Skip to content

GitLab

Commit f63f7274 authored by Tiago de Freitas Pereira's avatar Tiago de Freitas Pereira
Browse files

Fisher loss

parent 36c835c0
Pipeline #7688 failed with stages
in 4 minutes and 21 seconds
Hide whitespace changes
Inline Side-by-side
bob/learn/tensorflow/datashuffler/TripletWithSelectionMemory.py
View file @ f63f7274
......@@ -9,8 +9,11 @@ import tensorflow as tf
from .OnlineSampling import OnlineSampling
from .Memory import Memory
from .Triplet import Triplet
from scipy.spatial.distance import euclidean
from bob.learn.tensorflow.datashuffler.Normalizer import Linear
from scipy.spatial.distance import euclidean, cdist
import logging
logger = logging.getLogger("bob.learn")
class TripletWithSelectionMemory(Triplet, Memory, OnlineSampling):
......@@ -61,7 +64,6 @@ class TripletWithSelectionMemory(Triplet, Memory, OnlineSampling):
def __init__(self, data, labels,
input_shape,
input_dtype="float64",
scale=True,
batch_size=1,
seed=10,
data_augmentation=None,
......@@ -73,7 +75,6 @@ class TripletWithSelectionMemory(Triplet, Memory, OnlineSampling):
labels=labels,
input_shape=input_shape,
input_dtype=input_dtype,
scale=scale,
batch_size=batch_size,
seed=seed,
data_augmentation=data_augmentation,
......@@ -86,26 +87,7 @@ class TripletWithSelectionMemory(Triplet, Memory, OnlineSampling):
self.data = self.data.astype(input_dtype)
self.total_identities = total_identities
self.first_batch = True
def get_random_batch(self):
"""
Get a random triplet
**Parameters**
is_target_set_train: Defining the target set to get the batch
**Return**
"""
data_a = numpy.zeros(shape=self.shape, dtype='float32')
data_p = numpy.zeros(shape=self.shape, dtype='float32')
data_n = numpy.zeros(shape=self.shape, dtype='float32')
for i in range(self.shape[0]):
data_a[i, ...], data_p[i, ...], data_n[i, ...] = self.get_one_triplet(self.data, self.labels)
return [data_a, data_p, data_n]
self.batch_increase_factor = 4
def get_batch(self):
"""
......@@ -117,54 +99,27 @@ class TripletWithSelectionMemory(Triplet, Memory, OnlineSampling):
**Return**
"""
if self.first_batch:
self.first_batch = False
return self.get_random_batch()
# Selecting the classes used in the selection
indexes = numpy.random.choice(len(self.possible_labels), self.total_identities, replace=False)
samples_per_identity = self.batch_size/self.total_identities
anchor_labels = numpy.ones(samples_per_identity) * indexes[0]
samples_per_identity = numpy.ceil(self.batch_size/float(self.total_identities))
anchor_labels = numpy.ones(samples_per_identity) * self.possible_labels[indexes[0]]
for i in range(1, self.total_identities):
anchor_labels = numpy.hstack((anchor_labels,numpy.ones(samples_per_identity) * indexes[i]))
anchor_labels = numpy.hstack((anchor_labels,numpy.ones(samples_per_identity) * self.possible_labels[indexes[i]]))
anchor_labels = anchor_labels[0:self.batch_size]
data_a = numpy.zeros(shape=self.shape, dtype='float32')
data_p = numpy.zeros(shape=self.shape, dtype='float32')
data_n = numpy.zeros(shape=self.shape, dtype='float32')
# Fetching the anchors
for i in range(self.shape[0]):
data_a[i, ...] = self.get_anchor(anchor_labels[i])
features_a = self.project(data_a)
samples_a = numpy.zeros(shape=self.shape, dtype='float32')
# Computing the embedding
for i in range(self.shape[0]):
label = anchor_labels[i]
#anchor = self.get_anchor(label)
positive, distance_anchor_positive = self.get_positive(label, features_a[i])
negative = self.get_negative(label, features_a[i], distance_anchor_positive)
data_p[i, ...] = positive
data_n[i, ...] = negative
# Applying the data augmentation
if self.data_augmentation is not None:
for i in range(data_a.shape[0]):
d = self.bob2skimage(self.data_augmentation(self.skimage2bob(data_a[i, ...])))
data_a[i, ...] = d
d = self.bob2skimage(self.data_augmentation(self.skimage2bob(data_p[i, ...])))
data_p[i, ...] = d
d = self.bob2skimage(self.data_augmentation(self.skimage2bob(data_n[i, ...])))
data_n[i, ...] = d
samples_a[i, ...] = self.get_anchor(anchor_labels[i])
embedding_a = self.project(samples_a)
# Scaling
data_a = self.normalize_sample(data_a)
data_p = self.normalize_sample(data_p)
data_n = self.normalize_sample(data_n)
# Getting the positives
samples_p, embedding_p, d_anchor_positive = self.get_positives(anchor_labels, embedding_a)
samples_n = self.get_negative(anchor_labels, embedding_a, d_anchor_positive)
return data_a, data_p, data_n
return samples_a, samples_p, samples_n
def get_anchor(self, label):
"""
......@@ -175,61 +130,67 @@ class TripletWithSelectionMemory(Triplet, Memory, OnlineSampling):
indexes = numpy.where(self.labels == label)[0]
numpy.random.shuffle(indexes)
return self.data[indexes[0], ...]
return self.normalize_sample(self.data[indexes[0], ...])
def get_positive(self, label, anchor_feature):
def get_positives(self, anchor_labels, embedding_a):
"""
Get the best positive sample given the anchor.
The best positive sample for the anchor is the farthest from the anchor
Get the a random set of positive pairs
"""
samples_p = numpy.zeros(shape=self.shape, dtype='float32')
for i in range(self.shape[0]):
l = anchor_labels[i]
indexes = numpy.where(self.labels == l)[0]
numpy.random.shuffle(indexes)
samples_p[i, ...] = self.normalize_sample(self.data[indexes[0], ...])
# Projecting the anchor
#anchor_feature = self.feature_extractor(self.reshape_for_deploy(anchor), session=self.session)
indexes = numpy.where(self.labels == label)[0]
numpy.random.shuffle(indexes)
indexes = indexes[
0:self.batch_size] # Limiting to the batch size, otherwise the number of comparisons will explode
distances = []
positive_features = self.project(self.data[indexes, ...])
embedding_p = self.project(samples_p)
# Projecting the positive instances
for p in positive_features:
distances.append(euclidean(anchor_feature, p))
# Computing the distances
d_anchor_positive = []
for i in range(self.shape[0]):
d_anchor_positive.append(euclidean(embedding_a[i, :], embedding_p[i, :]))
# Geting the max
index = numpy.argmax(distances)
return self.data[indexes[index], ...], distances[index]
return samples_p, embedding_p, d_anchor_positive
def get_negative(self, label, anchor_feature, distance_anchor_positive):
def get_negative(self, anchor_labels, embedding_a, d_anchor_positive):
"""
Get the best negative sample for a pair anchor-positive
Get the the semi-hard negative
"""
# Projecting the anchor
#anchor_feature = self.feature_extractor(self.reshape_for_deploy(anchor), session=self.session)
# Selecting the negative samples
indexes = numpy.where(self.labels != label)[0]
# Shuffling all the dataset
indexes = range(len(self.labels))
numpy.random.shuffle(indexes)
indexes = indexes[
0:self.batch_size] # Limiting to the batch size, otherwise the number of comparisons will explode
negative_features = self.project(self.data[indexes, ...])
distances = []
for n in negative_features:
d = euclidean(anchor_feature, n)
negative_samples_search = self.batch_size * self.batch_increase_factor
# Limiting to the batch size, otherwise the number of comparisons will explode
indexes = indexes[0:negative_samples_search]
# Loading samples for the semi-hard search
shape = tuple([len(indexes)] + list(self.shape[1:]))
temp_samples_n = numpy.zeros(shape=shape, dtype='float32')
samples_n = numpy.zeros(shape=self.shape, dtype='float32')
for i in range(shape[0]):
temp_samples_n[i, ...] = self.normalize_sample(self.data[indexes[i], ...])
# Computing all the embeddings
embedding_temp_n = self.project(temp_samples_n)
# Semi-hard samples criteria
if d > distance_anchor_positive:
distances.append(d)
else:
distances.append(numpy.inf)
# Computing the distances
d_anchor_negative = cdist(embedding_a, embedding_temp_n, metric='euclidean')
# Selecting the negative samples
for i in range(self.shape[0]):
label = anchor_labels[i]
possible_candidates = [d if d > d_anchor_positive[i] else numpy.inf for d in d_anchor_negative[i]]
# Getting the minimum negative sample as the reference for the pair
index = numpy.argmin(distances)
for j in numpy.argsort(possible_candidates):
# if the semi-hardest is inf take the first
if numpy.isinf(distances[index]):
index = 0
# Checking if they don't have the same label
if self.labels[indexes[j]] != label:
samples_n[i, ...] = temp_samples_n[j, ...]
if numpy.isinf(possible_candidates[j]):
logger.info("SEMI-HARD negative not found, took the first one")
break
return self.data[indexes[index], ...]
return samples_n
bob/learn/tensorflow/loss/TripletFisherLoss.py
View file @ f63f7274
......@@ -15,8 +15,8 @@ class TripletFisherLoss(BaseLoss):
"""
"""
def __init__(self, margin=0.2):
self.margin = margin
def __init__(self):
pass
def __call__(self, anchor_embedding, positive_embedding, negative_embedding):
......@@ -31,6 +31,7 @@ class TripletFisherLoss(BaseLoss):
tf.reduce_mean(negative_embedding, axis=0)), 2)
length = anchor_embedding.get_shape().as_list()[0]
dim = anchor_embedding.get_shape().as_list()[1]
split_positive = tf.unstack(positive_embedding, num=length, axis=0)
split_negative = tf.unstack(negative_embedding, num=length, axis=0)
......@@ -40,11 +41,11 @@ class TripletFisherLoss(BaseLoss):
positive = s[0]
negative = s[1]
buffer_sw = tf.reshape(tf.subtract(positive, average_class), shape=(2, 1))
buffer_sw = tf.matmul(buffer_sw, tf.reshape(buffer_sw, shape=(1, 2)))
buffer_sw = tf.reshape(tf.subtract(positive, average_class), shape=(dim, 1))
buffer_sw = tf.matmul(buffer_sw, tf.reshape(buffer_sw, shape=(1, dim)))
buffer_sb = tf.reshape(tf.subtract(negative, average_total), shape=(2, 1))
buffer_sb = tf.matmul(buffer_sb, tf.reshape(buffer_sb, shape=(1, 2)))
buffer_sb = tf.reshape(tf.subtract(negative, average_total), shape=(dim, 1))
buffer_sb = tf.matmul(buffer_sb, tf.reshape(buffer_sb, shape=(1, dim)))
if Sw is None:
Sw = buffer_sw
......
bob/learn/tensorflow/network/VGG16.py
View file @ f63f7274
......@@ -225,17 +225,22 @@ class VGG16(SequenceNetwork):
))
self.add(MaxPooling(name="pooling5", strides=[1, 2, 2, 1]))
self.add(FullyConnected(name="fc6", output_dim=fc6_output,
activation=tf.nn.relu,
weights_initialization=Xavier(seed=seed, use_gpu=self.use_gpu),
bias_initialization=Constant(use_gpu=self.use_gpu)
))
self.add(Dropout(name="dropout", keep_prob=0.5))
self.add(FullyConnected(name="fc7", output_dim=fc7_output,
activation=tf.nn.relu,
weights_initialization=Xavier(seed=seed, use_gpu=self.use_gpu),
bias_initialization=Constant(use_gpu=self.use_gpu)
))
self.add(Dropout(name="dropout", keep_prob=0.5))
self.add(FullyConnected(name="fc8", output_dim=n_classes,
activation=None,
......
bob/learn/tensorflow/script/train.py
View file @ f63f7274
......@@ -43,7 +43,7 @@ def main():
prefetch=PREFETCH,
learning_rate=config.learning_rate,
temp_dir=OUTPUT_DIR,
snapshot=10,
snapshot=100,
model_from_file=PRETRAINED_NET,
use_gpu=USE_GPU
)
......
bob/learn/tensorflow/utils/session.py
View file @ f63f7274
......@@ -11,8 +11,9 @@ from tensorflow.python import debug as tf_debug
class Session(object):
def __init__(self):
config = tf.ConfigProto(log_device_placement=True,
gpu_options=tf.GPUOptions(per_process_gpu_memory_fraction=0.333))
config = tf.ConfigProto(log_device_placement=False,
allow_soft_placement=True,
gpu_options=tf.GPUOptions(per_process_gpu_memory_fraction=0.5))
config.gpu_options.allow_growth = True
self.session = tf.Session()
#self.session = tf_debug.LocalCLIDebugWrapperSession(self.session)
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment