def split_raw_data(images,
labels,
test_ratio=0,
validation_ratio=0,
moderation_features=None,
augmentation_images=None,
augmentation_labels=None,
split_labels=True,
force_mask_creation=False):
"""Currently depends on test/validation_ratio being 0 when not using test/validation"""
# serialize labels if they are lists (e.g. for regression)
if isinstance(labels, list):
if split_labels:
labels = [' '.join(map(str, label)) for label in labels]
n_aug = len(
augmentation_labels
) if augmentation_images is not None and augmentation_labels is not None else 0
mask = _get_split_mask(test_ratio, validation_ratio, len(labels), n_aug,
force_mask_creation)
if augmentation_images is not None and augmentation_labels is not None:
images = images + augmentation_images
labels = labels + augmentation_labels
try:
if test_ratio != 0 and validation_ratio != 0:
train_images, test_images, val_images = tf.dynamic_partition(
images, mask, 3)
train_labels, test_labels, val_labels = tf.dynamic_partition(
labels, mask, 3)
elif test_ratio != 0 and validation_ratio == 0:
train_images, test_images = tf.dynamic_partition(images, mask, 2)
train_labels, test_labels = tf.dynamic_partition(labels, mask, 2)
val_images, val_labels = None, None
elif test_ratio == 0 and validation_ratio != 0:
train_images, val_images = tf.dynamic_partition(images, mask, 2)
train_labels, val_labels = tf.dynamic_partition(labels, mask, 2)
test_images, test_labels = None, None
else:
# We are just training, but we still need partitions for rest of the code to interact with.
# dynamic_partition returns a length 1 list in this case instead of just the training set.
train_images = tf.dynamic_partition(images, mask, 1)[0]
train_labels = tf.dynamic_partition(labels, mask, 1)[0]
test_images, test_labels = None, None
val_images, val_labels = None, None
except ValueError:
raise ValueError(
"Images/labels and partition mask have mismatched lengths")
# Also partition moderation features if present <-- NEEDS TO BE FIXED/IMPROVED
train_mf, test_mf, val_mf = None, None, None
if moderation_features is not None:
train_mf, test_mf, val_mf = tf.dynamic_partition(
moderation_features, mask, 2)
return train_images, train_labels, train_mf, test_images, test_labels, test_mf, val_images, val_labels, val_mf
def undo_mask(x, mask, pad_val=0.0):
"""Converts the output of boolean_mask to the original input dimensions.
The boolean_mask is usually used to condense items from multiple batches into
one large 'batch' for faster processing. This function is used to convert
back.
Args:
x: The input to reshape.
mask: The mask used in boolean_mask.
pad_val: value to pad with.
Returns:
x reshaped and padded.
"""
with tf.variable_scope('undo_mask'):
flat_x = tf.reshape(x, [-1])
x_shape = tf.shape(x)[1:]
expanded_mask = tf.tile(
tf.reshape(
mask,
tf.concat([[-1, tf.shape(mask)[1]],
tf.ones_like(x_shape)], 0)),
tf.concat([[1, 1], x_shape], 0))
flat_mask = tf.reshape(expanded_mask, [-1])
start_indices = tf.range(tf.shape(flat_mask)[0])
condition_indices = tf.dynamic_partition(start_indices,
tf.cast(flat_mask, tf.int32),
2)
stitched = tf.dynamic_stitch(condition_indices, [
tf.ones_like(condition_indices[0], tf.float32) * pad_val,
tf.reshape(flat_x, [-1])
])
final_shape = tf.shape(mask)
out_shape = tf.concat([[final_shape[0], final_shape[1]], x_shape], 0)
return tf.reshape(stitched, out_shape)
def random_substr(str_tensor, max_words):
"""Select random substring if the input has more than max_words."""
word_batch_r = tf.strings.split(str_tensor, result_type="RaggedTensor")
row_splits = word_batch_r.row_splits
words = word_batch_r.values
start_idx = row_splits[:-1]
end_idx = row_splits[1:]
words_per_example = end_idx - start_idx
ones = tf.ones_like(end_idx)
max_val = tf.maximum(ones, words_per_example - max_words)
max_words_batch = tf.reduce_max(words_per_example)
rnd = tf.random.uniform(tf.shape(start_idx),
minval=0,
maxval=max_words_batch,
dtype=tf.int64)
off_start_idx = tf.math.floormod(rnd, max_val)
new_words_per_example = tf.where(tf.equal(max_val, 1), words_per_example,
ones * max_words)
new_start_idx = start_idx + off_start_idx
new_end_idx = new_start_idx + new_words_per_example
indices = tf.expand_dims(tf.range(tf.size(words), dtype=tf.int64), axis=0)
within_limit = tf.logical_and(
tf.greater_equal(indices, tf.expand_dims(new_start_idx, axis=1)),
tf.less(indices, tf.expand_dims(new_end_idx, axis=1)))
keep_indices = tf.reduce_any(within_limit, axis=0)
keep_indices = tf.cast(keep_indices, dtype=tf.int32)
_, selected_words = tf.dynamic_partition(words, keep_indices, 2)
row_splits = tf.math.cumsum(new_words_per_example)
row_splits = tf.concat([[0], row_splits], axis=0)
new_tensor = tf.RaggedTensor.from_row_splits(values=selected_words,
row_splits=row_splits)
return tf.strings.reduce_join(new_tensor, axis=1, separator=" ")
def _ComputeLoss():
pair_a = tf.nn.l2_normalize(pregrasp_embedding-postgrasp_embedding, axis=1)
pair_b = tf.nn.l2_normalize(goal_embedding, axis=1)
distances = tf.losses.cosine_distance(
pair_a, pair_b, axis=1, reduction=tf.losses.Reduction.NONE)
_, mask1_data = tf.dynamic_partition(distances, mask, 2)
loss = tf.cast(tf.reduce_mean(mask1_data), tf.float32)
return loss
def __init__(self,
num_train,
lr=None,
batch_size=32,
num_inputs=1,
num_outputs=1,
w_threshold=0.3,
n_hidden=32,
hidden_layers=2,
ckpt_file='tmp.ckpt',
standardize=True,
reg_lambda=None,
reg_beta=None,
DAG_min=0.5):
self.w_threshold = w_threshold
self.DAG_min = DAG_min
if lr is None:
self.learning_rate = 0.001
else:
self.learning_rate = lr
if reg_lambda is None:
self.reg_lambda = 1.
else:
self.reg_lambda = reg_lambda
if reg_beta is None:
self.reg_beta = 1
else:
self.reg_beta = reg_beta
self.batch_size = batch_size
self.num_inputs = num_inputs
self.n_hidden = n_hidden
self.hidden_layers = hidden_layers
self.num_outputs = num_outputs
self.X = tf.placeholder("float", [None, self.num_inputs])
self.y = tf.placeholder("float", [None, 1])
self.rho = tf.placeholder("float", [1, 1])
self.alpha = tf.placeholder("float", [1, 1])
self.keep_prob = tf.placeholder("float")
self.Lambda = tf.placeholder("float")
self.noise = tf.placeholder("float")
self.is_train = tf.placeholder(tf.bool, name="is_train")
self.count = 0
self.max_steps = 200
self.saves = 50
self.patience = 30
self.metric = mean_squared_error
# One-hot vector indicating which nodes are trained
self.sample = tf.placeholder(tf.int32, [self.num_inputs])
# Store layers weight & bias
seed = 1
self.weights = {}
self.biases = {}
# Create the input and output weight matrix for each feature
for i in range(self.num_inputs):
self.weights['w_h0_' + str(i)] = tf.Variable(
tf.random_normal([self.num_inputs, self.n_hidden], seed=seed) *
0.01)
self.weights['out_' + str(i)] = tf.Variable(
tf.random_normal([self.n_hidden, self.num_outputs], seed=seed))
for i in range(self.num_inputs):
self.biases['b_h0_' + str(i)] = tf.Variable(
tf.random_normal([self.n_hidden], seed=seed) * 0.01)
self.biases['out_' + str(i)] = tf.Variable(
tf.random_normal([self.num_outputs], seed=seed))
# The first and second layers are shared
self.weights.update({
'w_h1':
tf.Variable(tf.random_normal([self.n_hidden, self.n_hidden]))
})
self.biases.update(
{'b_h1': tf.Variable(tf.random_normal([self.n_hidden]))})
self.hidden_h0 = {}
self.hidden_h1 = {}
self.layer_1 = {}
self.layer_1_dropout = {}
self.out_layer = {}
self.Out_0 = []
# Mask removes the feature i from the network that is tasked to construct feature i
self.mask = {}
self.activation = tf.nn.relu
for i in range(self.num_inputs):
indices = [i] * self.n_hidden
self.mask[str(i)] = tf.transpose(
tf.one_hot(indices,
depth=self.num_inputs,
on_value=0.0,
off_value=1.0,
axis=-1))
self.weights['w_h0_' +
str(i)] = self.weights['w_h0_' +
str(i)] * self.mask[str(i)]
self.hidden_h0['nn_' + str(i)] = self.activation(
tf.add(tf.matmul(self.X, self.weights['w_h0_' + str(i)]),
self.biases['b_h0_' + str(i)]))
self.hidden_h1['nn_' + str(i)] = self.activation(
tf.add(
tf.matmul(self.hidden_h0['nn_' + str(i)],
self.weights['w_h1']), self.biases['b_h1']))
self.out_layer['nn_' + str(i)] = tf.matmul(
self.hidden_h1['nn_' + str(i)],
self.weights['out_' + str(i)]) + self.biases['out_' + str(i)]
self.Out_0.append(self.out_layer['nn_' + str(i)])
# Concatenate all the constructed features
self.Out = tf.concat(self.Out_0, axis=1)
self.optimizer_subset = tf.train.AdamOptimizer(
learning_rate=self.learning_rate)
self.supervised_loss = tf.reduce_mean(tf.reduce_sum(
tf.square(self.out_layer['nn_0'] - self.y), axis=1),
axis=0)
self.regularization_loss = 0
self.W_0 = []
for i in range(self.num_inputs):
self.W_0.append(
tf.math.sqrt(
tf.reduce_sum(tf.square(self.weights['w_h0_' + str(i)]),
axis=1,
keepdims=True)))
self.W = tf.concat(self.W_0, axis=1)
#truncated power series
d = tf.cast(self.X.shape[1], tf.float32)
coff = 1.0
Z = tf.multiply(self.W, self.W)
dag_l = tf.cast(d, tf.float32)
Z_in = tf.eye(d)
for i in range(1, 10):
Z_in = tf.matmul(Z_in, Z)
dag_l += 1. / coff * tf.linalg.trace(Z_in)
coff = coff * (i + 1)
self.h = dag_l - tf.cast(d, tf.float32)
# Residuals
self.R = self.X - self.Out
# Average reconstruction loss
self.average_loss = 0.5 / num_train * tf.reduce_sum(tf.square(self.R))
#group lasso
L1_loss = 0.0
for i in range(self.num_inputs):
w_1 = tf.slice(self.weights['w_h0_' + str(i)], [0, 0], [i, -1])
w_2 = tf.slice(self.weights['w_h0_' + str(i)], [i + 1, 0],
[-1, -1])
L1_loss += tf.reduce_sum(tf.norm(w_1, axis=1)) + tf.reduce_sum(
tf.norm(w_2, axis=1))
# Divide the residual into untrain and train subset
_, subset_R = tf.dynamic_partition(tf.transpose(self.R),
partitions=self.sample,
num_partitions=2)
subset_R = tf.transpose(subset_R)
#Combine all the loss
self.mse_loss_subset = tf.cast(self.num_inputs, tf.float32) / tf.cast(
tf.reduce_sum(self.sample), tf.float32) * tf.reduce_sum(
tf.square(subset_R))
self.regularization_loss_subset = self.mse_loss_subset + self.reg_beta * L1_loss + 0.5 * self.rho * self.h * self.h + self.alpha * self.h
#Add in supervised loss
self.regularization_loss_subset += self.Lambda * self.rho * self.supervised_loss
self.loss_op_dag = self.optimizer_subset.minimize(
self.regularization_loss_subset)
self.loss_op_supervised = self.optimizer_subset.minimize(
self.supervised_loss + self.regularization_loss)
self.sess = tf.Session()
self.sess.run(tf.global_variables_initializer())
self.saver = tf.train.Saver(var_list=tf.global_variables())
self.tmp = ckpt_file
def split_raw_data(images,
labels,
test_ratio=0,
validation_ratio=0,
moderation_features=None,
augmentation_images=None,
augmentation_labels=None,
split_labels=True):
"""Currently depends on test/validation_ratio being 0 when not using test/validation"""
# serialize labels if they are lists (e.g. for regression)
if isinstance(labels, list):
if split_labels:
labels = [' '.join(map(str, label)) for label in labels]
# check if there is a previously saved mask to load from current directory
mask = []
try:
prev_mask_file = open("mask_ckpt.txt", "r", encoding='utf-8-sig')
found_prev_mask_file = True
print('{0}: {1}'.format(
datetime.datetime.now().strftime("%I:%M%p"),
"Previous mask found. Loading 'mask_ckpt.txt'"))
for line in prev_mask_file:
mask.append(int(line.rstrip()))
prev_mask_file.close()
except Exception:
found_prev_mask_file = False
if not found_prev_mask_file: # we build the mask
print('{0}: {1}'.format(datetime.datetime.now().strftime("%I:%M%p"),
'No previous mask found. Building new mask.'))
total_samples = len(labels)
mask = [0] * total_samples
val_mask_num = 1 # this changes depending on whether we are using testing or not
val_start_idx = 0 # if no testing then we idx from beginning, else we change this if there is testing
if test_ratio != 0:
# creating a mask [1,1,1,...,0,0,0]
num_test = int(total_samples * test_ratio)
mask[:num_test] = [1] * num_test
val_mask_num = 2
val_start_idx = num_test
if validation_ratio != 0:
# if test_ratio != 0 then val_num_mask = 2 and we will create a mask as [1,1,1,...,2,2,2,...,0,0,0,...]
# otherwise we will only have train and validation thus creating a mask as [1,1,1,...,0,0,0]
num_val = int(total_samples * validation_ratio)
mask[val_start_idx:val_start_idx +
num_val] = [val_mask_num] * num_val
# If we're using a training augmentation set, add them to the training portion
if augmentation_images is not None and augmentation_labels is not None:
images = images + augmentation_images
labels = labels + augmentation_labels
mask = mask + ([0] * len(augmentation_labels))
# make the split random <-- ESSENTIAL
random.shuffle(mask)
# save the mask file in current directory for future use
prev_mask_file = open('mask_ckpt.txt', 'w+', encoding='utf-8')
for entry in mask:
prev_mask_file.write(str(entry) + '\n')
prev_mask_file.close()
# create partitions, we set train/validation to None if they're not being used
if test_ratio != 0 and validation_ratio != 0:
train_images, test_images, val_images = tf.dynamic_partition(
images, mask, 3)
train_labels, test_labels, val_labels = tf.dynamic_partition(
labels, mask, 3)
elif test_ratio != 0 and validation_ratio == 0:
train_images, test_images = tf.dynamic_partition(images, mask, 2)
train_labels, test_labels = tf.dynamic_partition(labels, mask, 2)
val_images, val_labels = None, None
elif test_ratio == 0 and validation_ratio != 0:
train_images, val_images = tf.dynamic_partition(images, mask, 2)
train_labels, val_labels = tf.dynamic_partition(labels, mask, 2)
test_images, test_labels = None, None
else:
# must be just training, still need queues for rest of dpp code to load/interact with
# dynamic_partition returns a list, which is fine in the above cases but in the following case it returns
# a list of length 1, hence we index into it with [0] to get what we want
train_images = tf.dynamic_partition(images, mask, 1)[0]
train_labels = tf.dynamic_partition(labels, mask, 1)[0]
test_images, test_labels = None, None
val_images, val_labels = None, None
# Also partition moderation features if present <-- NEEDS TO BE FIXED/IMPROVED
train_mf, test_mf, val_mf = None, None, None
if moderation_features is not None:
train_mf, test_mf, val_mf = tf.dynamic_partition(
moderation_features, mask, 2)
return train_images, train_labels, train_mf, test_images, test_labels, test_mf, val_images, val_labels, val_mf
centroids = tf.Variable(tf.gather(vector_values, centroid_indices))
expanded_vectors = tf.expand_dims(vectors, 0)
expanded_centroids = tf.expand_dims(centroids, 1)
vectors_subtration = tf.subtract(expanded_vectors, expanded_centroids)
euclidean_distances = tf.reduce_sum(tf.square(vectors_subtration), 2)
assignments = tf.to_int32(tf.argmin(euclidean_distances, 0))
partitions = [0, 0, 1, 1, 0]
num_partitions = 2
data = [10, 20, 30, 40, 50]
#outputs[0] = [10, 20, 50]
#outputs[1] = [30, 40]
partitions = tf.dynamic_partition(vectors, assignments, num_clusters)
update_centroids = tf.concat(0, [
tf.expand_dims(tf.reduce_mean(partition, 0), 0) for partition in partitions
])
init_op = tf.initialize_all_variables()
sess = tf.Session()
sess.run(init_op)
for step in range(num_steps):
_, centroid_values, assignment_values =\
sess.run([update_centroids,\
centroids,\
assignments])
def _ComputeLoss():
raw_distances = pregrasp_embedding - goal_embedding - postgrasp_embedding
distances = tf.reduce_sum(raw_distances**2, axis=1)
_, mask1_data = tf.dynamic_partition(distances, mask, 2)
loss = tf.cast(tf.reduce_mean(mask1_data), tf.float32)
return loss
def _ComputeLoss():
distances = tf.norm(tensor, axis=1)
_, mask1_data = tf.dynamic_partition(distances, mask, 2)
loss = tf.cast(tf.reduce_mean(mask1_data), tf.float32)
return loss