deeplabv3架构_Deeplabv3+源码记录

Img_raw&＃61;tf.gfile.FastGFile(dir,&＃39;rb&＃39;).read() #dir具体到每张图片的地址&＃xff0c;得到的图片类型是bytes&＃xff0c;就不用转换类型了
# 变形记开始啦
Writer&＃61;tf.python_io.TFRecordWriter(output_dir) #output_dir是转换后文件地址&＃xff0c;文件后缀可以直接是.tfrecord
# 将得到的图片数据转换成example protocol buffer
Example&＃61;tf.train.Example(features&＃61;tf.train.Features(feature&＃61;{&＃39;img_raw&＃39;:_bytes_feature(Img_raw)
}))
Writer.write(Example.SerializeToString()) #将信息写入这个数据结构
Writer.close()

def create_tf_record(output_filename,image_dir,label_dir,examples):"""Creates a TFRecord file from examples.Args:output_filename: Path to where output file is saved.image_dir: Directory where image files are stored.label_dir: Directory where label files are stored.examples: Examples to parse and save to tf record."""# 创建一个类writerwriter &＃61; tf.python_io.TFRecordWriter(output_filename)for idx, example in enumerate(examples):if idx % 500 &＃61;&＃61; 0:tf.logging.info(&＃39;On image %d of %d&＃39;, idx, len(examples))# 得到图片image和label的具体地址image_path &＃61; os.path.join(image_dir, example &＃43; &＃39;.png&＃39;)label_path &＃61; os.path.join(label_dir, example &＃43; &＃39;.png&＃39;)if not os.path.exists(image_path):tf.logging.warning(&＃39;Could not find %s, ignoring example.&＃39;, image_path)continueelif not os.path.exists(label_path):tf.logging.warning(&＃39;Could not find %s, ignoring example.&＃39;, label_path)continue# 将两个地址都送入dict_to_tf_example中&＃xff0c;得到example对象try:tf_example &＃61; dict_to_tf_example(image_path, label_path)writer.write(tf_example.SerializeToString()) # 将对象写入文件地址中except ValueError:tf.logging.warning(&＃39;Invalid example: %s, ignoring.&＃39;, example)writer.close()# 将image和label都转换为TFrecord文件
def dict_to_tf_example(image_path, label_path):#以rb读二进制的方式打开图片所在地址with tf.gfile.GFile(image_path, &＃39;rb&＃39;) as fid:encoded_jpg &＃61; fid.read() # 读取图片内容encoded_jpg_io &＃61; io.BytesIO(encoded_jpg) # 创建一个临时二进制文件image &＃61; PIL.Image.open(encoded_jpg_io) # 读取这个二进制文件&＃xff0c;但是返回的不是numpy数据&＃xff0c;但可以通过numpy.array进行数据转换if image.format !&＃61; &＃39;png&＃39;:raise ValueError(&＃39;Image format not PNG&＃39;)# 同上&＃xff0c;对label数据进行转换。with tf.gfile.GFile(label_path, &＃39;rb&＃39;) as fid:encoded_label &＃61; fid.read()encoded_label_io &＃61; io.BytesIO(encoded_label)label &＃61; PIL.Image.open(encoded_label_io)if label.format !&＃61; &＃39;PNG&＃39;:raise ValueError(&＃39;Label format not PNG&＃39;)if image.size !&＃61; label.size:raise ValueError(&＃39;The size of image does not match with that of label.&＃39;)# 获得图片的宽高&＃xff08;480,360&＃xff09;width, height &＃61; image.size# 创建Example对象&＃xff0c;每一个example都有下面这些feature# bytes_feature和int64_feature是TF数据的两种类型&＃xff0c;即字符串型和int64example &＃61; tf.train.Example(features&＃61;tf.train.Features(feature&＃61;{&＃39;image/height&＃39;: dataset_util.int64_feature(height),&＃39;image/width&＃39;: dataset_util.int64_feature(width),&＃39;image/encoded&＃39;: dataset_util.bytes_feature(encoded_jpg),&＃39;image/format&＃39;: dataset_util.bytes_feature(&＃39;png&＃39;.encode(&＃39;utf8&＃39;)),&＃39;label/encoded&＃39;: dataset_util.bytes_feature(encoded_label),&＃39;label/format&＃39;: dataset_util.bytes_feature(&＃39;png&＃39;.encode(&＃39;utf8&＃39;)),}))return example

def main(unused_argv):# Using the Winograd non-fused algorithms provides a small performance boost.os.environ[&＃39;TF_ENABLE_WINOGRAD_NONFUSED&＃39;] &＃61; &＃39;1&＃39;if FLAGS.clean_model_dir:shutil.rmtree(FLAGS.model_dir, ignore_errors&＃61;True)#1.创建RunConfig来更改checkpoint的时间run_config&＃61;tf.estimator.RunConfig().replace(save_checkpoints_secs&＃61;1e9)#2.实例化Estimator,model_fn得到的是resnet-101的网络架构&＃xff0c;model_dir是预训练好的模型参数&＃xff0c;params参数会自动传送给deeplabv3_plus_model_fnmodel&＃61;tf.estimator.Estimator(model_fn&＃61;deeplab_model.deeplabv3_plus_model_fn, model_dir&＃61;FLAGS.model_dir,config&＃61;run_config, params&＃61;{
&＃39;output_stride&＃39;:FLAGS.output_stride,
&＃39;batch_size&＃39;:FLAGS.batch_size,
&＃39;base_architecture&＃39;:FLAGS.base_architecture,
&＃39;pre_trained_model&＃39;:FLAGS.pre_trained_model,
&＃39;batch_norm_decay&＃39;:_BATCH_NORM_DECAY,
&＃39;num_classes&＃39;:_NUM_CLASSES,
&＃39;tensorboard_images_max_outputs&＃39;:FLAGS.tensorboard_images_max_outputs,
&＃39;weight_decay&＃39;:FLAGS.weight_decay,
&＃39;learning_rate_policy&＃39;:FLAGS.learning_rate_policy,
&＃39;num_train&＃39;:_NUM_IMAGES[&＃39;train&＃39;],
&＃39;initial_learning_rate&＃39;:FLAGS.initial_learning_rate,
&＃39;max_iter&＃39;:FLAGS.max_iter,
&＃39;end_learning_rate&＃39;:FLAGS.end_learning_rate,
&＃39;power&＃39;:_POWER,
&＃39;momentum&＃39;:_MOMENTUM,
&＃39;freeze_batch_norm&＃39;:FLAGS.freeze_batch_norm,
&＃39;initial_global_step&＃39;:FLAGS.initial_global_step
})
# 会被打印出来的内容包括学习率&＃xff0c;交叉熵&＃xff0c;像素点的准确度和mIOU
for_inrange(FLAGS.train_epochs//FLAGS.epochs_per_eval):tensors_to_log&＃61;{&＃39;learning_rate&＃39;:&＃39;learning_rate&＃39;,&＃39;cross_entropy&＃39;:&＃39;cross_entropy&＃39;,&＃39;train_px_accuracy&＃39;:&＃39;train_px_accuracy&＃39;,&＃39;train_mean_iou&＃39;:&＃39;train_mean_iou&＃39;,}
# 设置每迭代10次就会打印一次
logging_hook&＃61;tf.train.LoggingTensorHook(
tensors&＃61;tensors_to_log,every_n_iter&＃61;10)
train_hooks&＃61;[logging_hook]
eval_hooks&＃61;None
# 调用TF的调试器
if FLAGS.debug:debug_hook&＃61;tf_debug.LocalCLIDebugHook()train_hooks.append(debug_hook)eval_hooks&＃61;[debug_hook]# 3.开始训练模型&＃xff0c;函数input_fn作为数据输入的来源
tf.logging.info("Starttraining.")
model.train(input_fn&＃61;lambda:input_fn(True,FLAGS.data_dir,FLAGS.batch_size,FLAGS.epochs_per_eval),hooks&＃61;train_hooks,)# 4.开始进行模型评估&＃xff0c;函数input_fn作为数据输入来源
tf.logging.info("Startevaluation.")
eval_results&＃61;model.evaluate(input_fn&＃61;lambda:input_fn(False,FLAGS.data_dir,1),hooks&＃61;eval_hooks,)
print(eval_results)

def input_fn(is_training,data_dir,batch_size,num_epochs&＃61;1):
# 1.切片处理&＃xff0c;输入数据&＃xff0c;在第一个维度内进行切片&＃xff01;&＃xff01;&＃xff01;!
dataset&＃61;tf.data.Dataset.from_tensor_slices(get_filenames(is_training,data_dir)) # 得到的是TFrecord的地址
# print(dataset.output_shapes)
#print(dataset.types) # 可以查看dataset里面的数据类型
dataset&＃61;dataset.flat_map(tf.data.TFRecordDataset) # 解析tfrecord文件的每一条记录&＃xff0c;序列化后#tf.train.Example&＃xff0c;解析函数是parse_record中的parse_single_example&＃xff08;&＃xff09;# 2.shuffle()操作&＃xff1a;随机化处理
ifis_training:dataset&＃61;dataset.shuffle(buffer_size&＃61;_NUM_IMAGES[&＃39;train&＃39;]) # 训练时&＃xff0c;将输入数据的顺序进行打乱dataset&＃61;dataset.map(parse_record) # 指定parse_record方法对数据进行改变&＃xff0c;返回的是image和label
dataset&＃61;dataset.map(lambdaimage,label:preprocess_image(image,label,is_training))
dataset&＃61;dataset.prefetch(batch_size)
# repeat操作
dataset&＃61;dataset.repeat(num_epochs) # 指定重复的次数
dataset&＃61;dataset.batch(batch_size)
# 3.创建迭代器
iterator&＃61;dataset.make_one_shot_iterator()
images,labels&＃61;iterator.get_next()
Return images,labels

def parse_record(raw_record):
keys_to_features&＃61;{
&＃39;image/height&＃39;:
tf.FixedLenFeature((),tf.int64),
&＃39;image/width&＃39;:
tf.FixedLenFeature((),tf.int64),
&＃39;image/encoded&＃39;:
tf.FixedLenFeature((),tf.string,default_value&＃61;&＃39;&＃39;),
&＃39;image/format&＃39;:
tf.FixedLenFeature((),tf.string,default_value&＃61;&＃39;png&＃39;),
&＃39;label/encoded&＃39;:
tf.FixedLenFeature((),tf.string,default_value&＃61;&＃39;&＃39;),
&＃39;label/format&＃39;:
tf.FixedLenFeature((),tf.string,default_value&＃61;&＃39;png&＃39;),
}
# 解析每一条记录
parsed&＃61;tf.parse_single_example(raw_record,keys_to_features) # 解析record的每条记录#height&＃61;tf.cast(parsed[&＃39;image/height&＃39;],tf.int32) # 进行类型转换&＃xff01;&＃xff01;&＃xff01;
#width&＃61;tf.cast(parsed[&＃39;image/width&＃39;],tf.int32)image&＃61;tf.image.decode_image(tf.reshape(parsed[&＃39;image/encoded&＃39;],shape&＃61;[]) ,_DEPTH)
image&＃61;tf.to_float(tf.image.convert_image_dtype(image,dtype&＃61;tf.uint8))
image.set_shape([None,None,3])label&＃61;tf.image.decode_image(
tf.reshape(parsed[&＃39;label/encoded&＃39;],shape&＃61;[]),1)
label&＃61;tf.to_int32(tf.image.convert_image_dtype(label,dtype&＃61;tf.uint8))
label.set_shape([None,None,1])return image,label

def preprocess_image(image,label,is_training):
"""Preprocessa single image of layout [height,width,depth]."""
ifis_training:
#Randomly scale thei mage and label.
image,label&＃61;preprocessing.random_rescale_image_and_label(image,label,_MIN_SCALE,_MAX_SCALE)#Randomly crop or pad a [_HEIGHT,_WIDTH] section of the image and label.
image,label&＃61;preprocessing.random_crop_or_pad_image_and_label(
image,label,_HEIGHT,_WIDTH,_IGNORE_LABEL)#Randomly flip the image and label horizontally.
image,label&＃61;preprocessing.random_flip_left_right_image_and_label(
image,label)image.set_shape([_HEIGHT,_WIDTH,3])
label.set_shape([_HEIGHT,_WIDTH,1])image&＃61;preprocessing.mean_image_subtraction(image)return image,label