参考链接：
AlexNet论文
csdn博客：AlexNet网络结构解析
Caffe AlexNet实现
ImageNet数据集下载：ILSVRC2012
ILSVRC2012数据集label信息下载
ImageNet数据集处理by Caffe
caffe针对imagenet相关文件详解
kratzert AlexNet TensorFlow开源实现
kratzert AlexNet TensorFlow开源实现解析
借鉴以上链接：从图像导入以及AlexNet网络结构搭建以及测试，完整过程如下：

版权声明：本程序修改自https://github.com/kratzert/finetune_alexnet_with_tensorflow/tree/5d751d62eb4d7149f4e3fd465febf8f07d4cea9d
。如果要应用本程序代码，请务必在评论区留言征得本人同意！

ImageNet图像处理

制作ImageNet2012数据集，并进行预处理
图像的均值文件下载地址：https://github.com/BVLC/caffe/blob/master/python/caffe/imagenet/ilsvrc_2012_mean.npy

datagenerator.py:

"""Containes a helper class for image input pipelines in tensorflow."""import tensorflow as tf
import numpy as np
import osfrom tensorflow.python.framework import dtypes
from tensorflow.python.framework.ops import convert_to_tensorimage_mean = np.load('ilsvrc_2012_mean.npy')
image_mean = np.swapaxes(image_mean, 0, 2)
image_mean = image_mean[:,:,::-1]class ImageDataGenerator(object):"""Wrapper class around the new Tensorflows dataset pipeline."""def __init__(self, txt_file, mode, imgfile_folder, batch_size, num_classes, shuffle=True,buffer_size=1000):self.txt_file = txt_fileself.imgfile_folder = imgfile_folderself.num_classes = num_classes# retrieve the data from the text fileself._read_txt_file()# number of samples in the datasetself.data_size = len(self.labels)# initial shuffling of the file and label lists (together!)if shuffle:self._shuffle_lists()# convert lists to TF tensorself.img_paths = convert_to_tensor(self.img_paths, dtype=dtypes.string)self.labels = convert_to_tensor(self.labels, dtype=dtypes.int32)# create datasetdata = tf.data.Dataset.from_tensor_slices((self.img_paths, self.labels))# distinguish between train/infer. when calling the parsing functionsif mode == 'training':# param: output_buffer_size = 100 * batch_sizedata = data.map(self._parse_function_train, num_parallel_calls=7)elif mode == 'inference':# param: output_buffer_size = 100 * batch_sizedata = data.map(self._parse_function_inference, num_parallel_calls=7)else:raise ValueError("Invalid mode '%s'." % (mode))# shuffle the first `buffer_size` elements of the datasetif shuffle:data = data.shuffle(buffer_size=buffer_size)# create a new dataset with batches of imagesdata = data.batch(batch_size)self.data = datadef _read_txt_file(self, ):"""Read the content of the text file and store it into lists."""cur_path = os.path.join(os.path.split(os.getcwd())[0], self.imgfile_folder)self.img_paths = []self.labels = []with open(self.txt_file, 'r') as f:lines = f.readlines()for line in lines:items = line.split(' ')self.img_paths.append(os.path.join(cur_path, items[0]))self.labels.append(int(items[1]))def _shuffle_lists(self):"""Conjoined shuffling of the list of paths and labels."""path = self.img_pathslabels = self.labelspermutation = np.random.permutation(self.data_size)self.img_paths = []self.labels = []for i in permutation:self.img_paths.append(path[i])self.labels.append(labels[i])def _parse_function_train(self, filename, label):"""Input parser for samples of the training set."""# convert label number into one-hot-encoding# one_hot = tf.one_hot(label, self.num_classes)img_string = tf.read_file(filename)img_decoded = tf.image.decode_png(img_string, channels=3)img_resize = tf.image.resize_images(img_decoded, [256, 256], method=3)  # 缩放尺寸为[256, 256]img_resize = tf.subtract(img_resize, image_mean)  # 减去平均值img_resized = tf.random_crop(img_resize, [227, 227, 3])  # 随机裁切为[227,227,3]# RGB -> BGRimg_bgr = img_resized[:, :, ::-1]return img_bgr, labeldef _parse_function_inference(self, filename, label):"""Input parser for samples of the validation/test set."""# convert label number into one-hot-encoding# one_hot = tf.one_hot(label, self.num_classes)img_string = tf.read_file(filename)img_decoded = tf.image.decode_png(img_string, channels=3)img_resize = tf.image.resize_images(img_decoded, [256,256], method=3)img_resize = tf.subtract(img_resize, image_mean)img_resized = tf.image.resize_image_with_crop_or_pad(img_resize, 227, 227)  # 中心裁剪[227,227]img_resized = tf.cast(img_resized, tf.float32)# RGB -> BGRimg_bgr = img_resized[:, :, ::-1]return img_bgr, label

#AlexNet网络结构定义为类

alexnet.py:

import tensorflow as tf
import numpy as npclass AlexNet(object):def __init__(self, x, keep_prob6, keep_prob7, num_classes, skip_layer, weights_path):# Parse input arguments into class variablesself.X = xself.NUM_CLASSES = num_classesself.KEEP_PROB6 = keep_prob6self.KEEP_PROB7 = keep_prob7self.SKIP_LAYER = skip_layerself.WEIGHTS_PATH = weights_path# Call the create function to build the computational graph of AlexNetself.create()def create(self):# 1st Layer: Conv (w ReLu) -> Lrn -> Poolconv1 = conv(self.X, 11, 11, 96, 4, 4, padding='VALID', name='conv1')norm1 = lrn(conv1, 2, 2e-05, 0.75, name='norm1')pool1 = max_pool(norm1, 3, 3, 2, 2, padding='VALID', name='pool1')# 2nd Layer: Conv (w ReLu)  -> Lrn -> Pool with 2 groupsconv2 = conv(pool1, 5, 5, 256, 1, 1, groups=2, name='conv2')norm2 = lrn(conv2, 2, 2e-05, 0.75, name='norm2')pool2 = max_pool(norm2, 3, 3, 2, 2, padding='VALID', name='pool2')# 3rd Layer: Conv (w ReLu)conv3 = conv(pool2, 3, 3, 384, 1, 1, name='conv3')# 4th Layer: Conv (w ReLu) splitted into two groupsconv4 = conv(conv3, 3, 3, 384, 1, 1, groups=2, name='conv4')# 5th Layer: Conv (w ReLu) -> Pool splitted into two groupsconv5 = conv(conv4, 3, 3, 256, 1, 1, groups=2, name='conv5')pool5 = max_pool(conv5, 3, 3, 2, 2, padding='VALID', name='pool5')# 6th Layer: Flatten -> FC (w ReLu) -> Dropoutflattened = tf.reshape(pool5, [-1, 6*6*256])fc6 = fc(flattened, 6*6*256, 4096, name='fc6')dropout6 = dropout(fc6, self.KEEP_PROB6)# 7th Layer: FC (w ReLu) -> Dropoutfc7 = fc(dropout6, 4096, 4096, name='fc7')dropout7 = dropout(fc7, self.KEEP_PROB7)# 8th Layer: FC and return unscaled activationsself.fc8 = fc(dropout7, 4096, self.NUM_CLASSES, relu=False, name='fc8')def load_initial_weights(self, session):weights = np.load(self.WEIGHTS_PATH)weights_dict = {}for k in weights.keys():layer_name = k.split('_')[0]if layer_name not in weights_dict.keys():weights_dict[layer_name] = []weights_dict[layer_name].append(weights[k].copy())# Loop over all layer names stored in the weights dictfor op_name in weights_dict:with tf.variable_scope(op_name, reuse=True):# Assign weights/biases to their corresponding tf variablefor data in weights_dict[op_name]:# Biasesif len(data.shape) == 1:var = tf.get_variable('biases')update = var.assign(data)session.run(update)# Weightselse:var = tf.get_variable('weights')update = var.assign(data)session.run(update)# load params into computational graphdef load_params(self, session, param_dict):for i,v in enumerate(tf.trainable_variables()):v.load(param_dict[i], session)# get params into computational graphdef get_params(self, session):layer = ['conv1', 'conv2', 'conv3', 'conv4', 'conv5', 'fc6', 'fc7', 'fc8']param = []for op_id in range(len(layer)):with tf.variable_scope(layer[op_id], reuse=True):varw = tf.get_variable('weights')param.append(session.run(varw))varb = tf.get_variable('biases')param.append(session.run(varb))return paramdef conv(x, filter_height, filter_width, num_filters, stride_y, stride_x, name,padding='SAME', groups=1):# Get number of input channelsinput_channels = int(x.get_shape()[-1])# Create lambda function for the convolutionconvolve = lambda i, k: tf.nn.conv2d(i, k,strides=[1, stride_y, stride_x, 1],padding=padding)with tf.variable_scope(name) as scope:# Create tf variables for the weights and biases of the conv layerweights = tf.get_variable('weights', shape=[filter_height,filter_width,input_channels/groups,num_filters])biases = tf.get_variable('biases', shape=[num_filters])l2_loss_conv = tf.multiply(tf.nn.l2_loss(weights), 0.001)tf.add_to_collection('losses', l2_loss_conv)if groups == 1:conv = convolve(x, weights)# In the cases of multiple groups, split inputs & weights andelse:# 拿conv2举例：# 权重由(5,5,48,256)分解为(5,5,48,128)和(5,5,48,128)# 数据由(1,27,27,96)分解为(1,27,27,48)和(1,27,27,48)# Split input and weights and convolve them separatelyinput_groups = tf.split(axis=3, num_or_size_splits=groups, value=x)weight_groups = tf.split(axis=3, num_or_size_splits=groups, value=weights)output_groups = [convolve(i, k) for i, k in zip(input_groups, weight_groups)]# Concat the convolved output together againconv = tf.concat(axis=3, values=output_groups)# Add biasesbias = tf.reshape(tf.nn.bias_add(conv, biases), tf.shape(conv))# Apply relu functionrelu = tf.nn.relu(bias, name=scope.name)return reludef fc(x, num_in, num_out, name, relu=True):"""Create a fully connected layer."""with tf.variable_scope(name) as scope:# Create tf variables for the weights and biasesweights = tf.get_variable('weights', shape=[num_in, num_out],trainable=True)biases = tf.get_variable('biases', [num_out], trainable=True)l2_loss_fc = tf.multiply(tf.nn.l2_loss(weights), 0.001)tf.add_to_collection('losses', l2_loss_fc)# Matrix multiply weights and inputs and add biasact = tf.nn.xw_plus_b(x, weights, biases, name=scope.name)if relu:relu = tf.nn.relu(act)return reluelse:return actdef max_pool(x, filter_height, filter_width, stride_y, stride_x, name, padding='SAME'):return tf.nn.max_pool(x, ksize=[1, filter_height, filter_width, 1],strides=[1, stride_y, stride_x, 1],padding=padding, name=name)def lrn(x, radius, alpha, beta, name, bias=1.0):return tf.nn.local_response_normalization(x, depth_radius=radius,alpha=alpha, beta=beta,bias=bias, name=name)def dropout(x, keep_prob):return tf.nn.dropout(x, keep_prob)

#运行和测试AlexNet
在images文件夹下，有3张图片：

图像数据的路径和label统一放在train.txt和val.txt中，形式为：

images/llama.jpeg 355
images/sealion.jpeg 150
images/zebra.jpeg 340

其中序号为该图片在ImageNet中的编号，完整的标注文件可从本文开头链接下载。
http://www.cs.toronto.edu/~guerzhoy/tf_alexnet/在此链接中可以下载AlexNet网络结构参数，在本程序中默认调用load_initial_weights方法就可以将参数导入，不需要训练。
在如下代码中，因为直接load进入参数，所以将训练操作注释掉，直接进行训练操作。

finetune.py

"""Script to finetune AlexNet using Tensorflow.With this script you can finetune AlexNet as provided in the alexnet.py
class on any given dataset. Specify the configuration settings at the
beginning according to your problem.
This script was written for TensorFlow >= version 1.2rc0 and comes with a blog
post, which you can find here:https://kratzert.github.io/2017/02/24/finetuning-alexnet-with-tensorflow.htmlAuthor: Frederik Kratzert
contact: f.kratzert(at)gmail.com
"""import osimport numpy as np
import tensorflow as tffrom alexnet import AlexNet
from datagenerator import ImageDataGenerator
from datetime import datetime"""
Configuration Part.
"""# Path to the textfiles for the trainings and validation set
train_file = '/path/to/train.txt'
val_file = 'val.txt'# Learning params
learning_rate = 0.01
num_epochs = 10
batch_size = 1# Network params
dropout_rate = 0.5
num_classes = 1000
train_layers = []# How often we want to write the tf.summary data to disk
display_step = 20"""
Main Part of the finetuning Script.
"""# Place data loading and preprocessing on the cpu
with tf.device('/cpu:0'):'''tr_data = ImageDataGenerator(train_file,mode='training',batch_size=batch_size,num_classes=num_classes,shuffle=True)'''val_data = ImageDataGenerator(val_file,mode='inference',batch_size=batch_size,num_classes=num_classes,shuffle=False)# create an reinitializable iterator given the dataset structureiterator = tf.data.Iterator.from_structure(val_data.data.output_types,val_data.data.output_shapes)next_batch = iterator.get_next()# Ops for initializing the two different iterators
# training_init_op = iterator.make_initializer(tr_data.data)
validation_init_op = iterator.make_initializer(val_data.data)# TF placeholder for graph input and output
x = tf.placeholder(tf.float32, [batch_size, 227, 227, 3])
y = tf.placeholder(tf.int32, [batch_size])
keep_prob = tf.placeholder(tf.float32)# Initialize model
model = AlexNet(x, keep_prob, num_classes, train_layers)# Link variable to model output
score = model.fc8
'''
# List of trainable variables of the layers we want to train
var_list = [v for v in tf.trainable_variables() if v.name.split('/')[0] in train_layers]# Op for calculating the loss
with tf.name_scope("cross_ent"):loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=score, labels=y))# Train op
with tf.name_scope("train"):# Get gradients of all trainable variablesgradients = tf.gradients(loss, var_list)gradients = list(zip(gradients, var_list))# Create optimizer and apply gradient descent to the trainable variablesoptimizer = tf.train.GradientDescentOptimizer(learning_rate)train_op = optimizer.apply_gradients(grads_and_vars=gradients)
'''# Evaluation op: Accuracy of the model
with tf.name_scope("accuracy"):accuracy_1 = tf.nn.in_top_k(score, y, 1)  # Top1 Accuracyaccuracy_5 = tf.nn.in_top_k(score, y, 5)  # Top5 Accuracy# Get the number of training/validation steps per epoch
# train_batches_per_epoch = int(np.floor(tr_data.data_size/batch_size))
val_batches_per_epoch = int(np.floor(val_data.data_size / batch_size))# Start Tensorflow session
with tf.Session() as sess:# Initialize all variablessess.run(tf.global_variables_initializer())# Load the pretrained weights into the non-trainable layermodel.load_initial_weights(sess)# print("{} Start training...".format(datetime.now()))# Loop over number of epochsfor epoch in range(num_epochs):print("{} Epoch number: {}".format(datetime.now(), epoch+1))'''# Initialize iterator with the training datasetsess.run(training_init_op)for step in range(train_batches_per_epoch):# get next batch of dataimg_batch, label_batch = sess.run(next_batch)# And run the training opsess.run(train_op, feed_dict={x: img_batch,y: label_batch,keep_prob: dropout_rate})'''# Validate the model on the entire validation setprint("{} Start validation".format(datetime.now()))sess.run(validation_init_op)test_acc1 = 0test_acc5 = 0test_count = 0for _ in range(val_batches_per_epoch):img_batch, label_batch = sess.run(next_batch)acc1, acc5 = sess.run([accuracy_1, accuracy_5],feed_dict={x: img_batch,y: label_batch,keep_prob: 1.})test_acc1 += np.sum(acc1)test_acc5 += np.sum(acc5)test_count += batch_sizetest_acc1 /= test_counttest_acc5 /= test_countprint("Validation Top1/Top5 Accuracy = {:.4f}/{:.4f}".format(test_acc1, test_acc5))

输出测试结果的Top1和Top5错误率：

2018-06-09 20:40:49.905062 Epoch number: 1
2018-06-09 20:40:49.905062 Start validation
Validation Top1/Top5 Accuracy = 1.0000/1.0000
2018-06-09 20:40:50.619143 Epoch number: 2
2018-06-09 20:40:50.619143 Start validation
Validation Top1/Top5 Accuracy = 1.0000/1.0000
......

Tensorflow实现AlexNet相关推荐

【Tensorflow】深度学习实战03——Tensorflow实现AlexNet
[fishing-pan:https://blog.csdn.net/u013921430转载请注明出处] 前言前两篇博文中分别利用卷积神经网络识别手写数字和对CIFAR-10数据集分类,在这两次的 ...
教你用TensorFlow搭建AlexNet
勿满足于眼前的小小成就.你要问自己,我这辈子只有这样吗. AlexNet模式是在2012年提出来的,并且在当年获得了ILSVRC的冠军,top-5的错误率为16.4%,比第二名的成绩26.2%高出许多 ...
Tensorflow fine-tunning AlexNet
包括Tensorflow下的 AlexNet 网络模型搭建.测试.使用自己的数据做 fine-tunning.批量测试训练结果主要参考文章<Finetuning AlexNet with Te ...
TensorFlow实战-AlexNet
1 # 导入数据 2 from tensorflow.examples.tutorials.mnist import input_data 3 # 读取数据 4 mnist=input_data.re ...
TensorFlow MNIST AlexNet
原始的AlexNet用来处理277*277*3的数据集, 并且采用5层卷积,3层全连接层来处理图像分类. 具体结构和参数信息见 http://blog.csdn.net/chenhaifeng2016 ...
利用tensorflow构建AlexNet模型，实现小数量级的猫狗分类（只有train）
首先看路径: data文件夹分为,model文件夹,train文件夹和文件夹,model文件夹存放模型文件,train存放cat和dog的两个文件夹图片, validation和train一样.con ...
TensorFlow实战5——TensorFlow实现AlexNet
1 #coding = utf-8 2 3 from datetime import datetime 4 import math 5 import time 6 import tensorflow ...
【深度学习】使用tensorflow实现AlexNet
转载注明出处:http://blog.csdn.net/accepthjp/article/details/69999309 AlexNet是2012年ImageNet比赛的冠军,虽然过去了很长时间, ...
用Tensorflow实现AlexNet识别猫狗数据集（猫狗大战）【附代码】
AlexNet识别猫狗数据集一.下载猫狗数据集二.AlexNet实现 1.划分训练集和测试集 2.将训练集和测试集图片放缩为224x224 3.AlexNet实现 4.训练过程 5.模型测试三. ...

Tensorflow实现AlexNet

ImageNet图像处理

Tensorflow实现AlexNet相关推荐

最新文章

热门文章