FCN8s 代码解析

FCN.py

# -×- coding: utf-8 -*-from __future__ import print_functionimport tensorflow as tfimport numpy as npimport TensorflowUtils as utilsimport read_MITSceneParsingData as scene_parsingimport datetimeimport BatchDatsetReader as datasetfrom six.moves import xrangeFLAGS = tf.flags.FLAGStf.flags.DEFINE_integer("batch_size", "2", "batch size for training")tf.flags.DEFINE_string("logs_dir", "logs/", "path to logs directory")tf.flags.DEFINE_string("data_dir", "Data_zoo/MIT_SceneParsing/", "path to dataset")tf.flags.DEFINE_float("learning_rate", "1e-4", "Learning rate for Adam Optimizer")tf.flags.DEFINE_string("model_dir", "Model_zoo/", "Path to vgg model mat")tf.flags.DEFINE_bool('debug', "False", "Debug mode: True/ False")tf.flags.DEFINE_string('mode', "train", "Mode train/ test/ visualize")MODEL_URL = 'http://www.vlfeat.org/matconvnet/models/beta16/imagenet-vgg-verydeep-19.mat'MAX_ITERATION = int(1e5 + 1)NUM_OF_CLASSESS = 151IMAGE_SIZE = 224## vgg 网络部分， weights 是vgg网络各层的权重集合， image是被预测的图像的向量def vgg_net(weights, image):    ## fcn的前五层网络就是vgg网络    layers = (        'conv1_1', 'relu1_1', 'conv1_2', 'relu1_2', 'pool1',        'conv2_1', 'relu2_1', 'conv2_2', 'relu2_2', 'pool2',        'conv3_1', 'relu3_1', 'conv3_2', 'relu3_2', 'conv3_3',        'relu3_3', 'conv3_4', 'relu3_4', 'pool3',        'conv4_1', 'relu4_1', 'conv4_2', 'relu4_2', 'conv4_3',        'relu4_3', 'conv4_4', 'relu4_4', 'pool4',        'conv5_1', 'relu5_1', 'conv5_2', 'relu5_2', 'conv5_3',        'relu5_3', 'conv5_4', 'relu5_4'    )    net = {}    current = image    for i, name in enumerate(layers):        kind = name[:4]        if kind == 'conv':            kernels, bias = weights[i][0][0][0][0]            # matconvnet: weights are [width, height, in_channels, out_channels]            # tensorflow: weights are [height, width, in_channels, out_channels]            kernels = utils.get_variable(np.transpose(kernels, (1, 0, 2, 3)), name=name + "_w")            bias = utils.get_variable(bias.reshape(-1), name=name + "_b")            current = utils.conv2d_basic(current, kernels, bias)        elif kind == 'relu':            current = tf.nn.relu(current, name=name)            if FLAGS.debug:                utils.add_activation_summary(current)        elif kind == 'pool':            ## vgg 的前5层的stride都是2，也就是前5层的size依次减小1倍            ## 这里处理了前4层的stride，用的是平均池化            ## 第5层的pool在下文的外部处理了，用的是最大池化            ## pool1 size缩小2倍            ## pool2 size缩小4倍            ## pool3 size缩小8倍            ## pool4 size缩小16倍            current = utils.avg_pool_2x2(current)  ## 平均池化        net[name] = current    return net  ## vgg每层的结果都保存再net中了## 预测流程，image是输入图像的向量，keep_prob是dropout ratedef inference(image, keep_prob):    """    Semantic segmentation network definition    ## 语义分割网络    :param image: input image. Should have values in range 0-255    :param keep_prob:    :return:    """    ## 获取训练好的vgg部分的model    print("setting up vgg initialized conv layers ...")    model_data = utils.get_model_data(FLAGS.model_dir, MODEL_URL)    mean = model_data['normalization'][0][0][0]    mean_pixel = np.mean(mean, axis=(0, 1))    weights = np.squeeze(model_data['layers'])    ## 将图像的向量值都减去平均像素值，进行 normalization    processed_image = utils.process_image(image, mean_pixel)    with tf.variable_scope("inference"):        ## 计算前5层vgg网络的输出结果        image_net = vgg_net(weights, processed_image)        conv_final_layer = image_net["conv5_3"]        ## pool1 size缩小2倍        ## pool2 size缩小4倍        ## pool3 size缩小8倍        ## pool4 size缩小16倍        ## pool5 size缩小32倍        pool5 = utils.max_pool_2x2(conv_final_layer)        ## 初始化第6层的w、b        ## 7*7 卷积核的视野很大        W6 = utils.weight_variable([7, 7, 512, 4096], name="W6")        b6 = utils.bias_variable([4096], name="b6")        conv6 = utils.conv2d_basic(pool5, W6, b6)        relu6 = tf.nn.relu(conv6, name="relu6")        if FLAGS.debug:            utils.add_activation_summary(relu6)        relu_dropout6 = tf.nn.dropout(relu6, keep_prob=keep_prob)        ## 在第6层没有进行池化，所以经过第6层后 size缩小仍为32倍        ## 初始化第7层的w、b        W7 = utils.weight_variable([1, 1, 4096, 4096], name="W7")        b7 = utils.bias_variable([4096], name="b7")        conv7 = utils.conv2d_basic(relu_dropout6, W7, b7)        relu7 = tf.nn.relu(conv7, name="relu7")        if FLAGS.debug:            utils.add_activation_summary(relu7)        relu_dropout7 = tf.nn.dropout(relu7, keep_prob=keep_prob)        ## 在第7层没有进行池化，所以经过第7层后 size缩小仍为32倍        ## 初始化第8层的w、b        ## 输出维度为NUM_OF_CLASSESS        W8 = utils.weight_variable([1, 1, 4096, NUM_OF_CLASSESS], name="W8")        b8 = utils.bias_variable([NUM_OF_CLASSESS], name="b8")        conv8 = utils.conv2d_basic(relu_dropout7, W8, b8)        # annotation_pred1 = tf.argmax(conv8, dimension=3, name="prediction1")        # now to upscale to actual image size        ## 开始将size提升为图像原始尺寸        deconv_shape1 = image_net["pool4"].get_shape()        W_t1 = utils.weight_variable([4, 4, deconv_shape1[3].value, NUM_OF_CLASSESS], name="W_t1")        b_t1 = utils.bias_variable([deconv_shape1[3].value], name="b_t1")        ## 对第8层的结果进行反卷积(上采样),通道数也由NUM_OF_CLASSESS变为第4层的通道数        conv_t1 = utils.conv2d_transpose_strided(conv8, W_t1, b_t1, output_shape=tf.shape(image_net["pool4"]))        ## 对应论文原文中的"2× upsampled prediction + pool4 prediction"        fuse_1 = tf.add(conv_t1, image_net["pool4"], name="fuse_1")        deconv_shape2 = image_net["pool3"].get_shape()        ## 对上一层上采样的结果进行反卷积(上采样),通道数也由上一层的通道数变为第3层的通道数        W_t2 = utils.weight_variable([4, 4, deconv_shape2[3].value, deconv_shape1[3].value], name="W_t2")        b_t2 = utils.bias_variable([deconv_shape2[3].value], name="b_t2")        conv_t2 = utils.conv2d_transpose_strided(fuse_1, W_t2, b_t2, output_shape=tf.shape(image_net["pool3"]))        ## 对应论文原文中的"2× upsampled prediction + pool3 prediction"        fuse_2 = tf.add(conv_t2, image_net["pool3"], name="fuse_2")        ## 原始图像的height、width和通道数        shape = tf.shape(image)        ## 既形成一个列表，形式为[height, width, in_channels, out_channels]        deconv_shape3 = tf.stack([shape[0], shape[1], shape[2], NUM_OF_CLASSESS])        W_t3 = utils.weight_variable([16, 16, NUM_OF_CLASSESS, deconv_shape2[3].value], name="W_t3")        b_t3 = utils.bias_variable([NUM_OF_CLASSESS], name="b_t3")        ## 再进行一次反卷积，将上一层的结果转化为和原始图像相同size、通道数为分类数的形式数据        conv_t3 = utils.conv2d_transpose_strided(fuse_2, W_t3, b_t3, output_shape=deconv_shape3, stride=8)        ## 目前conv_t3的形式为size为和原始图像相同的size，通道数与分类数相同        ## 这句我的理解是对于每个像素位置，根据第3维度（通道数）通过argmax能计算出这个像素点属于哪个分类        ## 也就是对于每个像素而言，NUM_OF_CLASSESS个通道中哪个数值最大，这个像素就属于哪个分类        annotation_pred = tf.argmax(conv_t3, dimension=3, name="prediction")    return tf.expand_dims(annotation_pred, dim=3), conv_t3## 训练def train(loss_val, var_list):    optimizer = tf.train.AdamOptimizer(FLAGS.learning_rate)    ## 下面是参照tf api    ## Compute gradients of loss_val for the variables in var_list.    ## This is the first part of minimize().    ## loss: A Tensor containing the value to minimize.    ## var_list: Optional list of tf.Variable to update to minimize loss.    ##   Defaults to the list of variables collected in the graph under the key GraphKey.TRAINABLE_VARIABLES.    grads = optimizer.compute_gradients(loss_val, var_list=var_list)    if FLAGS.debug:        # print(len(var_list))        for grad, var in grads:            utils.add_gradient_summary(grad, var)    ## 下面是参照tf api    ## Apply gradients to variables.    ## This is the second part of minimize(). It returns an Operation that applies gradients.    return optimizer.apply_gradients(grads)def main(argv=None):    ## dropout 的保留率    keep_probability = tf.placeholder(tf.float32, name="keep_probabilty")    ## 原始图像的向量    image = tf.placeholder(tf.float32, shape=[None, IMAGE_SIZE, IMAGE_SIZE, 3], name="input_image")    ## 原始图像对应的标注图像的向量    annotation = tf.placeholder(tf.int32, shape=[None, IMAGE_SIZE, IMAGE_SIZE, 1], name="annotation")    ## 输入原始图像向量、保留率，得到预测的标注图像和随后一层的网络输出    pred_annotation, logits = inference(image, keep_probability)    ## 为了方便查看图像预处理的效果，可以利用 TensorFlow 提供的 tensorboard 工具进行可视化，直接用 tf.summary.image 将图像写入 summary    tf.summary.image("input_image", image, max_outputs=2)    tf.summary.image("ground_truth", tf.cast(annotation, tf.uint8), max_outputs=2)    tf.summary.image("pred_annotation", tf.cast(pred_annotation, tf.uint8), max_outputs=2)    ## 计算预测标注图像和真实标注图像的交叉熵    loss = tf.reduce_mean((tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits,                                                                          labels=tf.squeeze(annotation, squeeze_dims=[3]),                                                                          name="entropy")))    tf.summary.scalar("entropy", loss)    ## 返回需要训练的变量列表    trainable_var = tf.trainable_variables()    if FLAGS.debug:        for var in trainable_var:            utils.add_to_regularization_and_summary(var)    ## 定义损失    train_op = train(loss, trainable_var)    print("Setting up summary op...")    ## 定义合并变量操作，一次性生成所有摘要数据    summary_op = tf.summary.merge_all()    print("Setting up image reader...")    ## 读取训练数据集、验证数据集    train_records, valid_records = scene_parsing.read_dataset(FLAGS.data_dir)    print(len(train_records))    print(len(valid_records))    print("Setting up dataset reader")    ## 将训练数据集、验证数据集的格式转换为网络需要的格式    image_options = {'resize': True, 'resize_size': IMAGE_SIZE}    if FLAGS.mode == 'train':        train_dataset_reader = dataset.BatchDatset(train_records, image_options)    validation_dataset_reader = dataset.BatchDatset(valid_records, image_options)    sess = tf.Session()    print("Setting up Saver...")    saver = tf.train.Saver()    summary_writer = tf.summary.FileWriter(FLAGS.logs_dir, sess.graph)    sess.run(tf.global_variables_initializer())    ## 加载之前的checkpoint    ckpt = tf.train.get_checkpoint_state(FLAGS.logs_dir)    if ckpt and ckpt.model_checkpoint_path:        saver.restore(sess, ckpt.model_checkpoint_path)        print("Model restored...")    if FLAGS.mode == "train":        for itr in xrange(MAX_ITERATION):            ## 读取训练集的一个batch            train_images, train_annotations = train_dataset_reader.next_batch(FLAGS.batch_size)            feed_dict = {image: train_images, annotation: train_annotations, keep_probability: 0.85}            ## 执行计算损失操作，网络跑起来了            sess.run(train_op, feed_dict=feed_dict)            if itr % 10 == 0:                train_loss, summary_str = sess.run([loss, summary_op], feed_dict=feed_dict)                print("Step: %d, Train_loss:%g" % (itr, train_loss))                summary_writer.add_summary(summary_str, itr)            if itr % 500 == 0:                valid_images, valid_annotations = validation_dataset_reader.next_batch(FLAGS.batch_size)                valid_loss = sess.run(loss, feed_dict={image: valid_images, annotation: valid_annotations,                                                       keep_probability: 1.0})                print("%s ---> Validation_loss: %g" % (datetime.datetime.now(), valid_loss))                saver.save(sess, FLAGS.logs_dir + "model.ckpt", itr)    elif FLAGS.mode == "visualize":        valid_images, valid_annotations = validation_dataset_reader.get_random_batch(FLAGS.batch_size)        pred = sess.run(pred_annotation, feed_dict={image: valid_images, annotation: valid_annotations,                                                    keep_probability: 1.0})        valid_annotations = np.squeeze(valid_annotations, axis=3)        pred = np.squeeze(pred, axis=3)        for itr in range(FLAGS.batch_size):            utils.save_image(valid_images[itr].astype(np.uint8), FLAGS.logs_dir, name="inp_" + str(5+itr))            utils.save_image(valid_annotations[itr].astype(np.uint8), FLAGS.logs_dir, name="gt_" + str(5+itr))            utils.save_image(pred[itr].astype(np.uint8), FLAGS.logs_dir, name="pred_" + str(5+itr))            print("Saved image: %d" % itr)if __name__ == "__main__":    tf.app.run()