將Caffe train.txt轉換爲Tensorflow

Question

我正在嘗試使我的Caffe代碼適應tensorflow。我想知道什麼是轉換我的train.txt和test.txt以便爲tensorflow工作的最佳方法。

在我train.txt，文件看起來像：

/folder/filename1.jpg 1 

/folder/filename2.jpg 2 
... 

第一列是圖像的名稱，第二列是類標籤

謝謝！

Answer 1

我假設你想獲得一批具有數字標籤的相同大小的圖像。我們將使用tf.decode_csv()分析文本，tf.read_file()加載JPEG數據作爲一個字符串，tf.image.decode_jpeg()解析它變成一個緻密的張量，最後tf.train.batch()的解析數據建成一批圖像。許多這些功能都有很多選項可供配置，因此請參閱文檔以獲取更多定製細節。

# Set options here for whether to repeat, etc. 
filename_producer = tf.string_input_producer(["train.txt"], ...) 

# Read lines from the file, one at a time. 
line_reader = tf.TextLineReader() 
next_line = line_reader.read(filename_producer) 

# Parse line into a filename and an integer label. 
image_filename, label = tf.decode_csv(
    next_line, [tf.constant([], dtype=tf.string), tf.constant([], dtype=tf.int32)], 
    field_delim=" ") 

# Read the image as a string. 
image_bytes = tf.read_file(image_filename) 

# Convert the image into a 3-D tensor (height x width x channels). 
image_tensor = tf.image.decode_jpeg(image_bytes, ...) 

# OPTIONAL: Resize your image to a standard size if they are not already. 
HEIGHT = ... 
WIDTH = ... 
image_tensor = tf.image.resize_image_with_crop_or_pad(image_tensor, HEIGHT, WIDTH) 

# Create a batch of images. 
BATCH_SIZE = 32 
images, labels = tf.train.batch([image_tensor, label], BATCH_SIZE, ...) 

# [...build the rest of your model...] 

本示例大量使用TensorFlow預取來加載示例。該TensorFlow文檔有一個how-to that explains how to use the prefetching feature，但是要注意的最重要的事情是，你必須在你的會話開始打電話tf.train.start_queue_runners()開始預取：

sess = tf.Session() 

# You must execute this statement to begin prefetching data. 
tf.train.start_queue_runners(sess)