tensorflow实现AIchallenger scene classification baseline

主要是为了写出pipeline，至于模型啥的都可以自己改的，有什么问题可以评论，不多说了，放代码

#!/usr/bin/env python
# -*- coding: utf-8 -*-
# @Time    : 2018/4/7 16:18
# @Author  : Zehan Song
# @Site    :
# @File    : cnn.py
# @Software: PyCharm

# set your config class
import tensorflow as tf
import warnings
import os
import time
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
os.environ['CUDA_VISIBLE_DEVICES'] = '1'
#默认为0：输出所有log信息
#设置为1：进一步屏蔽INFO信息
#设置为2：进一步屏蔽WARNING信息
#设置为3：进一步屏蔽ERROR信息
class Config(object):
    # data_path parameters
    class_path = "/home/szh/TIP/scene_classes.csv"
    data_path = "/home/szh/AIchallenger/baseline/final/"
    train_data_json = "/home/szh/TIP/scene_train_annotations_20170904.json"
    val_data_json = "/home/szh/TIP/scene_validation_annotations_20170908.json"
    #train parameters
    lr = 0.001
    batch_size = 64
    max_epoch = 50
    #evaluation parameters
    save_path = "/home/szh/TIP/model_path"
    evaluation_period = 10

def parse(self,kwargs):
    for k,v in kwargs.items():
        if not hasattr(self,k):
            warnings.warn("your config doesn't have that key(%s)" %(k))
        else:
            setattr(self,k,v)
    print('user config:')
    print('#################################')
    for k in dir(self):
        if not k.startswith('_') and k != 'parse' and k != 'state_dict':
            print k, getattr(self, k)
    print('#################################')
    return self

def state_dict(self):
    return {k: getattr(self, k) for k in dir(self) if not k.startswith('_') and k != 'parse' and k != 'state_dict'}

Config.parse = parse
Config.state_dict = state_dict
opt = Config()

#set your batcher
import json
import pandas as pd
import numpy as np
import random
from PIL import Image
import cv2
class_names = pd.read_csv(opt.class_path,header=None)
number_ids = class_names[0]
chinese_names = class_names[1]
english_names = class_names[2]


class Batcher(object):
    def __init__(self,train,data_path,json_path,batch_size=opt.batch_size):
        with open(json_path,'r') as f:
            data = json.load(f)
        self.train = train
        self.batch_size = batch_size
        self.all_img_paths = [os.path.join(opt.data_path,d["image_id"]) for d in data]
        self.all_label_ids = [d["label_id"] for d in data]
        self.start = 0
        self.data_index = list(range(len(self.all_img_paths)))
        if self.train:
            random.shuffle(self.data_index)
    def reset(self):
        self.start = 0
        if self.train:
            random.shuffle(self.data_index)
    def img_resize(self, imgpath, img_size):
        # resize the image to the specific size
        img = Image.open(imgpath)

        if (img.width > img.height):
            scale = float(img_size) / float(img.height)
            img = np.array(cv2.resize(np.array(img), (
            int(img.width * scale + 1), img_size))).astype(np.float32)
        else:
            scale = float(img_size) / float(img.width)
            img = np.array(cv2.resize(np.array(img), (
            img_size, int(img.height * scale + 1)))).astype(np.float32)
        # crop the proper size and scale to [-1, 1]
        img = (img[
                  (img.shape[0] - img_size) // 2:
                  (img.shape[0] - img_size) // 2 + img_size,
                  (img.shape[1] - img_size) // 2:
                  (img.shape[1] - img_size) // 2 + img_size,
                  :]-127)/255
        return img
    def get_data(self):
        # begin = time.time()
        if not self.train:
            self.batch_size = len(self.all_img_paths)

        # config = tf.ConfigProto()
        # config.gpu_options.allow_growth = True
        # config.allow_soft_placement = True
        # sess = tf.Session(config=config)
        image_batch = []
        label_batch = []
        if self.start>len(self.all_img_paths):
            self.reset()
        self.end = min(self.start+self.batch_size,len(self.all_img_paths))
        for i in range(self.start,self.end):
            img = self.img_resize(self.all_img_paths[i],128)
            # img = tf.image.convert_image_dtype(img,dtype=tf.float32)
            # img = tf.image.resize_images(img,[128,128])
            # if self.train:
            #     img = tf.image.random_flip_up_down(img)
            #     img = tf.image.random_flip_left_right(img)
            #     img = tf.image.random_brightness(img,0.1)
            # img = sess.run(img)
            #u can find other methods to implement data augmentation like some predefined functions
            #of cv2. I don't want to use functions wrt tf because a session would be neccessary adding
            #extra much more overhead of time

            image_batch.append(img)
            label_batch.append(self.all_label_ids[i])
        self.start += self.batch_size
        # print("start")
        # print self.start
        # sess.close()
        # end = time.time()
        # print("use time")
        # print(end-begin)
        # print np.array(image_batch).size
        return np.array(image_batch),np.array(label_batch)


#set your model
class Model(object):
    def weight_variable(self,shape, stddev=0.1):
        initial = tf.truncated_normal(shape, stddev=stddev)
        return tf.Variable(initial)
    def bias_variable(self,shape, bais=0.1):
        initial = tf.constant(bais, shape=shape)
        return tf.Variable(initial)
    def conv2d(self,x, w):
        return tf.nn.conv2d(x, w, [1, 1, 1, 1], 'SAME')
    def max_pool_2x2(self,x):
        return tf.nn.max_pool(x, [1, 2, 2, 1], [1, 2, 2, 1], 'SAME')
    def max_pool_3x3(self,x):
        return tf.nn.max_pool(x, [1, 3, 3, 1], [1, 2, 2, 1], 'SAME')
    def avg_pool_3x3(self,x):
        return tf.nn.avg_pool(x, [1, 3, 3, 1], [1, 2, 2, 1], 'SAME')
    def bulid_model(self):
        # network structure
        # conv1
        # self.train = tf.placeholder(tf.bool)
        self.features = tf.placeholder(dtype=tf.float32, shape=(None,128, 128, 3))
        # self.features = tf.cond(self.train,lambda: self.preprocess(self.rawimage),lambda: self.nopreprocess(self.rawimage))
        # self.features = tf.placeholder(dtype=tf.float32,shape=(None,128,128,3),name='features-input')
        self.labels = tf.placeholder(dtype=tf.int32,shape=(None),name='labels-input')
        # self.labels = tf.one_hot(indices=self.labels,depth=80)
        W_conv1 = self.weight_variable([5, 5, 3, 64], stddev=1e-4)
        b_conv1 = self.bias_variable([64])
        h_conv1 = tf.nn.relu(self.conv2d(self.features, W_conv1) + b_conv1)
        h_pool1 = self.max_pool_3x3(h_conv1)
        # norm1
        norm1 = tf.nn.lrn(h_pool1, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75, name='norm1')
        # conv2
        W_conv2 = self.weight_variable([5, 5, 64, 64], stddev=1e-2)
        b_conv2 = self.bias_variable([64])
        h_conv2 = tf.nn.relu(self.conv2d(norm1, W_conv2) + b_conv2)
        # norm2
        norm2 = tf.nn.lrn(h_conv2, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75, name='norm2')
        h_pool2 = self.max_pool_3x3(norm2)
        # conv3
        W_conv3 = self.weight_variable([5, 5, 64, 64], stddev=1e-2)
        b_conv3 = self.bias_variable([64])
        h_conv3 = tf.nn.relu(self.conv2d(h_pool2, W_conv3) + b_conv3)
        h_pool3 = self.max_pool_3x3(h_conv3)
        # fc1
        W_fc1 = self.weight_variable([16 * 16 * 64, 128])
        b_fc1 = self.bias_variable([128])
        h_pool3_flat = tf.reshape(h_pool3, [-1, 16 * 16 * 64])
        h_fc1 = tf.nn.relu(tf.matmul(h_pool3_flat, W_fc1) + b_fc1)
        # introduce dropout
        self.keep_prob = tf.placeholder("float")
        h_fc1_drop = tf.nn.dropout(h_fc1, self.keep_prob)
        # fc2
        W_fc2 = self.weight_variable([128, 80])
        b_fc2 = self.bias_variable([80])
        self.y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2

#set train and validation process
from tqdm import tqdm
class Trainer(object):
    def __init__(self,train_batcher,val_batcher,model):
        self.train_batcher = train_batcher
        self.val_batcher = val_batcher
        self.model = model
        self.model.bulid_model()
        self.saver = tf.train.Saver()

    def train(self):
        config = tf.ConfigProto()
        config.gpu_options.allow_growth = True
        config.allow_soft_placement = True
        self.sess = tf.Session(config=config)
        cross_entropy_mean = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits\
                                                (labels=self.model.labels,logits=self.model.y_conv))
        variables = tf.trainable_variables()
        regularization_cost = tf.reduce_sum([tf.nn.l2_loss(v) for v in variables])
        loss = cross_entropy_mean + regularization_cost
        # one_hot_labels = tf.one_hot(indices=self.model.labels,depth=80,dtype=tf.int32)
        accuracy = tf.reduce_mean(tf.cast(tf.nn.in_top_k\
                    (predictions=self.model.y_conv,\
                     targets=self.model.labels,k=1)\
                           ,tf.float32))
        optimizer = tf.train.AdamOptimizer(opt.lr)
        train_op = optimizer.minimize(loss)
        TRAINING_STEPS = len(self.train_batcher.all_img_paths)/opt.batch_size
        self.sess.run(tf.global_variables_initializer())
        best_accuracy = 0
        for epoch in range(opt.max_epoch):
            total_loss = 0
            total_accuracy = 0
            for i in tqdm(range(TRAINING_STEPS)):
                # begin = time.time()
                imgs,labels = self.train_batcher.get_data()
                feed = {self.model.features:imgs,self.model.labels:labels,\
                        self.model.keep_prob:0.5}
                _,loss_value,accuracy_value = self.sess.run([train_op,loss,accuracy],feed_dict=feed)
                total_loss += loss_value
                total_accuracy += accuracy_value
                # end = time.time()
                # print("train use time")
                # print(end-begin)
                if i%10 == 0:
                    print("iter[%d/%d]:loss=%.6f,accuracy=%.6f" % (i+1, TRAINING_STEPS, \
                                                                    loss_value, \
                                                                    accuracy_value))
            print("epoch[%d/%d]:loss=%.6f,accuracy=%.6f"%(epoch+1,opt.max_epoch,\
                                                          total_loss/TRAINING_STEPS,\
                                                          total_accuracy/TRAINING_STEPS))
            if epoch%opt.evaluation_period==0:
                imgs,labels = self.val_batcher.get_data()
                feed = {self.model.features: imgs, self.model.labels: labels, \
                        self.model.keep_prob: 1.0}
                loss_value, accuracy_value = self.sess.run([loss, accuracy], feed_dict=feed)
                print("val_loss = %.6f, val_accuracy = %.6f"%\
                      (loss_value,accuracy_value))
                if best_accuracy<accuracy_value:
                    best_accuracy = accuracy_value
                    self.saver.save(self.sess,"bestmodel.ckpt")

        '''if u have the ground truths of test dataset,the u can write like that'''
        # img, labels = self.test_batcher.get_data()
        # feed = {self.model.features: img, self.model.labels: labels, \
        #         self.model.keep_prob: 1.0}
        # loss_value, accuracy_value = self.sess.run([loss, accuracy], feed_dict=feed)
        self.sess.close()

def main(**kwargs):
    opt.parse(kwargs)
    train_batcher = Batcher(train=True, data_path=opt.data_path, json_path=opt.train_data_json, \
                            batch_size=opt.batch_size)
    val_batcher = Batcher(train=False, data_path=opt.data_path, json_path=opt.val_data_json)
    model = Model()
    trainer = Trainer(train_batcher,val_batcher,model)
    trainer.train()

if __name__ == "__main__":
    import fire
    fire.Fire()

运行命令如下：

python cnn.py main --batch-size=32 --lr=0.001

后面的参数可以自己改