TripletWithFastSelectionDisk.py 8.06 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231
#!/usr/bin/env python
# vim: set fileencoding=utf-8 :
# @author: Tiago de Freitas Pereira <tiago.pereira@idiap.ch>
# @date: Wed 11 May 2016 09:39:36 CEST 

import numpy
import tensorflow as tf

from .Disk import Disk
from .Triplet import Triplet
from .OnlineSampling import OnLineSampling
from scipy.spatial.distance import euclidean, cdist

import logging
logger = logging.getLogger("bob.learn.tensorflow")


class TripletWithFastSelectionDisk(Triplet, Disk, OnLineSampling):
    """
    This data shuffler generates triplets from :py:class:`bob.learn.tensorflow.datashuffler.Memory` shufflers.

    The selection of the triplets is inspired in the paper:

    Schroff, Florian, Dmitry Kalenichenko, and James Philbin.
    "Facenet: A unified embedding for face recognition and clustering." Proceedings of the IEEE Conference on
    Computer Vision and Pattern Recognition. 2015.


    In this shuffler, the triplets are selected as the following:
      1. Select M identities
      2. Get N pairs anchor-positive (for each M identities) such that the argmax(anchor, positive)
      3. For each pair anchor-positive, find the "semi-hard" negative samples such that
      argmin(||f(x_a) - f(x_p)||^2 < ||f(x_a) - f(x_n)||^2

     **Parameters**
       data:
       labels:
       perc_train:
       scale:
       train_batch_size:
       validation_batch_size:
       data_augmentation:
       total_identities: Number of identities inside of the batch
    """

    def __init__(self, data, labels,
                 input_shape,
                 input_dtype="float64",
                 scale=True,
                 batch_size=1,
                 seed=10,
                 data_augmentation=None,
                 total_identities=10):

        super(TripletWithFastSelectionDisk, self).__init__(
            data=data,
            labels=labels,
            input_shape=input_shape,
            input_dtype=input_dtype,
            scale=scale,
            batch_size=batch_size,
            seed=seed,
            data_augmentation=data_augmentation
        )
        self.clear_variables()
        # Seting the seed
        numpy.random.seed(seed)

        self.total_identities = total_identities
        self.first_batch = True

        # For the negative search I'll load `N` times the batch
        self.batch_increase_factor = 4

    def get_random_batch(self):
        """
        Get a random triplet

        **Parameters**
            is_target_set_train: Defining the target set to get the batch

        **Return**
        """

        sample_a = numpy.zeros(shape=self.shape, dtype='float32')
        sample_p = numpy.zeros(shape=self.shape, dtype='float32')
        sample_n = numpy.zeros(shape=self.shape, dtype='float32')

        for i in range(self.shape[0]):
            file_name_a, file_name_p, file_name_n = self.get_one_triplet(self.data, self.labels)
            sample_a[i, ...] = self.load_from_file(str(file_name_a))
            sample_p[i, ...] = self.load_from_file(str(file_name_p))
            sample_n[i, ...] = self.load_from_file(str(file_name_n))

        if self.scale:
            sample_a *= self.scale_value
            sample_p *= self.scale_value
            sample_n *= self.scale_value

        return [sample_a, sample_p, sample_n]

    def get_batch(self):
        """
        Get SELECTED triplets

        **Parameters**
            is_target_set_train: Defining the target set to get the batch

        **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 = 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) * self.possible_labels[indexes[i]]))
        anchor_labels = anchor_labels[0:self.batch_size]

        samples_a = numpy.zeros(shape=self.shape, dtype='float32')

        # Computing the embedding
        for i in range(self.shape[0]):
            samples_a[i, ...] = self.get_anchor(anchor_labels[i])
        embedding_a = self.project(samples_a)

        # 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)

        #import bob.io.base
        #import bob.io.image
        #for i in range(self.shape[0]):
            #bob.io.base.save((self.skimage2bob(samples_a[i, ...]) / 0.00390625).astype("uint8"), "{0}_a.jpg".format(i))
            #bob.io.base.save((self.skimage2bob(samples_p[i, ...]) / 0.00390625).astype("uint8"), "{0}_p.jpg".format(i))
            #bob.io.base.save((self.skimage2bob(samples_n[i, ...]) / 0.00390625).astype("uint8"), "{0}_n.jpg".format(i))

        return samples_a, samples_p, samples_n

    def get_anchor(self, label):
        """
        Select random samples as anchor
        """

        # Getting the indexes of the data from a particular client
        indexes = numpy.where(self.labels == label)[0]
        numpy.random.shuffle(indexes)

        file_name = self.data[indexes[0], ...]
        anchor = self.load_from_file(str(file_name))

        if self.scale:
            anchor *= self.scale_value

        return anchor

    def get_positives(self, anchor_labels, embedding_a):
        """
        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)
            file_name = self.data[indexes[0], ...]
            samples_p[i, ...] = self.load_from_file(str(file_name))

        if self.scale:
            samples_p *= self.scale_value

        embedding_p = self.project(samples_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, :]))

        return samples_p, embedding_p, d_anchor_positive

    def get_negative(self, anchor_labels, embedding_a, d_anchor_positive):
        """
        Get the the semi-hard negative
        """

        # Shuffling all the dataset
        indexes = range(len(self.labels))
        numpy.random.shuffle(indexes)

        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]):
            file_name = self.data[indexes[i], ...]
            temp_samples_n[i, ...] = self.load_from_file(str(file_name))

        if self.scale:
            temp_samples_n *= self.scale_value

        # Computing all the embeddings
        embedding_temp_n = self.project(temp_samples_n)

        # 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]]

            for j in numpy.argsort(possible_candidates):

                # Checking if they don't have the same label
                if 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 samples_n