code
代码地址：https://github.com/davidsandberg/facenet
这是人脸识别算法的TensorFlow实现，在论文“ FaceNet：人脸识别和聚类的统一嵌入”中进行了介绍。 该项目还使用了牛津大学视觉几何学小组“深度面部识别”一文中的想法。
第一部分
图1 光照和位姿不变性 光照和位姿不变性。姿态和光照是人脸识别中长期存在的问题。该图显示了FaceNet在不同的姿势和光照组合下对相同的人脸和不同的人脸之间的输出距离。距离为0.0表示两张脸是相同的，4.0表示相反的光谱，两种不同的身份。可以看到，阈值为1.1将正确地对每一对进行分类。
第二部分
模型结构。我们的网络由一个批处理输入层和一个深度CNN和L2 归一化组成，然后输出结果是人脸嵌入，接下来是训练中三元组损失函数。
图2 模型结构The network consists of a batch input layer and a deep CNN followed by L2 normalization, which results in the face embedding. This is followed by the triplet loss during training.
这里附上附上训练三元组损失的代码
代码语言：javascript复制from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from datetime import datetime
import os.path
import time
import sys
import tensorflow as tf
import numpy as np
import importlib
import itertools
import argparse
import facenet
import lfw
from tensorflow.python.ops import data_flow_ops
from six.moves import xrange # @UnresolvedImport
def main(args):
network = importlib.import_module(args.model_def)
subdir = datetime.strftime(datetime.now(), '%Y%m%d-%H%M%S')
log_dir = os.path.join(os.path.expanduser(args.logs_base_dir), subdir)
if not os.path.isdir(log_dir): # Create the log directory if it doesn't exist
os.makedirs(log_dir)
model_dir = os.path.join(os.path.expanduser(args.models_base_dir), subdir)
if not os.path.isdir(model_dir): # Create the model directory if it doesn't exist
os.makedirs(model_dir)
# Write arguments to a text file
facenet.write_arguments_to_file(args, os.path.join(log_dir, 'arguments.txt'))
# Store some git revision info in a text file in the log directory
src_path,_ = os.path.split(os.path.realpath(__file__))
facenet.store_revision_info(src_path, log_dir, ' '.join(sys.argv))
np.random.seed(seed=args.seed)
train_set = facenet.get_dataset(args.data_dir)
print('Model directory: %s' % model_dir)
print('Log directory: %s' % log_dir)
if args.pretrained_model:
print('Pre-trained model: %s' % os.path.expanduser(args.pretrained_model))
if args.lfw_dir:
print('LFW directory: %s' % args.lfw_dir)
# Read the file containing the pairs used for testing
pairs = lfw.read_pairs(os.path.expanduser(args.lfw_pairs))
# Get the paths for the corresponding images
lfw_paths, actual_issame = lfw.get_paths(os.path.expanduser(args.lfw_dir), pairs)
with tf.Graph().as_default():
tf.set_random_seed(args.seed)
global_step = tf.Variable(0, trainable=False)
# Placeholder for the learning rate
learning_rate_placeholder = tf.placeholder(tf.float32, name='learning_rate')
batch_size_placeholder = tf.placeholder(tf.int32, name='batch_size')
phase_train_placeholder = tf.placeholder(tf.bool, name='phase_train')
image_paths_placeholder = tf.placeholder(tf.string, shape=(None,3), name='image_paths')
labels_placeholder = tf.placeholder(tf.int64, shape=(None,3), name='labels')
input_queue = data_flow_ops.FIFOQueue(capacity=100000,
dtypes=[tf.string, tf.int64],
shapes=[(3,), (3,)],
shared_name=None, name=None)
enqueue_op = input_queue.enqueue_many([image_paths_placeholder, labels_placeholder])
nrof_preprocess_threads = 4
images_and_labels = []
for _ in range(nrof_preprocess_threads):
filenames, label = input_queue.dequeue()
images = []
for filename in tf.unstack(filenames):
file_contents = tf.read_file(filename)
image = tf.image.decode_image(file_contents, channels=3)
if args.random_crop:
image = tf.random_crop(image, [args.image_size, args.image_size, 3])
else:
image = tf.image.resize_image_with_crop_or_pad(image, args.image_size, args.image_size)
if args.random_flip:
image = tf.image.random_flip_left_right(image)
#pylint: disable=no-member
image.set_shape((args.image_size, args.image_size, 3))
images.append(tf.image.per_image_standardization(image))
images_and_labels.append([images, label])
image_batch, labels_batch = tf.train.batch_join(
images_and_labels, batch_size=batch_size_placeholder,
shapes=[(args.image_size, args.image_size, 3), ()], enqueue_many=True,
capacity=4 * nrof_preprocess_threads * args.batch_size,
allow_smaller_final_batch=True)
image_batch = tf.identity(image_batch, 'image_batch')
image_batch = tf.identity(image_batch, 'input')
labels_batch = tf.identity(labels_batch, 'label_batch')
# Build the inference graph
prelogits, _ = network.inference(image_batch, args.keep_probability,
phase_train=phase_train_placeholder, bottleneck_layer_size=args.embedding_size,
weight_decay=args.weight_decay)
embeddings = tf.nn.l2_normalize(prelogits, 1, 1e-10, name='embeddings')
# Split embeddings into anchor, positive and negative and calculate triplet loss
anchor, positive, negative = tf.unstack(tf.reshape(embeddings, [-1,3,args.embedding_size]), 3, 1)
triplet_loss = facenet.triplet_loss(anchor, positive, negative, args.alpha)
learning_rate = tf.train.exponential_decay(learning_rate_placeholder, global_step,
args.learning_rate_decay_epochs*args.epoch_size, args.learning_rate_decay_factor, staircase=True)
tf.summary.scalar('learning_rate', learning_rate)
# Calculate the total losses
regularization_losses = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
total_loss = tf.add_n([triplet_loss] regularization_losses, name='total_loss')
# Build a Graph that trains the model with one batch of examples and updates the model parameters
train_op = facenet.train(total_loss, global_step, args.optimizer,
learning_rate, args.moving_average_decay, tf.global_variables())
# Create a saver
saver = tf.train.Saver(tf.trainable_variables(), max_to_keep=3)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.summary.merge_all()
# Start running operations on the Graph.
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=args.gpu_memory_fraction)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
# Initialize variables
sess.run(tf.global_variables_initializer(), feed_dict={phase_train_placeholder:True})
sess.run(tf.local_variables_initializer(), feed_dict={phase_train_placeholder:True})
summary_writer = tf.summary.FileWriter(log_dir, sess.graph)
coord = tf.train.Coordinator()
tf.train.start_queue_runners(coord=coord, sess=sess)
with sess.as_default():
if args.pretrained_model:
print('Restoring pretrained model: %s' % args.pretrained_model)
saver.restore(sess, os.path.expanduser(args.pretrained_model))
# Training and validation loop
epoch = 0
while epoch < args.max_nrof_epochs:
step = sess.run(global_step, feed_dict=None)
epoch = step // args.epoch_size
# Train for one epoch
train(args, sess, train_set, epoch, image_paths_placeholder, labels_placeholder, labels_batch,
batch_size_placeholder, learning_rate_placeholder, phase_train_placeholder, enqueue_op, input_queue, global_step,
embeddings, total_loss, train_op, summary_op, summary_writer, args.learning_rate_schedule_file,
args.embedding_size, anchor, positive, negative, triplet_loss)
# Save variables and the metagraph if it doesn't exist already
save_variables_and_metagraph(sess, saver, summary_writer, model_dir, subdir, step)
# Evaluate on LFW
if args.lfw_dir:
evaluate(sess, lfw_paths, embeddings, labels_batch, image_paths_placeholder, labels_placeholder,
batch_size_placeholder, learning_rate_placeholder, phase_train_placeholder, enqueue_op, actual_issame, args.batch_size,
args.lfw_nrof_folds, log_dir, step, summary_writer, args.embedding_size)
return model_dir
def train(args, sess, dataset, epoch, image_paths_placeholder, labels_placeholder, labels_batch,
batch_size_placeholder, learning_rate_placeholder, phase_train_placeholder, enqueue_op, input_queue, global_step,
embeddings, loss, train_op, summary_op, summary_writer, learning_rate_schedule_file,
embedding_size, anchor, positive, negative, triplet_loss):
batch_number = 0
if args.learning_rate>0.0:
lr = args.learning_rate
else:
lr = facenet.get_learning_rate_from_file(learning_rate_schedule_file, epoch)
while batch_number < args.epoch_size:
# Sample people randomly from the dataset
image_paths, num_per_class = sample_people(dataset, args.people_per_batch, args.images_per_person)
print('Running forward pass on sampled images: ', end='')
start_time = time.time()
nrof_examples = args.people_per_batch * args.images_per_person
labels_array = np.reshape(np.arange(nrof_examples),(-1,3))
image_paths_array = np.reshape(np.expand_dims(np.array(image_paths),1), (-1,3))
sess.run(enqueue_op, {image_paths_placeholder: image_paths_array, labels_placeholder: labels_array})
emb_array = np.zeros((nrof_examples, embedding_size))
nrof_batches = int(np.ceil(nrof_examples / args.batch_size))
for i in range(nrof_batches):
batch_size = min(nrof_examples-i*args.batch_size, args.batch_size)
emb, lab = sess.run([embeddings, labels_batch], feed_dict={batch_size_placeholder: batch_size,
learning_rate_placeholder: lr, phase_train_placeholder: True})
emb_array[lab,:] = emb
print('%.3f' % (time.time()-start_time))
# Select triplets based on the embeddings
print('Selecting suitable triplets for training')
triplets, nrof_random_negs, nrof_triplets = select_triplets(emb_array, num_per_class,
image_paths, args.people_per_batch, args.alpha)
selection_time = time.time() - start_time
print('(nrof_random_negs, nrof_triplets) = (%d, %d): time=%.3f seconds' %
(nrof_random_negs, nrof_triplets, selection_time))
# Perform training on the selected triplets
nrof_batches = int(np.ceil(nrof_triplets*3/args.batch_size))
triplet_paths = list(itertools.chain(*triplets))
labels_array = np.reshape(np.arange(len(triplet_paths)),(-1,3))
triplet_paths_array = np.reshape(np.expand_dims(np.array(triplet_paths),1), (-1,3))
sess.run(enqueue_op, {image_paths_placeholder: triplet_paths_array, labels_placeholder: labels_array})
nrof_examples = len(triplet_paths)
train_time = 0
i = 0
emb_array = np.zeros((nrof_examples, embedding_size))
loss_array = np.zeros((nrof_triplets,))
summary = tf.Summary()
step = 0
while i < nrof_batches:
start_time = time.time()
batch_size = min(nrof_examples-i*args.batch_size, args.batch_size)
feed_dict = {batch_size_placeholder: batch_size, learning_rate_placeholder: lr, phase_train_placeholder: True}
err, _, step, emb, lab = sess.run([loss, train_op, global_step, embeddings, labels_batch], feed_dict=feed_dict)
emb_array[lab,:] = emb
loss_array[i] = err
duration = time.time() - start_time
print('Epoch: [%d][%d/%d]tTime %.3ftLoss %2.3f' %
(epoch, batch_number 1, args.epoch_size, duration, err))
batch_number = 1
i = 1
train_time = duration
summary.value.add(tag='loss', simple_value=err)
# Add validation loss and accuracy to summary
#pylint: disable=maybe-no-member
summary.value.add(tag='time/selection', simple_value=selection_time)
summary_writer.add_summary(summary, step)
return step
def select_triplets(embeddings, nrof_images_per_class, image_paths, people_per_batch, alpha):
""" Select the triplets for training
"""
trip_idx = 0
emb_start_idx = 0
num_trips = 0
triplets = []
# VGG Face: Choosing good triplets is crucial and should strike a balance between
# selecting informative (i.e. challenging) examples and swamping training with examples that
# are too hard. This is achieve by extending each pair (a, p) to a triplet (a, p, n) by sampling
# the image n at random, but only between the ones that violate the triplet loss margin. The
# latter is a form of hard-negative mining, but it is not as aggressive (and much cheaper) than
# choosing the maximally violating example, as often done in structured output learning.
for i in xrange(people_per_batch):
nrof_images = int(nrof_images_per_class[i])
for j in xrange(1,nrof_images):
a_idx = emb_start_idx j - 1
neg_dists_sqr = np.sum(np.square(embeddings[a_idx] - embeddings), 1)
for pair in xrange(j, nrof_images): # For every possible positive pair.
p_idx = emb_start_idx pair
pos_dist_sqr = np.sum(np.square(embeddings[a_idx]-embeddings[p_idx]))
neg_dists_sqr[emb_start_idx:emb_start_idx nrof_images] = np.NaN
#all_neg = np.where(np.logical_and(neg_dists_sqr-pos_dist_sqr<alpha, pos_dist_sqr<neg_dists_sqr))[0] # FaceNet selection
all_neg = np.where(neg_dists_sqr-pos_dist_sqr<alpha)[0] # VGG Face selecction
nrof_random_negs = all_neg.shape[0]
if nrof_random_negs>0:
rnd_idx = np.random.randint(nrof_random_negs)
n_idx = all_neg[rnd_idx]
triplets.append((image_paths[a_idx], image_paths[p_idx], image_paths[n_idx]))
#print('Triplet %d: (%d, %d, %d), pos_dist=%2.6f, neg_dist=%2.6f (%d, %d, %d, %d, %d)' %
# (trip_idx, a_idx, p_idx, n_idx, pos_dist_sqr, neg_dists_sqr[n_idx], nrof_random_negs, rnd_idx, i, j, emb_start_idx))
trip_idx = 1
num_trips = 1
emb_start_idx = nrof_images
np.random.shuffle(triplets)
return triplets, num_trips, len(triplets)
def sample_people(dataset, people_per_batch, images_per_person):
nrof_images = people_per_batch * images_per_person
# Sample classes from the dataset
nrof_classes = len(dataset)
class_indices = np.arange(nrof_classes)
np.random.shuffle(class_indices)
i = 0
image_paths = []
num_per_class = []
sampled_class_indices = []
# Sample images from these classes until we have enough
while len(image_paths)<nrof_images:
class_index = class_indices[i]
nrof_images_in_class = len(dataset[class_index])
image_indices = np.arange(nrof_images_in_class)
np.random.shuffle(image_indices)
nrof_images_from_class = min(nrof_images_in_class, images_per_person, nrof_images-len(image_paths))
idx = image_indices[0:nrof_images_from_class]
image_paths_for_class = [dataset[class_index].image_paths[j] for j in idx]
sampled_class_indices = [class_index]*nrof_images_from_class
image_paths = image_paths_for_class
num_per_class.append(nrof_images_from_class)
i =1
return image_paths, num_per_class
def evaluate(sess, image_paths, embeddings, labels_batch, image_paths_placeholder, labels_placeholder,
batch_size_placeholder, learning_rate_placeholder, phase_train_placeholder, enqueue_op, actual_issame, batch_size,
nrof_folds, log_dir, step, summary_writer, embedding_size):
start_time = time.time()
# Run forward pass to calculate embeddings
print('Running forward pass on LFW images: ', end='')
nrof_images = len(actual_issame)*2
assert(len(image_paths)==nrof_images)
labels_array = np.reshape(np.arange(nrof_images),(-1,3))
image_paths_array = np.reshape(np.expand_dims(np.array(image_paths),1), (-1,3))
sess.run(enqueue_op, {image_paths_placeholder: image_paths_array, labels_placeholder: labels_array})
emb_array = np.zeros((nrof_images, embedding_size))
nrof_batches = int(np.ceil(nrof_images / batch_size))
label_check_array = np.zeros((nrof_images,))
for i in xrange(nrof_batches):
batch_size = min(nrof_images-i*batch_size, batch_size)
emb, lab = sess.run([embeddings, labels_batch], feed_dict={batch_size_placeholder: batch_size,
learning_rate_placeholder: 0.0, phase_train_placeholder: False})
emb_array[lab,:] = emb
label_check_array[lab] = 1
print('%.3f' % (time.time()-start_time))
assert(np.all(label_check_array==1))
_, _, accuracy, val, val_std, far = lfw.evaluate(emb_array, actual_issame, nrof_folds=nrof_folds)
print('Accuracy: %1.3f -%1.3f' % (np.mean(accuracy), np.std(accuracy)))
print('Validation rate: %2.5f -%2.5f @ FAR=%2.5f' % (val, val_std, far))
lfw_time = time.time() - start_time
# Add validation loss and accuracy to summary
summary = tf.Summary()
#pylint: disable=maybe-no-member
summary.value.add(tag='lfw/accuracy', simple_value=np.mean(accuracy))
summary.value.add(tag='lfw/val_rate', simple_value=val)
summary.value.add(tag='time/lfw', simple_value=lfw_time)
summary_writer.add_summary(summary, step)
with open(os.path.join(log_dir,'lfw_result.txt'),'at') as f:
f.write('%dt%.5ft%.5fn' % (step, np.mean(accuracy), val))
def save_variables_and_metagraph(sess, saver, summary_writer, model_dir, model_name, step):
# Save the model checkpoint
print('Saving variables')
start_time = time.time()
checkpoint_path = os.path.join(model_dir, 'model-%s.ckpt' % model_name)
saver.save(sess, checkpoint_path, global_step=step, write_meta_graph=False)
save_time_variables = time.time() - start_time
print('Variables saved in %.2f seconds' % save_time_variables)
metagraph_filename = os.path.join(model_dir, 'model-%s.meta' % model_name)
save_time_metagraph = 0
if not os.path.exists(metagraph_filename):
print('Saving metagraph')
start_time = time.time()
saver.export_meta_graph(metagraph_filename)
save_time_metagraph = time.time() - start_time
print('Metagraph saved in %.2f seconds' % save_time_metagraph)
summary = tf.Summary()
#pylint: disable=maybe-no-member
summary.value.add(tag='time/save_variables', simple_value=save_time_variables)
summary.value.add(tag='time/save_metagraph', simple_value=save_time_metagraph)
summary_writer.add_summary(summary, step)
def get_learning_rate_from_file(filename, epoch):
with open(filename, 'r') as f:
for line in f.readlines():
line = line.split('#', 1)[0]
if line:
par = line.strip().split(':')
e = int(par[0])
lr = float(par[1])
if e <= epoch:
learning_rate = lr
else:
return learning_rate
def parse_arguments(argv):
parser = argparse.ArgumentParser()
parser.add_argument('--logs_base_dir', type=str,
help='Directory where to write event logs.', default='~/logs/facenet')
parser.add_argument('--models_base_dir', type=str,
help='Directory where to write trained models and checkpoints.', default='~/models/facenet')
parser.add_argument('--gpu_memory_fraction', type=float,
help='Upper bound on the amount of GPU memory that will be used by the process.', default=1.0)
parser.add_argument('--pretrained_model', type=str,
help='Load a pretrained model before training starts.')
parser.add_argument('--data_dir', type=str,
help='Path to the data directory containing aligned face patches.',
default='~/datasets/casia/casia_maxpy_mtcnnalign_182_160')
parser.add_argument('--model_def', type=str,
help='Model definition. Points to a module containing the definition of the inference graph.', default='models.inception_resnet_v1')
parser.add_argument('--max_nrof_epochs', type=int,
help='Number of epochs to run.', default=500)
parser.add_argument('--batch_size', type=int,
help='Number of images to process in a batch.', default=90)
parser.add_argument('--image_size', type=int,
help='Image size (height, width) in pixels.', default=160)
parser.add_argument('--people_per_batch', type=int,
help='Number of people per batch.', default=45)
parser.add_argument('--images_per_person', type=int,
help='Number of images per person.', default=40)
parser.add_argument('--epoch_size', type=int,
help='Number of batches per epoch.', default=1000)
parser.add_argument('--alpha', type=float,
help='Positive to negative triplet distance margin.', default=0.2)
parser.add_argument('--embedding_size', type=int,
help='Dimensionality of the embedding.', default=128)
parser.add_argument('--random_crop',
help='Performs random cropping of training images. If false, the center image_size pixels from the training images are used. '
'If the size of the images in the data directory is equal to image_size no cropping is performed', action='store_true')
parser.add_argument('--random_flip',
help='Performs random horizontal flipping of training images.', action='store_true')
parser.add_argument('--keep_probability', type=float,
help='Keep probability of dropout for the fully connected layer(s).', default=1.0)
parser.add_argument('--weight_decay', type=float,
help='L2 weight regularization.', default=0.0)
parser.add_argument('--optimizer', type=str, choices=['ADAGRAD', 'ADADELTA', 'ADAM', 'RMSPROP', 'MOM'],
help='The optimization algorithm to use', default='ADAGRAD')
parser.add_argument('--learning_rate', type=float,
help='Initial learning rate. If set to a negative value a learning rate '
'schedule can be specified in the file "learning_rate_schedule.txt"', default=0.1)
parser.add_argument('--learning_rate_decay_epochs', type=int,
help='Number of epochs between learning rate decay.', default=100)
parser.add_argument('--learning_rate_decay_factor', type=float,
help='Learning rate decay factor.', default=1.0)
parser.add_argument('--moving_average_decay', type=float,
help='Exponential decay for tracking of training parameters.', default=0.9999)
parser.add_argument('--seed', type=int,
help='Random seed.', default=666)
parser.add_argument('--learning_rate_schedule_file', type=str,
help='File containing the learning rate schedule that is used when learning_rate is set to to -1.', default='data/learning_rate_schedule.txt')
# Parameters for validation on LFW
parser.add_argument('--lfw_pairs', type=str,
help='The file containing the pairs to use for validation.', default='data/pairs.txt')
parser.add_argument('--lfw_dir', type=str,
help='Path to the data directory containing aligned face patches.', default='')
parser.add_argument('--lfw_nrof_folds', type=int,
help='Number of folds to use for cross validation. Mainly used for testing.', default=10)
return parser.parse_args(argv)
if __name__ == '__main__':
main(parse_arguments(sys.argv[1:]))Tript loss简介
我这里将Triplet Loss翻译为三元组损失，其中的三元也就是如下图的Anchor、Negative、Positive，如下图所示通过Triplet Loss的学习后使得Positive元和Anchor元之间的距离最小，而和Negative之间距离最大。其中Anchor为训练数据集中随机选取的一个样本，Positive为和Anchor属于同一类的样本，而Negative则为和Anchor不同类的样本。
重要的理解说三遍
三元组损失函数使锚点和正样本之间的距离最小，两者具有相同的特性，并使锚点和负样本之间的距离最大，两者具有不同的特性。
使得同类样本的positive样本更靠近Anchor，而不同类的样本Negative则远离Anchor。
也就是类似于缩小类内距离，扩大类间距离。
图3第三部分
3.1 三元组损失函数
嵌入用f(x)∈R表示d。它将图像x嵌入到d维的欧几里得空间中。此外，我们还将这种嵌入限制在d维超球面上，即kf(x)k上2 = 1。这个损失在最近邻分类的背景下[19]的动机。在这里，我们希望确保一个特定的人的图像xai(锚)与同一个人的所有其他图像xpi(正)的距离比与任何其他人的图像xni(负)的距离更近。
公式1
公式2
公式33.2 Triplet 选择
在整个训练集上计算argmin和argmax是不可行的。此外，这可能会导致糟糕的训练，因为标签错误和图像欠佳的面孔将会占据硬阳性和阴性。有两个明显的选择可以避免这个问题:
•每n步离线生成三胞胎，使用最近的网络检查点和计算数据子集的argmin和argmax。
•在线生成三胞胎。这可以通过从一个小批量中选择硬的正面/负面范例来实现。
- Generate triplets offline every n steps, using the most recent network checkpoint and computing the argmin and argmax on a subset of the data.
- Generate triplets online. This can be done by selecting the hard positive/negative exemplars from within a mini-batch.
在实践中，选择最困难的负面因素会在训练早期导致糟糕的局部极小值，特别是它会导致一个崩溃的模型(即f(x) = 0)。为了减轻这种情况，它有助于选择这样的样本。
公式4如前所述，正确的三元组选择是快速收敛的关键。一方面，我们想使用小批量，因为这些倾向于改善收敛在随机梯度下降(SGD)[20]。另一方面，实现细节使得成批的几十到几百个范例更加有效。然而，关于批大小的主要限制是我们从小批中选择硬相关的三胞胎的方式。在大多数实验中，我们使用大约1800个样本的批量大小。
3.3 深度卷积神经网络的结构
NN1 这个表格展示了我们的Zeiler&Fergus基于文章的Visualizing and understanding convolutional networks的模型的结构，该模型受到Network in network论文的启发，具有1×1的回旋。输入和输出大小以行* cols * #过滤器来描述。内核被指定为rows×cols、stride, maxout[6]池大小为p = 2。
Visualizing and understanding convolutional networks文章地址：https://arxiv.org/abs/1311.2901
表格1第四部分 数据集和评价
Flops vs.准确性取舍。本文展示了对于各种不同的模型大小和架构，在Flops和准确性之间的权衡。突出显示的是我们在实验中关注的四个模型。
图4第五部分 实验
NN2：NN2 Inception化身的细节。这个模型几乎与[16]中描述的模型相同。两个主要区别是L的使用2 池，而不是最大池(m)，其中指定。也就是说，不是取空间最大值，而是取L2 标准计算。池总是3×3(除了最终的平均池)，并且与每个初始模块中的卷积模块并行。如果池化后降维，记为p. 1×1,3×3，然后将5×5池连接起来得到最终输出。
表2网络架构。这张图显示了4.2节中我们的个人照片测试集的四种不同模型的完整ROC。10E-4距离的急剧下降可以用groundtruth标签上的噪音来解释。模型性能排序为:NN2: 224×224输入在先启模型;NN1:基于Zeiler&Fergus的1×1卷积网络;NNS1:小型盗梦空间模式，只有2.2亿次失败;《ns2》:微型盗梦空间模型，只有20万次失败。
图5网络架构。这个表比较了我们的模型架构在保持测试集上的性能(参见4.1节)。报告的是平均验证率VAL为10E-3假接受率。另外还显示了五次测试分割的平均值的标准误差。
表3
