fake,
# labels=target_gaussian1))
real_category_loss = tf.reduce_sum(
tf.abs(target_gaussian1_fake[:, :self.latent_dim] -
target_gaussian1_shuffle[:, :self.latent_dim]))
#fake_category_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=fake_category_logits,
# labels=true_labels))
#category_loss = self.Lcategory_penalty * (real_category_loss + fake_category_loss)
# binary real/fake loss
#d_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=real_D_logits,
# labels=tf.ones_like(real_D)))
#d_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=fake_D_logits,
# labels=tf.zeros_like(fake_D)))
# L1 loss between real and generated images
l1_loss = self.L1_penalty * tf.reduce_mean(
tf.abs(fake_target_shuffle - real_A), [1, 2, 3])
l1_loss2 = self.L1_penalty * tf.reduce_mean(
tf.abs(fake_target_shuffle2 - real_A), [1, 2, 3])
l1_loss3 = self.L1_penalty * tf.reduce_mean(
tf.abs(fake_target_shuffle3 - real_A), [1, 2, 3])
l1_loss4 = self.L1_penalty * tf.reduce_mean(
tf.abs(fake_target_shuffle4 - real_A), [1, 2, 3])
l1_loss5 = self.L1_penalty * tf.reduce_mean(
tf.abs(fake_target_shuffle5 - real_A), [1, 2, 3])
# total variation loss
#width = self.output_width
#tv_loss = (tf.nn.l2_loss(fake_B[:, 1:, :, :] - fake_B[:, :width - 1, :, :]) / width
# + tf.nn.l2_loss(fake_B[:, :, 1:, :] - fake_B[:, :, :width - 1, :]) / width) * self.Ltv_penalty
# maximize the chance generator fool the discriminator
#cheat_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=fake_D_logits,
# labels=tf.ones_like(fake_D)))
#d_loss = d_loss_real + d_loss_fake + category_loss / 2.0
q_z = distributions.Normal(
loc=target_gaussian1_shuffle[:, :self.latent_dim],
scale=tf.nn.softplus(target_gaussian1_shuffle[:,
self.latent_dim:]))
q_z2 = distributions.Normal(
loc=target_gaussian1_shuffle2[:, :self.latent_dim],
scale=tf.nn.softplus(target_gaussian1_shuffle2[:,
self.latent_dim:]))
q_z3 = distributions.Normal(
loc=target_gaussian1_shuffle3[:, :self.latent_dim],
scale=tf.nn.softplus(target_gaussian1_shuffle3[:,
self.latent_dim:]))
q_z4 = distributions.Normal(
loc=target_gaussian1_shuffle4[:, :self.latent_dim],
scale=tf.nn.softplus(target_gaussian1_shuffle4[:,
self.latent_dim:]))
q_z5 = distributions.Normal(
loc=target_gaussian1_shuffle5[:, :self.latent_dim],
scale=tf.nn.softplus(target_gaussian1_shuffle5[:,
self.latent_dim:]))
# print(output.get_shape())
p_z = distributions.Normal(loc=np.zeros(self.latent_dim,
dtype=np.float32),
scale=np.ones(self.latent_dim,
dtype=np.float32))
kl_loss = tf.reduce_sum(distributions.kl_divergence(q_z, p_z), 1)
kl_loss2 = tf.reduce_sum(distributions.kl_divergence(q_z2, p_z), 1)
kl_loss3 = tf.reduce_sum(distributions.kl_divergence(q_z3, p_z), 1)
kl_loss4 = tf.reduce_sum(distributions.kl_divergence(q_z4, p_z), 1)
kl_loss5 = tf.reduce_sum(distributions.kl_divergence(q_z5, p_z), 1)
kl_loss_1 = tf.reduce_sum(
tf.reduce_sum(distributions.kl_divergence(q_z, p_z), 1))
#const_loss = (tf.reduce_mean(tf.reduce_sum(tf.square(lay_AB["e1"] - lay_A["e1"]))+tf.reduce_sum(tf.square(lay_AB["e2"] - lay_A["e2"]))+tf.reduce_sum(tf.square(lay_AB["e3"] - lay_A["e3"]))+tf.reduce_sum(tf.square(lay_AB["e4"] - lay_A["e4"]))+tf.reduce_sum(tf.square(lay_AB["e5"] - lay_A["e5"]))+tf.reduce_sum(tf.square(lay_AB["e6"] - lay_A["e6"]))+tf.reduce_sum(tf.square(lay_AB["e7"] - lay_A["e7"])))) * self.Lconst_penalty
const_loss1 = tf.reduce_mean(
tf.square(layers_source_fake["e8"] - layers_source["e8"]),
[1, 2, 3])
const_loss2 = 10.0 * tf.reduce_mean(
tf.square(layers_source_fake["e7"] - layers_source["e7"]),
[1, 2, 3])
const_loss3 = 5.0 * tf.reduce_mean(
tf.square(layers_source_fake["e6"] - layers_source["e6"]),
[1, 2, 3])
const_loss4 = tf.reduce_mean(
tf.square(layers_source_fake["e5"] - layers_source["e5"]),
[1, 2, 3])
const_loss = tf.reduce_sum(const_loss1 + const_loss2 + const_loss3 +
const_loss4) * self.Lconst_penalty
const_loss = (tf.reduce_sum(const_loss1)) * self.Lconst_penalty
T_loss = tf.reduce_sum(l1_loss + kl_loss)
T_loss2 = tf.reduce_sum(l1_loss2 + kl_loss2)
T_loss3 = tf.reduce_sum(l1_loss3 + kl_loss3)
T_loss4 = tf.reduce_sum(l1_loss4 + kl_loss4)
T_loss5 = tf.reduce_sum(l1_loss5 + kl_loss5)
#d_loss_real_summary = tf.summary.scalar("d_loss_real", d_loss_real)
#d_loss_fake_summary = tf.summary.scalar("d_loss_fake", d_loss_fake)
#category_loss_summary = tf.summary.scalar("category_loss", category_loss)
#cheat_loss_summary = tf.summary.scalar("cheat_loss", cheat_loss)
#l1_loss_summary = tf.summary.scalar("l1_loss", l1_loss)
real_category_loss_summary = tf.summary.scalar(
"real_category_loss", tf.reduce_sum(real_category_loss))
#const_loss_summary = tf.summary.scalar("const_loss", const_loss)
#d_loss_summary = tf.summary.scalar("d_loss", d_loss)
T_loss_summary = tf.summary.scalar("T_loss", T_loss)
T_loss2_summary = tf.summary.scalar("T_loss2", T_loss2)
T_loss3_summary = tf.summary.scalar("T_loss3", T_loss3)
T_loss4_summary = tf.summary.scalar("T_loss4", T_loss4)
T_loss5_summary = tf.summary.scalar("T_loss5", T_loss5)
l1_loss_summary = tf.summary.scalar("l1_loss", tf.reduce_sum(l1_loss))
l1_loss2_summary = tf.summary.scalar("l1_loss2",
tf.reduce_sum(l1_loss2))
kl_loss_summary = tf.summary.scalar("kl_loss", kl_loss_1)
#kl_loss2_summary = tf.summary.scalar("kl_loss2", kl_loss2_1)
#kl_loss_fake_AB_summary = tf.summary.scalar("kl_loss_fake_AB", kl_loss_fake_AB_1)
const_loss_summary = tf.summary.scalar("const_loss_loss", const_loss)
All_merged_summary = tf.summary.merge([
l1_loss_summary, l1_loss2_summary, T_loss_summary, kl_loss_summary,
const_loss_summary, real_category_loss_summary, T_loss2_summary,
T_loss3_summary, T_loss4_summary, T_loss5_summary
])
#tv_loss_summary = tf.summary.scalar("tv_loss", tv_loss)
#d_merged_summary = tf.summary.merge([d_loss_real_summary, d_loss_fake_summary,
# category_loss_summary, d_loss_summary])
#g_merged_summary = tf.summary.merge([cheat_loss_summary, l1_loss_summary,
# fake_category_loss_summary,
# const_loss_summary,
# g_loss_summary, tv_loss_summary])
# expose useful nodes in the graph as handles globally
input_handle = InputHandle(real_data=real_data,
embedding_ids=embedding_ids,
no_target_data=no_target_data,
no_target_ids=no_target_ids)
loss_handle = LossHandle(
T_loss=T_loss,
T_loss2=T_loss2,
T_loss3=T_loss3,
T_loss4=T_loss4,
T_loss5=T_loss5,
#const_loss=const_loss,
kl_loss=kl_loss,
l1_loss=l1_loss
#category_loss=category_loss,
#cheat_loss=cheat_loss,
)
eval_handle = EvalHandle(encoder=target_gaussian2,
generator=fake_target,
generator2=fake_target_shuffle,
target=real_A,
source=real_B,
embedding=embedding,
gaussian_params=target_gaussian1_shuffle)
summary_handle = SummaryHandle(T_sum=All_merged_summary)
# those operations will be shared, so we need
# to make them visible globally
setattr(self, "input_handle", input_handle)
setattr(self, "loss_handle", loss_handle)
setattr(self, "eval_handle", eval_handle)
setattr(self, "summary_handle", summary_handle)
def detection_targets_graph(proposals, gt_class_ids, gt_boxes, config):
"""Generates detection targets for one image. Subsamples proposals and
generates target class IDs, bounding box deltas for each.
Inputs:
proposals: [N, (y1, x1, y2, x2)] in normalized coordinates. Might
be zero padded if there are not enough proposals.
gt_class_ids: [MAX_GT_INSTANCES] int class IDs
gt_boxes: [MAX_GT_INSTANCES, (y1, x1, y2, x2)] in normalized coordinates.
Returns: Target ROIs and corresponding class IDs, bounding box shifts
rois: [TRAIN_ROIS_PER_IMAGE, (y1, x1, y2, x2)] in normalized coordinates
class_ids: [TRAIN_ROIS_PER_IMAGE]. Integer class IDs. Zero padded.
deltas: [TRAIN_ROIS_PER_IMAGE, NUM_CLASSES, (dy, dx, log(dh), log(dw))]
Class-specific bbox refinments.
Note: Returned arrays might be zero padded if not enough target ROIs.
"""
# Assertions
asserts = [
tf.Assert(tf.greater(tf.shape(proposals)[0], 0), [proposals],
name="roi_assertion"),
]
with tf.control_dependencies(asserts):
proposals = tf.identity(proposals)
# Remove zero padding
proposals, _ = trim_zeros_graph(proposals, name="trim_proposals")
gt_boxes, non_zeros = trim_zeros_graph(gt_boxes, name="trim_gt_boxes")
gt_class_ids = tf.boolean_mask(gt_class_ids,
non_zeros,
name="trim_gt_class_ids")
# Compute overlaps matrix [proposals, gt_boxes]
overlaps = overlaps_graph(proposals, gt_boxes)
# Determine postive and negative ROIs
roi_iou_max = tf.reduce_max(overlaps, axis=1)
# 1. Positive ROIs are those with >= 0.5 IoU with a GT box
positive_roi_bool = (roi_iou_max >= 0.5)
positive_indices = tf.where(positive_roi_bool)[:, 0]
# 2. Negative ROIs are those with < 0.5 with every GT box. Skip crowds.
negative_indices = tf.where(roi_iou_max < 0.5)[:, 0]
# Subsample ROIs. Aim for 33% positive
# Positive ROIs
positive_count = int(config.TRAIN_ROIS_PER_IMAGE *
config.ROI_POSITIVE_RATIO)
positive_indices = tf.random_shuffle(positive_indices)[:positive_count]
positive_count = tf.shape(positive_indices)[0]
# Negative ROIs. Add enough to maintain positive:negative ratio.
r = 1.0 / config.ROI_POSITIVE_RATIO
negative_count = tf.cast(r * tf.cast(positive_count, tf.float32),
tf.int32) - positive_count
negative_indices = tf.random_shuffle(negative_indices)[:negative_count]
# Gather selected ROIs
positive_rois = tf.gather(proposals, positive_indices)
negative_rois = tf.gather(proposals, negative_indices)
# Assign positive ROIs to GT boxes.
positive_overlaps = tf.gather(overlaps, positive_indices)
roi_gt_box_assignment = tf.argmax(positive_overlaps, axis=1)
roi_gt_boxes = tf.gather(gt_boxes, roi_gt_box_assignment)
roi_gt_class_ids = tf.gather(gt_class_ids, roi_gt_box_assignment)
# Compute bbox refinement for positive ROIs
deltas = KerasRFCN.Utils.box_refinement_graph(positive_rois, roi_gt_boxes)
deltas /= config.BBOX_STD_DEV
# Append negative ROIs and pad bbox deltas and masks that
# are not used for negative ROIs with zeros.
rois = tf.concat([positive_rois, negative_rois], axis=0)
N = tf.shape(negative_rois)[0]
P = tf.maximum(config.TRAIN_ROIS_PER_IMAGE - tf.shape(rois)[0], 0)
rois = tf.pad(rois, [(0, P), (0, 0)])
roi_gt_boxes = tf.pad(roi_gt_boxes, [(0, N + P), (0, 0)])
roi_gt_class_ids = tf.pad(roi_gt_class_ids, [(0, N + P)])
deltas = tf.pad(deltas, [(0, N + P), (0, 0)])
return rois, roi_gt_class_ids, deltas
def _get_randomized_indices(self):
"""Generates randomized indices into a sequence of a specific length."""
indices = tf.range(0, self._dataset_info.sequence_size)
indices = tf.random_shuffle(indices)
indices = tf.slice(indices, begin=[0], size=[self._example_size])
return indices
def crop_proposal():
rand_vec = lambda minval, maxval: tf.random_uniform(shape=(
ssd_constants.NUM_CROP_PASSES, 1),
minval=minval,
maxval=maxval,
dtype=tf.float32)
width, height = rand_vec(0.3, 1), rand_vec(0.3, 1)
left, top = rand_vec(0, 1 - width), rand_vec(0, 1 - height)
right = left + width
bottom = top + height
ltrb = tf.concat([left, top, right, bottom], axis=1)
min_iou = tf.random_shuffle(ssd_constants.CROP_MIN_IOU_CHOICES)[0]
ious = calc_iou_tensor(ltrb, boxes)
# discard any bboxes whose center not in the cropped image
xc, yc = [
tf.tile(0.5 * (boxes[:, i + 0] + boxes[:, i + 2])[tf.newaxis, :],
(ssd_constants.NUM_CROP_PASSES, 1)) for i in range(2)
]
masks = tf.reduce_all(tf.stack([
tf.greater(xc, tf.tile(left, (1, num_boxes))),
tf.less(xc, tf.tile(right, (1, num_boxes))),
tf.greater(yc, tf.tile(top, (1, num_boxes))),
tf.less(yc, tf.tile(bottom, (1, num_boxes))),
],
axis=2),
axis=2)
# Checks of whether a crop is valid.
valid_aspect = tf.logical_and(tf.less(height / width, 2),
tf.less(height / width, 2))
valid_ious = tf.reduce_all(tf.greater(ious, min_iou),
axis=1,
keepdims=True)
valid_masks = tf.reduce_any(masks, axis=1, keepdims=True)
valid_all = tf.cast(
tf.reduce_all(tf.concat([valid_aspect, valid_ious, valid_masks],
axis=1),
axis=1), tf.int32)
# One indexed, as zero is needed for the case of no matches.
index = tf.range(1, 1 + ssd_constants.NUM_CROP_PASSES, dtype=tf.int32)
# Either one-hot, or zeros if there is no valid crop.
selection = tf.equal(tf.reduce_max(index * valid_all), index)
use_crop = tf.reduce_any(selection)
output_ltrb = tf.reduce_sum(tf.multiply(
ltrb,
tf.tile(tf.cast(selection, tf.float32)[:, tf.newaxis], (1, 4))),
axis=0)
output_masks = tf.reduce_any(tf.logical_and(
masks, tf.tile(selection[:, tf.newaxis], (1, num_boxes))),
axis=0)
return use_crop, output_ltrb, output_masks
def detection_targets_graph(proposals, gt_class_ids, gt_boxes, gt_masks,
config):
"""Generates detection targets for one image. Subsamples proposals and
generates target class IDs, bounding box deltas, and masks for each.
Inputs:
proposals: [N, (y1, x1, y2, x2)] in normalized coordinates. Might
be zero padded if there are not enough proposals.
gt_class_ids: [MAX_GT_INSTANCES] int class IDs
gt_boxes: [MAX_GT_INSTANCES, (y1, x1, y2, x2)] in normalized coordinates.
gt_masks: [height, width, MAX_GT_INSTANCES] of boolean type.
Returns: Target ROIs and corresponding class IDs, bounding box shifts,
and masks.
rois: [TRAIN_ROIS_PER_IMAGE, (y1, x1, y2, x2)] in normalized coordinates
class_ids: [TRAIN_ROIS_PER_IMAGE]. Integer class IDs. Zero padded.
deltas: [TRAIN_ROIS_PER_IMAGE, NUM_CLASSES, (dy, dx, log(dh), log(dw))]
Class-specific bbox refinements.
masks: [TRAIN_ROIS_PER_IMAGE, height, width). Masks cropped to bbox
boundaries and resized to neural network output size.
Note: Returned arrays might be zero padded if not enough target ROIs.
"""
# Assertions
asserts = [
tf.Assert(tf.greater(tf.shape(proposals)[0], 0), [proposals],
name="roi_assertion"),
]
with tf.control_dependencies(asserts):
proposals = tf.identity(proposals)
# Remove zero padding
proposals, _ = trim_zeros_graph(proposals, name="trim_proposals")
gt_boxes, non_zeros = trim_zeros_graph(gt_boxes, name="trim_gt_boxes")
gt_class_ids = tf.boolean_mask(gt_class_ids,
non_zeros,
name="trim_gt_class_ids")
gt_masks = tf.gather(gt_masks,
tf.where(non_zeros)[:, 0],
axis=2,
name="trim_gt_masks")
# Handle COCO crowds
# A crowd box in COCO is a bounding box around several instances. Exclude
# them from training. A crowd box is given a negative class ID.
crowd_ix = tf.where(gt_class_ids < 0)[:, 0]
non_crowd_ix = tf.where(gt_class_ids > 0)[:, 0]
crowd_boxes = tf.gather(gt_boxes, crowd_ix)
crowd_masks = tf.gather(gt_masks, crowd_ix, axis=2)
gt_class_ids = tf.gather(gt_class_ids, non_crowd_ix)
gt_boxes = tf.gather(gt_boxes, non_crowd_ix)
gt_masks = tf.gather(gt_masks, non_crowd_ix, axis=2)
# Compute overlaps matrix [proposals, gt_boxes]
overlaps = overlaps_graph(proposals, gt_boxes)
# Compute overlaps with crowd boxes [anchors, crowds]
crowd_overlaps = overlaps_graph(proposals, crowd_boxes)
crowd_iou_max = tf.reduce_max(crowd_overlaps, axis=1)
no_crowd_bool = (crowd_iou_max < 0.001)
# Determine positive and negative ROIs
roi_iou_max = tf.reduce_max(overlaps, axis=1)
# 1. Positive ROIs are those with >= 0.5 IoU with a GT box
positive_roi_bool = (roi_iou_max >= 0.5)
positive_indices = tf.where(positive_roi_bool)[:, 0]
# 2. Negative ROIs are those with < 0.5 with every GT box. Skip crowds.
negative_indices = tf.where(
tf.logical_and(roi_iou_max < 0.5, no_crowd_bool))[:, 0]
# Subsample ROIs. Aim for 33% positive
# Positive ROIs
positive_count = int(config.TRAIN_ROIS_PER_IMAGE *
config.ROI_POSITIVE_RATIO)
positive_indices = tf.random_shuffle(positive_indices)[:positive_count]
positive_count = tf.shape(positive_indices)[0]
# Negative ROIs. Add enough to maintain positive:negative ratio.
r = 1.0 / config.ROI_POSITIVE_RATIO
negative_count = tf.cast(r * tf.cast(positive_count, tf.float32),
tf.int32) - positive_count
negative_indices = tf.random_shuffle(negative_indices)[:negative_count]
# Gather selected ROIs
positive_rois = tf.gather(proposals, positive_indices)
negative_rois = tf.gather(proposals, negative_indices)
# Assign positive ROIs to GT boxes.
positive_overlaps = tf.gather(overlaps, positive_indices)
roi_gt_box_assignment = tf.cond(
tf.greater(tf.shape(positive_overlaps)[1], 0),
true_fn=lambda: tf.argmax(positive_overlaps, axis=1),
false_fn=lambda: tf.cast(tf.constant([]), tf.int64))
roi_gt_boxes = tf.gather(gt_boxes, roi_gt_box_assignment)
roi_gt_class_ids = tf.gather(gt_class_ids, roi_gt_box_assignment)
# Compute bbox refinement for positive ROIs
deltas = utils.box_refinement_graph(positive_rois, roi_gt_boxes)
deltas /= config.BBOX_STD_DEV
# Assign positive ROIs to GT masks
# Permute masks to [N, height, width, 1]
transposed_masks = tf.expand_dims(tf.transpose(gt_masks, [2, 0, 1]), -1)
# Pick the right mask for each ROI
roi_masks = tf.gather(transposed_masks, roi_gt_box_assignment)
# Compute mask targets
boxes = positive_rois
if config.USE_MINI_MASK:
# Transform ROI coordinates from normalized image space
# to normalized mini-mask space.
y1, x1, y2, x2 = tf.split(positive_rois, 4, axis=1)
gt_y1, gt_x1, gt_y2, gt_x2 = tf.split(roi_gt_boxes, 4, axis=1)
gt_h = gt_y2 - gt_y1
gt_w = gt_x2 - gt_x1
y1 = (y1 - gt_y1) / gt_h
x1 = (x1 - gt_x1) / gt_w
y2 = (y2 - gt_y1) / gt_h
x2 = (x2 - gt_x1) / gt_w
boxes = tf.concat([y1, x1, y2, x2], 1)
box_ids = tf.range(0, tf.shape(roi_masks)[0])
masks = tf.image.crop_and_resize(tf.cast(roi_masks, tf.float32), boxes,
box_ids, config.MASK_SHAPE)
# Remove the extra dimension from masks.
masks = tf.squeeze(masks, axis=3)
# Threshold mask pixels at 0.5 to have GT masks be 0 or 1 to use with
# binary cross entropy loss.
masks = tf.round(masks)
# Append negative ROIs and pad bbox deltas and masks that
# are not used for negative ROIs with zeros.
rois = tf.concat([positive_rois, negative_rois], axis=0)
N = tf.shape(negative_rois)[0]
P = tf.maximum(config.TRAIN_ROIS_PER_IMAGE - tf.shape(rois)[0], 0)
rois = tf.pad(rois, [(0, P), (0, 0)])
roi_gt_boxes = tf.pad(roi_gt_boxes, [(0, N + P), (0, 0)])
roi_gt_class_ids = tf.pad(roi_gt_class_ids, [(0, N + P)])
deltas = tf.pad(deltas, [(0, N + P), (0, 0)])
masks = tf.pad(masks, [[0, N + P], (0, 0), (0, 0)])
return rois, roi_gt_class_ids, deltas, masks
def _make_dataset(self,
binaries_fname_pattern,
data_augmentation=False,
shuffle=True):
"""Creates a CIFAR-100 data set (helper used by ``.make_*_datset`` below).
Args:
binaries_fname_pattern (str): Pattern of the ``.bin`` files from which
to load images and labels (e.g. ``some/path/data_batch_*.bin``).
data_augmentation (bool): Whether to apply data augmentation operations.
shuffle (bool): Switch to turn on or off shuffling of the data set.
Defaults to ``True``.
Returns:
A tf.data.Dataset yielding batches of CIFAR-100 data.
"""
# Set number of bytes to read.
label_bytes = 1
label_offset = 1
num_classes = 100
depth = 3
image_size = 32
image_bytes = image_size * image_size * depth
record_bytes = label_bytes + label_offset + image_bytes
def parse_func(raw_record):
"""Function parsing data from raw binary records."""
# Decode raw_record.
record = tf.reshape(tf.decode_raw(raw_record, tf.uint8),
[record_bytes])
label = tf.cast(tf.slice(record, [label_offset], [label_bytes]),
tf.int32)
depth_major = tf.reshape(
tf.slice(record, [label_bytes], [image_bytes]),
[depth, image_size, image_size])
image = tf.cast(tf.transpose(depth_major, [1, 2, 0]), tf.float32)
# Add image pre-processing.
if data_augmentation:
image = tf.image.resize_image_with_crop_or_pad(
image, image_size + 4, image_size + 4)
image = tf.random_crop(image, [32, 32, 3])
image = tf.image.random_flip_left_right(image)
image = tf.image.random_brightness(image, max_delta=63. / 255.)
image = tf.image.random_saturation(image, lower=0.5, upper=1.5)
image = tf.image.random_contrast(image, lower=0.2, upper=1.8)
else:
image = tf.image.resize_image_with_crop_or_pad(image, 32, 32)
image = tf.image.per_image_standardization(image)
label = tf.squeeze(tf.one_hot(label, depth=num_classes))
return image, label
with tf.name_scope(self._name):
with tf.device('/cpu:0'):
filenames = tf.matching_files(binaries_fname_pattern)
filenames = tf.random_shuffle(filenames)
data = tf.data.FixedLengthRecordDataset(
filenames=filenames, record_bytes=record_bytes)
data = data.map(
parse_func,
num_parallel_calls=(8 if data_augmentation else 4))
if shuffle:
data = data.shuffle(buffer_size=20000)
data = data.batch(self._batch_size, drop_remainder=True)
data = data.prefetch(buffer_size=4)
return data
else:
x = np.random.normal(3.0, 0.6)
y = np.random.normal(1.0, 0.6)
xy.append([x, y])
return xy
# input data tensor를 생성한다
n = 1000
tf.reset_default_graph()
inputXY = tf.constant(createData(n))
# 초기 중심 좌표를 설정한다.
k = 5
tmpXY = tf.random_shuffle(inputXY) # input data를 shuffling 한다
tmpCent = tf.slice(tmpXY, [0, 0], [k, -1]) # 앞의 k개를 중심좌표로 설정한다
initCent = tf.Variable(tmpCent) # 초기 중심좌표 텐서
# 데이터 ~ 중심좌표 사이의 거래 계산을 위해 dimension을 맞춘다. 3-차원 텐서.
#expXY.get_shape() --> (D0, D1, D2) = (1, n, 2) : D0을 확장함
#expCent.get_shape() --> (D0, D1, D2) = (k, 1, 2) : D1을 확장함
expXY = tf.expand_dims(inputXY, 0) # D0-축을 확장한다
expCent = tf.expand_dims(initCent, 1) # D1-축을 확장한다
# 데이터와 중삼좌표 사이의 거리를 계산한다
tmpDist = tf.square(tf.subtract(expXY, expCent))
dist = tf.sqrt(tf.reduce_sum(tmpDist, 2)) # D2 축으로 합침
error = tf.reduce_sum(dist) # 거리의 총합을 error로 정의한다
# 각 데이터를 거리가 작은 중점에 assign 한다.
def main():
parser = argparse.ArgumentParser()
parser.add_argument("-l",
"--learner-hosts",
type=str,
required=True,
help="Comma-separated list of hostname:port pairs")
parser.add_argument("-e",
"--explorer-hosts",
type=str,
required=True,
help="Comma-separated list of hostname:port pairs")
parser.add_argument("-j",
"--job-name",
type=str,
default="learner",
help="One of 'learner', 'explorer'")
parser.add_argument("-t",
"--task-index",
type=int,
default=0,
help="Index of task within the job")
parser.add_argument(
"-s",
"--steps-per-explorer",
type=int,
default=1000000,
help=
"Max. steps an explorer should make in its env (in total before stopping)."
)
parser.add_argument(
"-m",
"--max-steps-per-episode",
type=int,
default=100,
help="Max. steps an explorer should make in one episode.")
parser.add_argument(
"-b",
"--buffer-size",
type=int,
default=1000,
help=
"Number of time steps to store (round robin) in the local buffers of each explorer. "
"The size of the global buffer will be this number times the number of explorers."
)
parser.add_argument(
"--learn-batch-size",
type=int,
default=16,
help=
"Size of a batch (number of episodes) to pull randomly from the main buffer "
"for each learner iteration.")
parser.add_argument(
"-f",
"--upload-frequency",
type=int,
default=4,
help=
"Every how many episodes does an explorer upload its local buffer of episodes "
"to the learners?")
parser.add_argument("--num-hidden",
type=int,
default=10,
help="Number of hidden nodes.")
parser.add_argument("-g",
"--gamma",
type=float,
default=0.97,
help="The discount factor gamma (default 0.9).")
parser.add_argument(
"-a",
"--learning-rate",
type=float,
default=0.0005,
help="The learning rate (alpha) to use for optimizing the cost.")
args = parser.parse_args()
learner_hosts = args.learner_hosts.split(",")
explorer_hosts = args.explorer_hosts.split(",")
# Create a cluster from the given hosts (learners and explorers).
cluster = tf.train.ClusterSpec({
"learner": learner_hosts,
"explorer": explorer_hosts
})
# Create and start a server for the local task.
server = tf.train.Server(cluster,
job_name=args.job_name,
task_index=args.task_index)
# simple 1 hidden layer, feed-forward network
num_inputs = 1 # input is always [1.0] -> blind env
num_hidden = args.num_hidden
num_out = 2 + 1 # left and right actions + value function
# 1 disc. return, 1 action (0=left, 1=right),
len_buffer_record = 2
#config = tf.ConfigProto(log_device_placement=True)
##config.gpu_options.allow_growth = True # need to set this in order to be able to run locally with GPU
#config.gpu_options.per_process_gpu_memory_fraction = 0.4
# the global main policy: pi (live on the learner(s))
# - explorers sync their own policies (mu) with this at the beginning of an episode
with tf.device(
tf.train.replica_device_setter(
ps_tasks=len(learner_hosts),
ps_device="/job:learner",
worker_device="/job:explorer/task:0")):
weights_1_pi = tf.Variable(tf.truncated_normal(shape=(num_inputs,
num_hidden),
stddev=0.01),
name="pi-W1")
biases_1_pi = tf.Variable(tf.zeros(shape=(num_hidden, )), name="pi-b1")
weights_2_pi = tf.Variable(tf.truncated_normal(shape=(num_hidden,
num_out),
stddev=0.01),
name="pi-W2")
biases_2_pi = tf.Variable(tf.zeros(shape=(num_out, )), name="pi-b2")
# main experience buffer (on central learner)
# - create experience buffer in both explorer and learner, but host it on learner
# - each explorer writes to a certain chunk in round robin fashion
size_main_experience_buffer = args.buffer_size * len(explorer_hosts)
with tf.device("/job:learner/task:0"):
main_experience_buffer = tf.Variable(tf.zeros(
[size_main_experience_buffer, len_buffer_record]),
name="global-episode-buffer")
global_step = tf.train.get_or_create_global_step()
if args.job_name == "explorer":
# local policy (mu) -> all zero; will be sync'd with pi anyway at start of each episode
weights_1_mu = tf.Variable(tf.zeros(shape=(num_inputs, num_hidden)),
name="mu-W1")
biases_1_mu = tf.Variable(tf.zeros(shape=(num_hidden, )), name="mu-b1")
weights_2_mu = tf.Variable(tf.zeros(shape=(num_hidden, num_out)),
name="mu-W2")
biases_2_mu = tf.Variable(tf.zeros(shape=(num_out, )), name="mu-b2")
# ops that sync from the main policy (pi) (using locking)
# - must fetch these after a reset of the env (before querying the first action in each episode)
sync_ops = [
tf.assign(weights_1_mu, weights_1_pi, name="sync-W1"),
tf.assign(biases_1_mu, biases_1_pi, name="sync-b1"),
tf.assign(weights_2_mu, weights_2_pi, name="sync-W2"),
tf.assign(biases_2_mu, biases_2_pi, name="sync-b2")
]
# number of actions (log-action-probs) to sample
num_action_samples = tf.placeholder(dtype=tf.int32, shape=())
# Buffer to store n (capacity) episodes of experiences (round-robin).
# This one gets inserted into the learners global memory after each m (upload-frequency) episodes.
# rank0=episode, rank1=step in episode, rank2=[action(0=left, 1=right), reward, mu(a)]
experience_buffer = tf.Variable(tf.zeros(
[args.buffer_size, len_buffer_record]),
name="episode-buffer")
experience_buffer_idx = tf.placeholder(dtype=tf.int32, shape=())
# In case we would like to use LSTM -> need to pass the initial internal state to learner
# so it can replay the episode through pi (instead of mu).
# init_internal_buffer = tf.Variable(tf.zeros([args.buffer_size, num_internal_state]),
# name="init-internal-buffer")
# upload op (from local experience buffer to global one)
start = args.task_index * args.buffer_size
stop = start + args.buffer_size
experience_upload = tf.assign(main_experience_buffer[start:stop, :],
experience_buffer)
# forward pass -> let local explorer handle this (as it's needed right here for querying actions)
hidden_out = tf.add(
tf.matmul(tf.ones(shape=(num_action_samples, num_inputs)),
weights_1_mu), biases_1_mu)
logits = tf.add(tf.matmul(hidden_out, weights_2_mu), biases_2_mu)
action_prob = tf.nn.softmax(
logits[:, :2]) # first two outputs are action logits
action_distr = tf.distributions.Bernoulli(probs=action_prob[:, 1])
actions = action_distr.sample(sample_shape=num_action_samples)
# store incoming step (a, R) in local experience buffer
a_in = tf.placeholder(dtype=tf.float32, shape=(None, ),
name="a-in") # 0.0=left, 1.0=right
returns_in = tf.placeholder(
dtype=tf.float32, shape=(None, ),
name="returns-in") # None=timesteps in the episode
# concat returns and log_aps within each timestep
episode = tf.concat(
[tf.expand_dims(a_in, 1),
tf.expand_dims(returns_in, 1)], 1)
episode_len = tf.shape(episode)[0]
stop = experience_buffer_idx + episode_len
# don't have to lock as the only one that's ever touching the local buffer is ourselves
add_episode = tf.cond(
stop <= args.buffer_size,
# true fn
lambda: tf.group(
tf.assign(experience_buffer[experience_buffer_idx:stop],
episode,
use_locking=False)),
# false fn
lambda: tf.group(
tf.assign(experience_buffer[experience_buffer_idx:],
episode[:args.buffer_size - experience_buffer_idx],
use_locking=False),
tf.assign(experience_buffer[:episode_len - (
args.buffer_size - experience_buffer_idx)],
episode[args.buffer_size - experience_buffer_idx:],
use_locking=False)))
# create our own private env
env = Env()
total_steps = 0
num_episodes = 0
with tf.train.MonitoredTrainingSession(
master=server.target,
is_chief=(args.task_index == 0),
#config=config,
hooks=[]) as mon_sess:
while total_steps < args.steps_per_explorer:
rs = [] # discounted accum. rewards over one episode
as_ = [] # the actual actions taken
# reset the env
env.reset()
episode_steps = 0
buffer_idx = 0
# update our mu with pi from learner
fetches = mon_sess.run(sync_ops)
while episode_steps < args.max_steps_per_episode and total_steps < args.steps_per_explorer:
s = env.state
a = mon_sess.run(actions,
feed_dict={num_action_samples: 1})
a = a[0][0] # a=0 for 'left', a=1 for 'right'
as_.append(a)
r, is_terminal = env.execute(a)
rs.append(r)
print(
"{:03d} explorer {}: s={} action {} (1=right) s'={} is-term={}"
.format(total_steps, args.task_index, s, a, env.state,
is_terminal))
total_steps += 1
episode_steps += 1
if is_terminal:
# calculate discounted accumulated rewards (returns)
returns = discount(rs, args.gamma)
# add episode to our buffer
mon_sess.run(
[add_episode],
feed_dict={
returns_in: returns,
a_in: as_,
experience_buffer_idx: buffer_idx
})
env.reset()
buffer_idx = (buffer_idx + len(rs)) % args.buffer_size
rs = []
as_ = []
episode_steps = 0
num_episodes += 1
fetches = sync_ops
if num_episodes % args.upload_frequency == 0:
print(
"num_episodes={}: uploading local buffer to global buffer"
.format(num_episodes))
fetches.append(experience_upload)
fetches = mon_sess.run(fetches)
print("Explorer {} is done!".format(args.task_index))
# - every learner iteration, it samples randomly from the main buffer and learns
else:
# build the pi-network (similar to mu-network above)
#with tf.device("/gpu:0"):
hidden_out = tf.add(
tf.matmul(tf.ones(shape=(args.learn_batch_size, num_inputs)),
weights_1_pi), biases_1_pi)
logits = tf.add(tf.matmul(hidden_out, weights_2_pi), biases_2_pi)
action_right_prob = tf.nn.softmax(
logits[:, :2])[:,
1] # first two outputs are action logits (1=right)
avg_action_right_prob = tf.reduce_mean(action_right_prob)
values = logits[:, 2:3] # last output is the state-value
# for now: do simple REINFORCE (add v-trace later or directly to tensorforce as it's not really different)
# get a random batch from the main buffer
indexes = tf.random_shuffle(tf.range(size_main_experience_buffer))
sample = tf.gather(main_experience_buffer,
indexes[:args.learn_batch_size])
# separate action (0=left, 1=right) and discounted accum. reward
actions, returns = tf.split(sample,
num_or_size_splits=len_buffer_record,
axis=1)
# probability of the action actually taken
action_prob = tf.abs(
tf.subtract(actions, tf.ones(tf.shape(actions))) +
action_right_prob)
log_action_prob = tf.log(action_prob)
advantage = (returns - values)
# define our loss function
alpha_1 = 0.5 # log-action-prob loss
alpha_2 = 0.4 # value function loss
alpha_3 = 0.1 # regularization
loss = - alpha_1 * tf.reduce_mean(tf.multiply(log_action_prob, advantage)) + \
alpha_2 * tf.reduce_mean(0.5 * tf.square(advantage)) + \
alpha_3 * 0 # TODO: add regularization
train_op = tf.train.AdagradOptimizer(args.learning_rate).minimize(
loss, global_step=global_step)
with tf.train.MonitoredTrainingSession(master=server.target,
is_chief=(args.task_index == 0),
hooks=[]) as mon_sess:
while not mon_sess.should_stop():
sample_out, _, loss_out, g_step, avg_right_prob_out = mon_sess.run(
[
sample, train_op, loss, global_step,
avg_action_right_prob
])
print("task {} step {} loss {} avg-right-prob={}".format(
args.task_index, g_step, loss_out, avg_right_prob_out))
print("Learner {} is done!".format(args.task_index))
def batch_inputs(self, dataset, train):
"""Contruct batches of training or evaluation examples from the image input_data.
Args:
dataset: instance of Dataset class specifying the input_data.
See input_data.py for details.
batch_size: integer
train: boolean
num_preprocess_threads: integer, total number of preprocessing threads
num_readers: integer, number of parallel readers
Returns:
images: 4-D float Tensor of a batch of images
labels: 1-D integer Tensor of [batch_size].
Raises:
ValueError: if data is not found
"""
with tf.name_scope('batch_processing'):
data_files = dataset.data_files()
if data_files is None:
raise ValueError('No data files found for this input_data')
# Create filename_queue
if train:
filename_queue = tf.train.string_input_producer(data_files,
shuffle=True,
capacity=16)
else:
filename_queue = tf.train.string_input_producer(data_files,
shuffle=False,
capacity=1)
# Approximate number of examples per shard.
examples_per_shard = 1024
# Size the random shuffle queue to balance between good global
# mixing (more examples) and memory use (fewer examples).
# 1 image uses 299*299*3*4 bytes = 1MB
# The default input_queue_memory_factor is 16 implying a shuffling queue
# size: examples_per_shard * 16 * 1MB = 17.6GB
min_queue_examples = examples_per_shard * self.input_queue_memory_factor
if train:
examples_queue = tf.RandomShuffleQueue(
capacity=min_queue_examples + 3 * self.batch_size,
min_after_dequeue=min_queue_examples,
dtypes=[tf.string])
else:
examples_queue = tf.FIFOQueue(capacity=examples_per_shard +
3 * self.batch_size,
dtypes=[tf.string])
# Create multiple readers to populate the queue of examples.
if self.num_readers > 1:
enqueue_ops = []
for _ in range(self.num_readers):
reader = dataset.reader()
_, value = reader.read(filename_queue)
enqueue_ops.append(examples_queue.enqueue([value]))
tf.train.queue_runner.add_queue_runner(
tf.train.queue_runner.QueueRunner(examples_queue,
enqueue_ops))
example_serialized = examples_queue.dequeue()
else:
reader = dataset.reader()
_, example_serialized = reader.read(filename_queue)
pos_queue = None
neg_queue = None
if self.batch_size < 2:
pos_queue = tf.RandomShuffleQueue(
name="pos-queue",
capacity=10,
min_after_dequeue=5,
dtypes=[tf.float32, tf.float32, tf.string])
neg_queue = tf.RandomShuffleQueue(
name="neg-queue",
capacity=10,
min_after_dequeue=5,
dtypes=[tf.float32, tf.float32, tf.string])
pos_queue_enq = []
neg_queue_enq = []
with tf.name_scope('split-merge'):
if train and self.ensure_posneg_balance:
images_and_masks = []
for thread_id in range(self.num_preprocess_threads):
# Parse a serialized Example proto to extract the image and metadata.
image_buffer, mask_buffer, img_name_ = self.parse_example_proto(
example_serialized)
image_ = self.image_preprocessing(
image_buffer,
img_size=(self.input_size[0], self.input_size[1]),
num_channels=self.input_size[2])
mask_ = self.image_preprocessing(
mask_buffer,
img_size=(self.mask_size[0], self.mask_size[1]),
num_channels=self.mask_size[2])
image_ = tf.expand_dims(image_, 0)
mask_ = tf.expand_dims(mask_, 0)
img_name_ = tf.expand_dims(img_name_, 0)
img_shape = tf.TensorShape([
image_.shape[1], image_.shape[2], image_.shape[3]
])
mask_shape = tf.TensorShape(
[mask_.shape[1], mask_.shape[2], mask_.shape[3]])
img_name_shape = tf.TensorShape([])
# initialize pos/neg queues with proper shape size on first
if pos_queue is None or neg_queue is None:
pos_queue = tf.RandomShuffleQueue(
name="pos-queue",
capacity=10,
min_after_dequeue=5,
dtypes=[tf.float32, tf.float32, tf.string],
shapes=[img_shape, mask_shape, img_name_shape])
neg_queue = tf.RandomShuffleQueue(
name="neg-queue",
capacity=10,
min_after_dequeue=5,
dtypes=[tf.float32, tf.float32, tf.string],
shapes=[img_shape, mask_shape, img_name_shape])
is_pos = tf.squeeze(
tf.reduce_sum(mask_, [1, 2], keep_dims=False))
neg_mask = tf.less_equal(is_pos, 0)
pos_idx = tf.reshape(
tf.where([tf.logical_not(neg_mask)]), [-1])
neg_idx = tf.reshape(tf.where([neg_mask]), [-1])
pos_data = [
tf.gather(image_, pos_idx),
tf.gather(mask_, pos_idx),
tf.gather(img_name_, pos_idx)
]
neg_data = [
tf.gather(image_, neg_idx),
tf.gather(mask_, neg_idx),
tf.gather(img_name_, neg_idx)
]
pos_queue_enq.append(pos_queue.enqueue_many(pos_data))
neg_queue_enq.append(neg_queue.enqueue_many(neg_data))
tf.train.queue_runner.add_queue_runner(
tf.train.queue_runner.QueueRunner(
pos_queue, pos_queue_enq))
tf.train.queue_runner.add_queue_runner(
tf.train.queue_runner.QueueRunner(
neg_queue, neg_queue_enq))
if self.batch_size >= 2:
if self.batch_size % 2 != 0:
raise Exception(
"'batch_size' mod 2 != 0 ! only even batch sizes supported at the moment"
)
num_deque = int(self.batch_size / 2)
pos_data = pos_queue.dequeue_many(num_deque)
neg_data = neg_queue.dequeue_many(num_deque)
concat_data = [
tf.concat([pos_data[0], neg_data[0]],
axis=0,
name='Concat-img'),
tf.concat([pos_data[1], neg_data[1]],
axis=0,
name='Concat-mask'),
tf.concat([pos_data[2], neg_data[2]],
axis=0,
name='Concat-img-name')
]
# randomly permute within batch size (is this even necessary ??)
idx = tf.Variable(range(0, self.batch_size),
trainable=False,
dtype=tf.int32)
idx = tf.random_shuffle(idx)
images = tf.gather(concat_data[0], idx)
masks = tf.gather(concat_data[1], idx)
img_names = tf.gather(concat_data[2], idx)
else:
# positive only
#images, masks, img_names = pos_queue.dequeue()
# negative only
#images, masks, img_names = neg_queue.dequeue()
# mix 50/50
counter = tf.Variable(initial_value=0,
trainable=False,
dtype=tf.int32)
counter = tf.assign_add(counter, 1)
condition_term = tf.equal(tf.mod(counter, 2),
tf.constant(0))
images, masks, img_names = tf.cond(
condition_term, lambda: pos_queue.dequeue(),
lambda: neg_queue.dequeue())
if self.use_random_rotation:
images.set_shape(
tensor_shape.as_shape([None, None, 1]))
masks.set_shape(
tensor_shape.as_shape([None, None, 1]))
# randomly rotate image by 90 degrees
rot_factor = tf.random_uniform([1],
minval=0,
maxval=3,
dtype=tf.int32)
rot_factor = tf.gather(rot_factor, 0)
images = tf.image.rot90(images, k=rot_factor)
masks = tf.image.rot90(masks, k=rot_factor)
images = tf.expand_dims(images, axis=0)
masks = tf.expand_dims(masks, axis=0)
img_names = tf.expand_dims(img_names, axis=0)
else:
# Parse a serialized Example proto to extract the image and metadata.
image_buffer, mask_buffer, img_names = self.parse_example_proto(
example_serialized)
images = self.image_preprocessing(
image_buffer,
img_size=(self.input_size[0], self.input_size[1]),
num_channels=self.input_size[2])
masks = self.image_preprocessing(
mask_buffer,
img_size=(self.mask_size[0], self.mask_size[1]),
num_channels=1)
images = tf.expand_dims(images, axis=0)
masks = tf.expand_dims(masks, axis=0)
img_names = tf.expand_dims(img_names, axis=0)
# Reshape images into these desired dimensions.
images = tf.cast(images, tf.float32)
masks = tf.cast(masks, tf.float32)
images.set_shape(
tensor_shape.as_shape(
[self.batch_size, None, None, self.input_size[2]]))
masks.set_shape(
tensor_shape.as_shape([
self.batch_size, self.input_size[0], self.input_size[1],
self.mask_size[2]
]))
# Display the training images in the visualizer.
tf.summary.image('images', images)
tf.summary.image('masks', masks)
return images, masks, img_names
for i in range(num_vectors):
if np.random.random() > 0.5:
x_values.append(np.random.normal(0.4, 0.7))
y_values.append(np.random.normal(0.2, 0.8))
else:
x_values.append(np.random.normal(0.6, 0.4))
y_values.append(np.random.normal(0.8, 0.5))
# 將Ｘ陣列與Ｙ陣列合併為一項量陣列-> [[x0,y0],[x1,y1]..]
vector_values = list(zip(x_values, y_values))
vectors = tf.constant(vector_values)
# 取得陣列維度大小
n_samples = tf.shape(vector_values)[0]
# 製作一n_samples數量的編號陣列，範圍為0~n_samples-1
sample_range = tf.range(0, n_samples)
# 將編號陣列進行洗牌
random_indices = tf.random_shuffle(sample_range)
begin = [
0,
]
size = [
num_clusters,
]
size[0] = num_clusters
# 從洗牌後的編號陣列中取出num_clusters個值，型態為[a1,a2,a3,...]
centroid_indices = tf.slice(random_indices, begin, size)
# 再利用上述值從原陣列中取得num_clusters個中心點，型態為 [[xa,ya],[xb,yb],...]
centroids = tf.Variable(tf.gather(vector_values, centroid_indices))
# 增加維度
def model(params, examples, labels, epsilon):
serp_len = params['serp_len']
doc_emb_size = params['doc_emb'][-1]
hidden_state_size = params['hidden_state_size']
docs = examples['doc_tensors']
batch_size = docs.shape[0].value
batch_max_docs = tf.shape(docs)[1]
docs_per_query = examples['n_docs']
if params['context_input']:
to_shuffle = tf.concat([tf.cast(labels[:, :, None], tf.float32), docs],
axis=2)
shuffled = tf.random_shuffle(tf.transpose(to_shuffle, [1, 0, 2]))
shuffled = tf.transpose(shuffled, [1, 0, 2])
labels = tf.cast(tf.slice(shuffled, [0, 0, 0], [-1, -1, 1]), tf.int32)
labels = labels[:, :, 0]
docs = tf.slice(shuffled, [0, 0, 1], [-1, -1, -1])
# assert not params['context_input'], 'Context not supported for GRU.'
result = {
'docs_per_query': docs_per_query,
}
doc_emb = mu._shared_doc_embeddings(docs,
params,
'/main/doc_emb',
inference=True)
hidden_init = tf.zeros([batch_size, hidden_state_size])
if params['context_input']:
context_gru_fn = ru.get_gru_layer(params,
'/main/gru/context',
label_network=False,
inference=True,
reuse_variable_scope=False)
scan_input = tf.transpose(doc_emb, [1, 0, 2])
context = tf.scan(context_gru_fn, scan_input, hidden_init)
ind_nd = tf.concat([docs_per_query - 1,
tf.range(batch_size)[:, None]],
axis=1)
hidden_init = tf.gather_nd(context, ind_nd)
gru_fn = ru.get_gru_layer(params,
'/main/gru',
label_network=False,
inference=True,
reuse_variable_scope=False)
policy = mu.EpsilonGreedy(epsilon, batch_size, batch_max_docs,
docs_per_query)
hidden_state = hidden_init
#tf.zeros([n_docs, hidden_state_size])
serp = []
serp_labels = []
serp_ind = []
for i in range(serp_len):
hidden_states = tf.tile(hidden_state[:, None, :],
[1, batch_max_docs, 1])
score_input = tf.concat([hidden_states, doc_emb], axis=2)
scores = mu._create_subnetwork(score_input,
params,
subnetwork_name='/main/scoring',
label_network=False,
reuse_variable_scope=i > 0,
inference=True)
action = policy.choose(scores)
serp_ind.append(action)
nd_ind = tf.stack([tf.range(batch_size, dtype=tf.int64), action],
axis=1)
select_doc = tf.gather_nd(docs, nd_ind)
select_labels = tf.gather_nd(labels, nd_ind)[:, None]
tf.summary.scalar('policy/scores/pos_%d' % i,
tf.reduce_mean(tf.gather_nd(scores, nd_ind)))
serp_labels.append(
tf.where(
tf.less(i, docs_per_query),
select_labels,
tf.zeros([batch_size, 1], dtype=tf.int32),
))
serp.append(select_doc)
if i < serp_len - 1:
select_emb = tf.gather_nd(doc_emb, nd_ind)
hidden_state = gru_fn(hidden_state, select_emb)
result['serp'] = tf.stack(serp, axis=1)
result['serp_ind'] = tf.stack(serp_ind, axis=1)
result['labels'] = tf.concat(serp_labels, axis=1)
tf.summary.histogram("label/output", result['labels'])
# if params['context_input']:
max_docs = params['max_docs']
padding = tf.convert_to_tensor([[0, 0], [0, max_docs - batch_max_docs],
[0, 0]])
padded_docs = tf.pad(docs, padding, "CONSTANT")
padded_docs = tf.reshape(padded_docs,
[batch_size, max_docs, docs.shape[2].value])
result['docs'] = padded_docs
return result
def generate_curves(self):
'''
生成数据
生成函数输出为浮点型，输入为-2到2之间
返回：NPRegressionDescription 命名元组
'''
# 从均匀分布中随机采样
num_context = tf.random_uniform(shape=[], minval=3, maxval=self._max_num_context, dtype=tf.int32)
# 测试过程中，生成更多的点以绘制图像
if self._testing:
# 此处需要配合 x_size 进行变换
num_target = tf.cast(400 / self._x_size, dtype=tf.int32)
num_total_points = num_target
# tf.tile: 在同一维度上复制
# 生成 -2. 到 2. 的步长为 0.01 的数据
x_values = tf.tile(tf.expand_dims(tf.range(-2., 2., 1. / 100, dtype=tf.float32), axis=0), [self._batch_size, 1])
x_values = tf.expand_dims(x_values, axis=-1)
# reshape 成 x_size 的形状
x_values = tf.reshape(x_values, (self._batch_size, num_total_points, self._x_size))
else:
# 训练过程中，随机选择目标点个数与他们的 x 坐标
num_target = tf.random_uniform(shape=(), minval=0, maxval=self._max_num_context - num_context, dtype=tf.int32)
num_total_points = num_context + num_target
x_values = tf.random_uniform([self._batch_size, num_total_points, self._x_size], minval=-2, maxval=2)
# 设置核函数参数
if self._random_kernel_parameters:
# 随机选择参数
l1 = tf.random_uniform([self._batch_size, self._y_size, self._x_size], 0.1, self._l1_scale)
sigma_f = tf.random_uniform([self._batch_size, self._y_size], 0.1, self._sigma_scale)
else:
# 使用固定参数
l1 = tf.ones([self._batch_size, self._y_size, self._x_size]) * self._l1_scale
sigma_f = tf.ones([self._batch_size, self._y_size]) * self._sigma_scale
# [B, y_size, num_total_points, num_total_points]
kernel = self._gaussian_kernels(x_values, l1, sigma_f)
# Cholesky 分解
cholesky = tf.cast(tf.cholesky(tf.cast(kernel, tf.float64)), tf.float32)
# 采样
# [B, y_size, num_total_points, 1]
y_values = tf.matmul(cholesky, tf.random_uniform([self._batch_size, self._y_size, num_total_points, 1]))
# 矩阵转置，按照 perm 排列原始维度；tf.squeeze 删除所有大小为 1 的维度
# [B, num_total_points, y_size]
y_values = tf.transpose(tf.squeeze(y_values, 3), perm=[0, 2, 1])
if self._testing:
target_x = x_values
target_y = y_values
idx = tf.random_shuffle(tf.range(num_target))
# tf.gather: 根据索引从 params 中读取数据
context_x = tf.gather(params=x_values, indices=idx[:num_context], axis=1)
context_y = tf.gather(params=y_values, indices=idx[:num_context], axis=1)
else:
target_x = x_values[:, :num_target + num_context, :]
target_y = y_values[:, :num_target + num_context, :]
context_x = x_values[:, :num_context, :]
context_y = y_values[:, :num_context, :]
query = ((context_x, context_y), target_x)
return NPRegressionDescription(
query=query,
target_y=target_y,
num_total_points=tf.shape(target_x)[1],
num_context_points=num_context
)
def _subsampling(self,
normalized_rois,
gt_bboxes,
gt_labels,
pos_iou_thresh=0.5,
exclusive_iou_tresh=0.1,
pos_ratio=0.25):
"""正解データとのIoUを基にRoIをサンプリングする。
IoUがpos_iou_thresh以上であるRoIをオブジェクトとみなす。
オブジェクトはサンプルの25%以内とする。（n_samples_per_batch * pos_ratio 以内）
pos_iou_thresh未満、exclusive_iou_thresh以上は非オブジェクトとみなす。
exclusive_iou_thresh未満は偶然の一致であり意味なし（難解）なので無視。
※論文ではheuristic for hard example mining.と記載されている点。
バッチ毎のサンプル数はn_samples_per_batch以内とする。
（n_samples_per_batch未満の場合は、n_samples_per_batchになるよう0パディングする。）
上記のサンプリングに対応する正解データのラベル、また、BBoxとのオフセットも得る。
Args:
normalized_rois (tensor) : RegionProposalLayerで得られたRoI。
(N, n_rois, 4)
3軸目は領域の左上と右下の座標が0〜1に正規化された値。
入力画像サイズの高さ、幅で除算することで正規化された値。
(y1, x1, y2, x2)
gt_bboxes (ndarray) : 正解BBox。
(N, config.n_max_gt_objects_per_image, 4)
座標は正規化されていない。
gt_labels (ndarray) : 正解ラベル。
(N, config.n_max_gt_objects_per_image)
==0:背景データ
>=1:オブジェクト
Returns:
sample_rois (tensor): サンプリングしたRoI。
(N, n_samples_per_batch, 4)
3軸目の座標は0〜1に正規化された値。
sample_gt_offset (tensor): サンプリングしたRoIに対応するBBoxとのオフセット。
(N, n_samples_per_batch, 4)
3軸目の座標は0〜1に正規化された値をself.config.bbox_refinement_stdで割ることで標準化した値。
sample_gt_labels (tensor): サンプリングしたRoIに対応するBBoxのラベル。
(N, n_samples_per_batch)
"""
pos_roi_per_batch = round(self.n_samples_per_batch * pos_ratio)
# gt_bboxesをnormalized_roisに合わせて正規化する。
# これでIoUが評価出来るようになる。
input_h = self.config.image_shape[0]
input_w = self.config.image_shape[1]
normalized_gt_bboxes = bbox.normalize_bbox(gt_bboxes, input_h, input_w)
# 入力をバッチ毎に分割
normalized_rois = tf.split(normalized_rois, self.config.batch_size)
normalized_gt_bboxes = tf.split(normalized_gt_bboxes,
self.config.batch_size)
gt_labels = tf.split(gt_labels, self.config.batch_size)
sample_rois = []
sample_gt_offsets = []
sample_gt_labels = []
for roi, gt_bbox, gt_label in zip(normalized_rois,
normalized_gt_bboxes, gt_labels):
# 0次元目(バッチサイズ)は不要なので削除
roi = log.tfprint(roi, "roi: ")
gt_bbox = log.tfprint(gt_bbox, "gt_bbox: ")
gt_label = log.tfprint(gt_label, "gt_label: ")
roi = K.squeeze(roi, 0)
gt_bbox = K.squeeze(gt_bbox, 0)
gt_label = K.squeeze(gt_label, 0)
roi = log.tfprint(roi, "roi_squeezed: ")
gt_bbox = log.tfprint(gt_bbox, "gt_bbox_squeezed: ")
gt_label = log.tfprint(gt_label, "gt_label_squeezed: ")
# ゼロパディング行を除外
# K.gather(zero, K.squeeze(tf.where(K.any(zero, axis=1)), -1) )
idx_roi_row = K.flatten(tf.where(K.any(roi, axis=1)))
idx_gt_bbox = K.flatten(tf.where(K.any(gt_bbox, axis=1)))
roi = K.gather(roi, idx_roi_row)
# gt_bboxとgt_labelは行数と行の並びが同じなので同じidxを利用できる
gt_bbox = K.gather(gt_bbox, idx_gt_bbox)
gt_label = K.gather(gt_label, idx_gt_bbox)
gt_bbox = log.tfprint(gt_bbox, "gt_bbox_gathered: ")
gt_label = log.tfprint(gt_label, "gt_label_gathered: ")
# IoUを求める。
# (n_rois, )
ious = bbox.get_iou_K(roi, gt_bbox)
ious = log.tfprint(ious, "ious: ")
# 各RoI毎にIoU最大のBBoxの位置を得る
idx_max_gt = K.argmax(ious, axis=1)
idx_max_gt = log.tfprint(idx_max_gt, "idx_max_gt: ")
max_iou = K.max(ious, axis=1) # max_iouの行数はroiと同じになる
max_iou = log.tfprint(max_iou, "max_iou: ")
idx_pos = K.flatten(tf.where(max_iou >= pos_iou_thresh))
# positiveサンプル数をpos_roi_per_batch以内に制限
limit_pos = K.minimum(pos_roi_per_batch, K.shape(idx_pos)[0])
idx_pos = K.switch(
K.shape(idx_pos)[0] > 0,
tf.random_shuffle(idx_pos)[:limit_pos], idx_pos)
limit_pos = log.tfprint(limit_pos, "limit_pos: ")
idx_pos = log.tfprint(idx_pos, "idx_pos: ")
# negativeサンプル数を
# n_samples_per_batch - pos_roi_per_batch
# に制限
idx_neg = K.flatten(
tf.where((max_iou < pos_iou_thresh)
& (max_iou >= exclusive_iou_tresh)))
# negativeサンプル数は pos_roi_per_batch - limit_pos(つまり残り) 以内に制限
limit_neg = self.n_samples_per_batch - limit_pos
limit_neg = K.minimum(limit_neg, K.shape(idx_neg)[0])
idx_neg = K.switch(
K.shape(idx_neg)[0] > 0,
tf.random_shuffle(idx_neg)[:limit_neg], idx_neg)
limit_neg = log.tfprint(limit_neg, "limit_neg: ")
idx_neg = log.tfprint(idx_neg, "idx_neg: ")
# 返却するサンプルを抽出
# GTのoffsets, labelsは各roisに対応させる。つまり、同じ位置に格納する。
idx_keep = K.concatenate((idx_pos, idx_neg))
idx_keep = log.tfprint(idx_keep, "idx_keep: ")
# 各RoIの最大IoUを示すIndexについても、上記返却するサンプルのみを残す。
idx_gt_keep = K.gather(idx_max_gt, idx_keep)
# IoUが閾値以上のPositiveとみなされるサンプルのみを残すためのIndex。
idx_gt_keep_pos = K.gather(idx_max_gt, idx_pos)
idx_gt_keep = log.tfprint(idx_gt_keep, "idx_gt_keep: ")
sample_roi = K.gather(roi, idx_keep)
sample_gt_offset = bbox.get_offset_K(
sample_roi, K.gather(gt_bbox, idx_gt_keep))
# negativeな要素には0を設定
sample_gt_label = K.concatenate((
K.cast(K.gather(gt_label, idx_gt_keep_pos), dtype='int32'),
K.zeros(
[limit_neg], # K.zerosは0階テンソルを受け付けないので配列化。。。
dtype='int32')))
# 行数がn_samples_per_batch未満の場合は0パディング
remain = tf.maximum(
self.n_samples_per_batch - tf.shape(sample_roi)[0], 0)
sample_roi = tf.pad(sample_roi, [(0, remain), (0, 0)],
name='subsample_sample_roi')
sample_gt_offset = tf.pad(sample_gt_offset, [(0, remain), (0, 0)],
name='subsample_sample_gt_offset')
sample_gt_offset /= self.config.bbox_refinement_std
sample_gt_label = tf.pad(sample_gt_label, [(0, remain)],
name='subsample_sample_gt_label')
sample_roi = log.tfprint(sample_roi, "sample_roi: ")
sample_gt_offset = log.tfprint(sample_gt_offset,
"sample_gt_offset: ")
sample_gt_label = log.tfprint(sample_gt_label, "sample_gt_label: ")
sample_rois.append(sample_roi)
sample_gt_offsets.append(sample_gt_offset)
sample_gt_labels.append(sample_gt_label)
return [
K.stack(sample_rois),
K.stack(sample_gt_offsets),
K.stack(sample_gt_labels)
]
def main(self):
with tf.Graph().as_default() as graph, tf.device('/cpu:0'):
num_gpu = len(cfgs.GPU_GROUP.strip().split(','))
global_step = slim.get_or_create_global_step()
lr = self.warmup_lr(cfgs.LR, global_step, cfgs.WARM_SETP,
num_gpu * cfgs.BATCH_SIZE)
tf.summary.scalar('lr', lr)
optimizer = tf.train.MomentumOptimizer(lr, momentum=cfgs.MOMENTUM)
retinanet = build_whole_network_batch.DetectionNetworkRetinaNet(
cfgs=self.cfgs, is_training=True)
with tf.name_scope('get_batch'):
if cfgs.IMAGE_PYRAMID:
shortside_len_list = tf.constant(cfgs.IMG_SHORT_SIDE_LEN)
shortside_len = tf.random_shuffle(shortside_len_list)[0]
else:
shortside_len = cfgs.IMG_SHORT_SIDE_LEN
img_name_batch, img_batch, gtboxes_and_label_batch, num_objects_batch, img_h_batch, img_w_batch = \
self.reader.next_batch(dataset_name=cfgs.DATASET_NAME,
batch_size=cfgs.BATCH_SIZE * num_gpu,
shortside_len=shortside_len,
is_training=True)
# data processing
inputs_list = []
for i in range(num_gpu):
start = i * cfgs.BATCH_SIZE
end = (i + 1) * cfgs.BATCH_SIZE
img = img_batch[start:end, :, :, :]
pretrain_zoo = PretrainModelZoo()
if self.cfgs.NET_NAME in pretrain_zoo.pth_zoo or self.cfgs.NET_NAME in pretrain_zoo.mxnet_zoo:
img = img / tf.constant([cfgs.PIXEL_STD])
gtboxes_and_label_r = tf.py_func(
backward_convert,
inp=[
tf.reshape(gtboxes_and_label_batch[start:end], [-1, 9])
],
Tout=tf.float32)
gtboxes_and_label_r = tf.reshape(gtboxes_and_label_r,
[cfgs.BATCH_SIZE, -1, 6])
gtboxes_and_label_h = get_horizen_minAreaRectangle(
tf.reshape(gtboxes_and_label_batch[start:end], [-1, 9]))
gtboxes_and_label_h = tf.reshape(gtboxes_and_label_h,
[cfgs.BATCH_SIZE, -1, 5])
num_objects = num_objects_batch[start:end]
num_objects = tf.cast(
tf.reshape(num_objects, [
cfgs.BATCH_SIZE,
-1,
]), tf.float32)
img_h = img_h_batch[start:end]
img_w = img_w_batch[start:end]
inputs_list.append([
img, gtboxes_and_label_h, gtboxes_and_label_r, num_objects,
img_h, img_w
])
tower_grads = []
biases_regularizer = tf.no_regularizer
weights_regularizer = tf.contrib.layers.l2_regularizer(
cfgs.WEIGHT_DECAY)
with tf.variable_scope(tf.get_variable_scope()):
for i in range(num_gpu):
with tf.device('/gpu:%d' % i):
with tf.name_scope('tower_%d' % i):
with slim.arg_scope(
[slim.model_variable, slim.variable],
device='/device:CPU:0'):
with slim.arg_scope(
[
slim.conv2d, slim.conv2d_in_plane,
slim.conv2d_transpose,
slim.separable_conv2d,
slim.fully_connected
],
weights_regularizer=weights_regularizer,
biases_regularizer=biases_regularizer,
biases_initializer=tf.
constant_initializer(0.0)):
gtboxes_and_label_h, gtboxes_and_label_r = tf.py_func(
self.get_gtboxes_and_label,
inp=[
inputs_list[i][1],
inputs_list[i][2],
inputs_list[i][3]
],
Tout=[tf.float32, tf.float32])
gtboxes_and_label_h = tf.reshape(
gtboxes_and_label_h,
[cfgs.BATCH_SIZE, -1, 5])
gtboxes_and_label_r = tf.reshape(
gtboxes_and_label_r,
[cfgs.BATCH_SIZE, -1, 6])
img = inputs_list[i][0]
img_shape = inputs_list[i][-2:]
h_crop = tf.reduce_max(img_shape[0])
w_crop = tf.reduce_max(img_shape[1])
img = tf.image.crop_to_bounding_box(
image=img,
offset_height=0,
offset_width=0,
target_height=tf.cast(
h_crop, tf.int32),
target_width=tf.cast(w_crop, tf.int32))
outputs = retinanet.build_whole_detection_network(
input_img_batch=img,
gtboxes_batch_h=gtboxes_and_label_h,
gtboxes_batch_r=gtboxes_and_label_r,
gpu_id=i)
gtboxes_in_img_h = self.drawer.draw_boxes_with_categories(
img_batch=tf.expand_dims(
img[0, :, :, :], axis=0),
boxes=gtboxes_and_label_h[0, :, :-1],
labels=gtboxes_and_label_h[0, :, -1],
method=0)
gtboxes_in_img_r = self.drawer.draw_boxes_with_categories(
img_batch=tf.expand_dims(
img[0, :, :, :], axis=0),
boxes=gtboxes_and_label_r[0, :, :-1],
labels=gtboxes_and_label_r[0, :, -1],
method=1)
tf.summary.image(
'Compare/gtboxes_h_gpu:%d' % i,
gtboxes_in_img_h)
tf.summary.image(
'Compare/gtboxes_r_gpu:%d' % i,
gtboxes_in_img_r)
if cfgs.ADD_BOX_IN_TENSORBOARD:
detections_in_img = self.drawer.draw_boxes_with_categories_and_scores(
img_batch=tf.expand_dims(
img[0, :, :, :], axis=0),
boxes=outputs[0],
scores=outputs[1],
labels=outputs[2],
method=1)
tf.summary.image(
'Compare/final_detection_gpu:%d' %
i, detections_in_img)
loss_dict = outputs[-1]
total_loss_dict, total_losses = self.loss_dict(
loss_dict, num_gpu)
if i == num_gpu - 1:
regularization_losses = tf.get_collection(
tf.GraphKeys.REGULARIZATION_LOSSES)
# weight_decay_loss = tf.add_n(slim.losses.get_regularization_losses())
total_losses = total_losses + tf.add_n(
regularization_losses)
tf.get_variable_scope().reuse_variables()
grads = optimizer.compute_gradients(total_losses)
if cfgs.GRADIENT_CLIPPING_BY_NORM is not None:
grads = slim.learning.clip_gradient_norms(
grads, cfgs.GRADIENT_CLIPPING_BY_NORM)
tower_grads.append(grads)
self.log_printer(retinanet, optimizer, global_step, tower_grads,
total_loss_dict, num_gpu * cfgs.BATCH_SIZE, graph)
def detect_targets_graph(gt_boxes, gt_class_ids, proposals,
train_rois_per_image, roi_positive_ratio):
"""
每个图像生成检测网络的分类和回归目标
IoU>=0.5的为正样本；IoU<0.5的为负样本
:param gt_boxes: GT 边框坐标 [MAX_GT_BOXs, (y1,x1,y2,x2,tag)] ,tag=0 为padding
:param gt_class_ids: GT 类别 [MAX_GT_BOXs, 1+1] ;最后一位为tag, tag=0 为padding
:param proposals: [N,(y1,x1,y2,x2,tag)] ,tag=0 为padding
:param train_rois_per_image: 每张图像训练的proposal数量
:param roi_positive_ratio: proposal正负样本比
:return:
"""
# 去除padding
gt_boxes = tf_utils.remove_pad(gt_boxes)
gt_class_ids = tf_utils.remove_pad(gt_class_ids)[:, 0] # 从[N,1]变为[N]
proposals = tf_utils.remove_pad(proposals)
proposals_num = tf.shape(proposals)[0]
# 计算iou
iou = compute_iou(gt_boxes, proposals) # [gt_num,rois_num]
# iou >=0.5为正
proposals_iou_max = tf.reduce_max(iou, axis=0) # [rois_num]
positive_indices = tf.where(
tf.logical_and(tf.equal(iou, proposals_iou_max),
tf.greater_equal(iou, 0.5)))
gt_pos_idx = positive_indices[:, 0] # 第一维gt索引
proposal_pos_idx = positive_indices[:, 1] # 第二位rois索引
match_gt_num = tf.shape(tf.unique(gt_pos_idx)[0])[0] # shuffle 前匹配的gt num
gt_boxes_pos = tf.gather(gt_boxes, gt_pos_idx)
class_ids = tf.gather(gt_class_ids, gt_pos_idx)
proposal_pos = tf.gather(proposals, proposal_pos_idx)
# 根据正负样本比确定最终的正样本
positive_num = tf.minimum(
tf.shape(proposal_pos)[0],
int(train_rois_per_image * roi_positive_ratio))
gt_boxes_pos, class_ids, proposal_pos, gt_pos_idx = shuffle_sample(
[gt_boxes_pos, class_ids, proposal_pos, gt_pos_idx],
tf.shape(proposal_pos)[0], positive_num)
match_gt_num_after_shuffle = tf.shape(
tf.unique(gt_pos_idx)[0])[0] # shuffle 后匹配的gt num
# 计算回归目标
deltas = regress_target(proposal_pos, gt_boxes_pos)
# 负样本：与所有GT的iou<0.5且iou>0.1
proposal_iou_max = tf.reduce_max(iou, axis=0)
proposal_neg_idx = tf.cond( # 需要考虑GT个数为0的情况;全部都是负样本
tf.greater(tf.shape(gt_boxes)[0], 0),
true_fn=lambda: tf.where(
tf.logical_and(proposal_iou_max < 0.5, proposal_iou_max > 0.1))[:,
0],
false_fn=lambda: tf.cast(tf.range(proposals_num), dtype=tf.int64))
# 确定负样本数量
negative_num = tf.minimum(train_rois_per_image - positive_num,
tf.shape(proposal_neg_idx)[0])
proposal_neg_idx = tf.random_shuffle(proposal_neg_idx)[:negative_num]
# 收集负样本
proposal_neg = tf.gather(proposals, proposal_neg_idx)
class_ids_neg = tf.zeros(shape=[negative_num]) # 背景类，类别id为0
deltas_neg = tf.zeros(shape=[negative_num, 4])
# 合并正负样本
train_rois = tf.concat([proposal_pos, proposal_neg], axis=0)
deltas = tf.concat([deltas, deltas_neg], axis=0)
class_ids = tf.concat([class_ids, class_ids_neg], axis=0)
# 计算padding
class_ids, train_rois = tf_utils.pad_list_to_fixed_size(
[tf.expand_dims(class_ids, axis=1), train_rois],
train_rois_per_image) # class_ids分类扩一维
# 为后续处理方便负样本tag设置为-1
deltas = tf_utils.pad_to_fixed_size_with_negative(
deltas, train_rois_per_image, negative_num=negative_num)
# 其它统计指标
gt_num = tf.shape(gt_class_ids)[0] # GT数
miss_gt_num = gt_num - match_gt_num
miss_gt_num_shuffle = gt_num - match_gt_num_after_shuffle # shuffle后未分配roi的GT
gt_min_max_iou = tf.reduce_min(tf.reduce_max(iou, axis=1)) # gt 匹配最小最大值
return [
deltas, class_ids, train_rois,
tf_utils.scalar_to_1d_tensor(miss_gt_num),
tf_utils.scalar_to_1d_tensor(miss_gt_num_shuffle),
tf_utils.scalar_to_1d_tensor(gt_min_max_iou),
tf_utils.scalar_to_1d_tensor(positive_num),
tf_utils.scalar_to_1d_tensor(negative_num),
tf_utils.scalar_to_1d_tensor(proposals_num)
]
def f1():
# Choose a random window-length range from the sequence.
range_min = tf.random_shuffle(tf.range(seq_len - window))[0]
range_max = range_min + window
return tf.range(range_min, range_max)
