def quadratic_weighted_kappa_cm(conf_mat, num_ratings, cost_matrix):
"""
Compute QWK function using confusion matrix.
:param conf_mat: confusion matrix.
:param min_rating: lowest rating.
:param max_rating: highest rating.
:param cost_matrix: cost_matrix.
:return: QWK value.
"""
conf_mat = K.cast(conf_mat, dtype=K.floatx())
hist_rater_a = K.cast(K.reshape(K.reduce_sum(conf_mat, axis=1),
shape=[num_ratings, 1]),
dtype=K.floatx()) # Sum every row
hist_rater_b = K.cast(K.reshape(K.reduce_sum(conf_mat, axis=0),
shape=[1, num_ratings]),
dtype=K.floatx()) # Sum every column
num_scored_items = K.reduce_sum(conf_mat) # Sum all the elements
expected_count = K.matmul(hist_rater_a, hist_rater_b) / K.cast(
num_scored_items, dtype=K.floatx())
numerator = K.reduce_sum(cost_matrix * conf_mat)
denominator = K.reduce_sum(cost_matrix * expected_count)
return 1.0 - numerator / denominator
def keras_loss(y_true, y_pred, alpha=0.2):
# origin image
anchor = y_pred[0]
# same person
positive = y_pred[1]
# another person
negative = y_pred[2]
# operation
pos_dist = K.reduce_sum(K.square(K.subtract(positive, anchor)))
neg_dist = K.reduce_sum(K.square(K.subtract(negative, anchor)))
loss = K.maximum(
K.reduce_sum(K.add(K.subtract(pos_dist, neg_dist), alpha)), 0.0)
return loss
def logp(y_true, y_pred):
# Y is our n_instances x n_classes tensor of permutations
# F is n_instances x n_classes (n_classes many latent processes
def logsumexp(x):
masked = K.reshape(
K.boolean_mask(exped,
K.greater_equal(y_pred, K.cast(x, 'float32'))),
[n, -1])
return max_entry + K.log(K.reduce_sum(masked, axis=1))
n, m = K.shape(y_true)[0], K.shape(y_true)[1]
max_entry = K.reduce_max(y_true)
exped = K.exp(y_true - max_entry)
lse = K.map_fn(logsumexp, K.range(m), dtype='float32')
return K.reduce_sum(y_true) - K.reduce_sum(lse)
def kernel_pooling(translation_matrix, l_mean, l_sigma):
"""
function for Lambda layer
kernel pooling layer
:param translation_matrix: input translation matrix
:param l_mean: kp's mean
:param l_sigma: sigma for kernels
:return: output_shape = 1 + (hyper_parameter.l_kernel_pool_mean) # added 1 exact match pool
"""
#TODO
assert len(l_mean) == len(l_sigma)
mu = np.array(l_mean) # add exact match kernel
mu.reshape((1, 1, len(l_mean)))
sigma = np.array(l_sigma)
sigma.reshape((1, 1, len(l_sigma)))
m = TF.expand_dims(translation_matrix, -1)
raw_k_pool = TF.exp(
TF.div(TF.negative(TF.square(TF.sub(m, mu))),
TF.mul(TF.square(sigma), 2)))
k_pool = TF.reduce_sum(raw_k_pool, [0, 1])
k_pool = TF.log(TF.maximum(k_pool, 1e-10)) * 0.01
return k_pool
def gaussian_loss(y_true, y_pred):
# define loss function (eq 26 of http://arxiv.org/abs/1308.0850)
x_data, y_data = y_true[:, 1:]
eos_data = y_true[:, 0:1]
pi, mux, muy, sigmax, sigmay, rho = K.split(y_pred[:, 1:], 6, 1)
eos = y_pred[:, 0:1]
gaussian = gaussian2d(x_data, y_data, mux, muy, sigmax, sigmay,
rho)
term1 = K.multiply(gaussian, pi)
term1 = K.reduce_sum(term1, 1,
keep_dims=True) # do inner summation
term1 = -K.log(K.maximum(
term1, 1e-20)) # some errors are zero -> numerical errors.
term2 = K.multiply(eos, eos_data) + K.multiply(
1 - eos, 1 - eos_data) # modified Bernoulli -> eos probability
term2 = -K.log(term2) # negative log error gives loss
return K.reduce_sum(term1 + term2) # do outer summation
def GPU_lid_eval_keras(logits, k=20):
import time
start_time = time.time()
"""
Calculate LID for a minibatch of training samples based on the outputs of the network.
:param logits:
:param k:
:return:
"""
print(logits.shape)
logits = K.constant(logits, dtype=tf.float32)
epsilon = 1e-12
batch_size = K.shape(logits)[0]
# n_samples = logits.get_shape().as_list()
# calculate pairwise distance
r = K.reduce_sum(logits * logits, 1)
# turn r into column vector
r1 = K.reshape(r, [-1, 1])
D = r1 - 2 * K.matmul(logits, K.transpose(logits)) + K.transpose(r1) + \
K.ones([batch_size, batch_size])
# find the k nearest neighbor
D1 = -K.sqrt(D)
D2, _ = tf.nn.top_k(D1, k=k, sorted=True)
D3 = -D2[:, 1:] # skip the x-to-x distance 0 by using [,1:]
m = K.transpose(K.multiply(K.transpose(D3), 1.0 / D3[:, -1]))
v_log = K.reduce_sum(tf.log(m + epsilon), axis=1) # to avoid nan
lids = -k / v_log
## l2 normalization
# lids = tf.nn.l2_normalize(lids, dim=0, epsilon=epsilon)
# import ipdb;
# ipdb.set_trace()
# sess = tf.Session()
# # with sess.as_default():
# # lids = sess.run(lids)
print('LID GPU Time:', time.time() - start_time)
return lids.eval()
def batch_all_triplet_loss(y_true, y_pred, margin, squared=False):
pairwise_dist = _pairwise_distances(y_pred, squared=squared)
anchor_positive_dist = K.expand_dims(pairwise_dist, 2)
anchor_negative_dist = K.expand_dims(pairwise_dist, 1)
triplet_loss = anchor_positive_dist - anchor_negative_dist + margin
mask = _get_triplet_mask(labels)
mask = K.to_float(mask)
triplet_loss = K.multiply(mask, triplet_loss)
triplet_loss = K.maximum(triplet_loss, 0.0)
valid_triplets = K.to_float(K.greater(triplet_loss, 1e-16))
num_positive_triplets = K.reduce_sum(valid_triplets)
num_valid_triplets = K.reduce_sum(mask)
fraction_positive_triplets = num_positive_triplets / (num_valid_triplets +
1e-16)
triplet_loss = K.reduce_sum(triplet_loss) / (num_positive_triplets + 1e-16)
return triplet_loss, fraction_postive_triplets
def func(y_true, y_pred):
return K.reduce_sum(y_true * x + (y_true - 1) * x)
def K_parzen(x, mu, sigma):
d = (K.expand_dims(x, 1) - K.expand_dims(mu, 0)) / sigma
e = K_log_mean_exp(-0.5 * K.reduce_sum(K.multiply(d, d), axis=2), axis=1)
e = K.squeeze(e, axis=1)
z = K.to_float(K.shape(mu)[1]) * K.log(sigma * np.sqrt(np.pi * 2.0))
return e - z
def logsumexp(x):
masked = K.reshape(
K.boolean_mask(exped,
K.greater_equal(y_pred, K.cast(x, 'float32'))),
[n, -1])
return max_entry + K.log(K.reduce_sum(masked, axis=1))
def add_loss(locations, confidences, batched_bboxes, batched_num_bboxes,
bbox_priors, location_loss_alpha):
batch_size = tf.shape(locations)[0] # locations.get_shape().as_list()[0]
# ground truth bounding boxes:
# [batch_size, # of ground truth bounding boxes, 4]
# we also need to know the number of ground truth bounding boxes for each image in the batch
# (it can be different for each image...)
# We could assume 1 for now.
# Pass the locations, confidences, and ground truth labels to the matching function
locations = tf.reshape(locations, [-1, 4])
confidences = tf.reshape(confidences, [-1])
# add the priors to the predicted residuals
locations += tf.tile(bbox_priors, [batch_size, 1])
# add a small epsilon to the confidences
confidences += SMALL_EPSILON
# print "Shapes"
# print locations.get_shape().as_list()
# print confidences.get_shape().as_list()
# print batched_bboxes.get_shape().as_list()
# print batched_num_bboxes.get_shape().as_list()
params = [
locations, confidences, batched_bboxes, batched_num_bboxes, batch_size,
location_loss_alpha
]
matching, stacked_gt_bboxes = tf.py_func(compute_assignments,
params, [tf.int32, tf.float32],
name="bipartite_matching")
# matching: [num_predictions * batch_size] 0s and 1s for partitioning
# stacked_gt_bboxes : [total number of gt bboxes for this batch, 4]
# dynamic partition the bounding boxes and confidences into "positives" and "negatives"
unmatched_locations, matched_locations = tf.dynamic_partition(
locations, matching, 2)
unmatched_confidences, matched_confidences = tf.dynamic_partition(
confidences, matching, 2)
# Because we just did a dynamic partition, it could be the case that either the unmatched or matched matrices is empty.
# It could also be the case that there were no ground truth bboxes in this batch.
# Lets tack on some default values so that the loss calculations are well behaved.
matched_locations = tf.concat(0, [matched_locations, tf.zeros([1, 4])])
stacked_gt_bboxes = tf.concat(0, [stacked_gt_bboxes, tf.zeros([1, 4])])
matched_confidences = tf.concat(0, [matched_confidences, tf.ones([1])])
unmatched_confidences = tf.concat(
0, [unmatched_confidences, tf.zeros([1])])
location_loss = location_loss_alpha * tf.nn.l2_loss(matched_locations -
stacked_gt_bboxes)
confidence_loss = -1. * tf.reduce_sum(
tf.log(matched_confidences)) - tf.reduce_sum(
tf.log((1. - unmatched_confidences) + SMALL_EPSILON))
# It could be the case that there are no ground truth bounding boxes
# num_gt_bboxes = tf.reduce_sum(batched_num_bboxes)
# loc_loss = lambda: location_loss_alpha * tf.nn.l2_loss(matched_locations - stacked_gt_bboxes)
# zero_loc_loss = lambda: tf.zeros(shape=[])
# location_loss = tf.cond(num_gt_bboxes > 0, loc_loss, zero_loc_loss)
# conf_loss = lambda: -1. * tf.reduce_sum(tf.log(matched_confidences)) - tf.reduce_sum(tf.log((1. - unmatched_confidences) + SMALL_EPSILON))
# all_negative_conf_loss = lambda : -1. * tf.reduce_sum(tf.log((1. - unmatched_confidences) + SMALL_EPSILON))
# confidence_loss = tf.cond(num_gt_bboxes > 0, conf_loss, all_negative_conf_loss)
slim.losses.add_loss(location_loss)
slim.losses.add_loss(confidence_loss)
return location_loss, confidence_loss
def dotLayer(x):
# Check axis to reduce sum along
# Dot product operation
return Lambda(lambda x: K.reduce_sum(
K.multiply(x[0], x[1]), axis=-1, keep_dims=True),
name='DotLayer')
