TensorFlow：凍結圖形后顯着的精度損失？

Question

在凍結服務圖表后，是否常常看到嚴重的准確性損失？ 在使用預訓練的初始-resnet-v2訓練和評估花數據集期間，我的准確度為98-99％，正確預測的概率為90 +％。 然而，在凍結我的圖表並再次預測之后，我的模型不那么准確，只有30-40％的置信度才能預測正確的標簽。

模型訓練后，我有幾個項目：

1. 檢查點文件
2. model.ckpt.index文件
3. model.ckpt.meta文件
4. model.ckpt文件
5. graph.pbtxt文件。

由於我無法在GitHub上運行位於tensorflow存儲庫中的官方凍結圖文件（我想這是因為我在訓練后有一個pbtxt文件而不是pb文件），我正在重用本教程中的代碼。

這是我修改為凍結圖表的代碼：

import os, argparse

import tensorflow as tf
from tensorflow.python.framework import graph_util

dir = os.path.dirname(os.path.realpath(__file__))

def freeze_graph(model_folder, input_checkpoint):
    # We retrieve our checkpoint fullpath
    checkpoint = tf.train.get_checkpoint_state(model_folder)
    # input_checkpoint = checkpoint.model_checkpoint_path

    # We precise the file fullname of our freezed graph
    absolute_model_folder = "/".join(input_checkpoint.split('/')[:-1])
    output_graph = absolute_model_folder + "/frozen_model.pb"

    # Before exporting our graph, we need to precise what is our output node
    # This is how TF decides what part of the Graph he has to keep and what part it can dump
    # NOTE: this variable is plural, because you can have multiple output nodes
    output_node_names = "InceptionResnetV2/Logits/Predictions"

    # We clear devices to allow TensorFlow to control on which device it will load operations
    clear_devices = True

    # We import the meta graph and retrieve a Saver
    saver = tf.train.import_meta_graph(input_checkpoint + '.meta', clear_devices=clear_devices)

    # We retrieve the protobuf graph definition
    graph = tf.get_default_graph()
    input_graph_def = graph.as_graph_def()

    # We start a session and restore the graph weights
    with tf.Session() as sess:
        saver.restore(sess, input_checkpoint)

        # We use a built-in TF helper to export variables to constants
        output_graph_def = graph_util.convert_variables_to_constants(
            sess, # The session is used to retrieve the weights
            input_graph_def, # The graph_def is used to retrieve the nodes 
            output_node_names.split(",") # The output node names are used to select the usefull nodes
        ) 

        # Finally we serialize and dump the output graph to the filesystem
        with tf.gfile.GFile(output_graph, "wb") as f:
            f.write(output_graph_def.SerializeToString())
        print("%d ops in the final graph." % len(output_graph_def.node))


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument("--model_folder", type=str, help="Model folder to export")
    parser.add_argument("--input_checkpoint", type = str, help = "Input checkpoint name")
    args = parser.parse_args()

    freeze_graph(args.model_folder, args.input_checkpoint)

這是我用來運行我的預測的代碼，我只根據用戶的意圖輸入一個圖像：

import tensorflow as tf
from scipy.misc import imread, imresize
import numpy as np

img = imread("./dandelion.jpg")
img = imresize(img, (299,299,3))
img = img.astype(np.float32)
img = np.expand_dims(img, 0)

labels_dict = {0:'daisy', 1:'dandelion',2:'roses', 3:'sunflowers', 4:'tulips'}

#Define the filename of the frozen graph
graph_filename = "./frozen_model.pb"

#Create a graph def object to read the graph
with tf.gfile.GFile(graph_filename, "rb") as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())

#Construct the graph and import the graph from graphdef
with tf.Graph().as_default() as graph:
    tf.import_graph_def(graph_def)

    #We define the input and output node we will feed in
    input_node = graph.get_tensor_by_name('import/batch:0')
    output_node = graph.get_tensor_by_name('import/InceptionResnetV2/Logits/Predictions:0')

    with tf.Session() as sess:
        predictions = sess.run(output_node, feed_dict = {input_node: img})
        print predictions
        label_predicted = np.argmax(predictions[0])

    print 'Predicted Flower:', labels_dict[label_predicted]
    print 'Prediction probability:', predictions[0][label_predicted]

我通過運行我的預測收到的輸出：

2017-04-11 17:38:21.722217: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:901] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2017-04-11 17:38:21.722608: I tensorflow/core/common_runtime/gpu/gpu_device.cc:887] Found device 0 with properties: 
name: GeForce GTX 860M
major: 5 minor: 0 memoryClockRate (GHz) 1.0195
pciBusID 0000:01:00.0
Total memory: 3.95GiB
Free memory: 3.42GiB
2017-04-11 17:38:21.722624: I tensorflow/core/common_runtime/gpu/gpu_device.cc:908] DMA: 0 
2017-04-11 17:38:21.722630: I tensorflow/core/common_runtime/gpu/gpu_device.cc:918] 0:   Y 
2017-04-11 17:38:21.722642: I tensorflow/core/common_runtime/gpu/gpu_device.cc:977] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GTX 860M, pci bus id: 0000:01:00.0)
2017-04-11 17:38:22.183204: I tensorflow/compiler/xla/service/platform_util.cc:58] platform CUDA present with 1 visible devices
2017-04-11 17:38:22.183232: I tensorflow/compiler/xla/service/platform_util.cc:58] platform Host present with 8 visible devices
2017-04-11 17:38:22.184007: I tensorflow/compiler/xla/service/service.cc:183] XLA service 0xb85a1c0 executing computations on platform Host. Devices:
2017-04-11 17:38:22.184022: I tensorflow/compiler/xla/service/service.cc:191]   StreamExecutor device (0): <undefined>, <undefined>
2017-04-11 17:38:22.184140: I tensorflow/compiler/xla/service/platform_util.cc:58] platform CUDA present with 1 visible devices
2017-04-11 17:38:22.184149: I tensorflow/compiler/xla/service/platform_util.cc:58] platform Host present with 8 visible devices
2017-04-11 17:38:22.184610: I tensorflow/compiler/xla/service/service.cc:183] XLA service 0xb631ee0 executing computations on platform CUDA. Devices:
2017-04-11 17:38:22.184620: I tensorflow/compiler/xla/service/service.cc:191]   StreamExecutor device (0): GeForce GTX 860M, Compute Capability 5.0
[[ 0.1670652   0.46482906  0.12899996  0.12481128  0.11429448]]
Predicted Flower: dandelion
Prediction probability: 0.464829

潛在的問題來源：我首先使用TF 0.12訓練我的模型，但我相信它與我現在使用的版本Tf 1.01兼容。 作為安全預防措施，我將文件升級到TF 1.01並重新訓練模型以獲取新的檢查點文件集（具有相同的准確性），然后使用這些檢查點文件進行凍結。 我從源代碼編譯了我的張量流。 問題來自於我使用pbtxt文件而不是pb文件的事實嗎？ 我不知道如何通過訓練我的模型得到一個pb文件。

Answer 1

我認為問題與凍結模型無關。 相反，它與您預處理圖像的方式有關。

我建議您使用InceptionResnet V2中的默認預處理功能。

下面，我將發布一個采用圖像路徑（JPG或PNG）並返回預處理圖像的代碼。 您可以修改它以使其接收一批圖像。 它不是專業代碼。 它需要一些優化。 但是，它運作良好。

一，加載圖片：

def load_img(path_img):
    """
    Load an image to tensorflow
    :param path_img: image path on the disk
    :return: 3D tensorflow image
    """
    filename_queue = tf.train.string_input_producer([path_img])  # list of files to read

    reader = tf.WholeFileReader()
    key, value = reader.read(filename_queue)

    my_img = tf.image.decode_image(value)  # use png or jpg decoder based on your files.

    init_op = tf.global_variables_initializer()
    with tf.Session() as sess:
        sess.run(init_op)

        # Start populating the filename queue.

        coord = tf.train.Coordinator()
        threads = tf.train.start_queue_runners(coord=coord)

        for i in range(1):  # length of your filename list
            image = my_img.eval()  # here is your image Tensor :)

        print(image.shape)
        # Image.fromarray(np.asarray(image)).show()

        coord.request_stop()
        coord.join(threads)

        return image

然后，預處理代碼：

def preprocess(image, height, width,
               central_fraction=0.875, scope=None):
    """Prepare one image for evaluation.

    If height and width are specified it would output an image with that size by
    applying resize_bilinear.

    If central_fraction is specified it would cropt the central fraction of the
    input image.

    Args:
      image: 3-D Tensor of image. If dtype is tf.float32 then the range should be
        [0, 1], otherwise it would converted to tf.float32 assuming that the range
        is [0, MAX], where MAX is largest positive representable number for
        int(8/16/32) data type (see `tf.image.convert_image_dtype` for details)
      height: integer
      width: integer
      central_fraction: Optional Float, fraction of the image to crop.
      scope: Optional scope for name_scope.
    Returns:
      3-D float Tensor of prepared image.
    """

    image = tf.image.convert_image_dtype(image, dtype=tf.float32)
    # Crop the central region of the image with an area containing 87.5% of
    # the original image.
    if central_fraction:
        image = tf.image.central_crop(image, central_fraction=central_fraction)

    if height and width:
        # Resize the image to the specified height and width.
        image = tf.expand_dims(image, 0)
        image = tf.image.resize_bilinear(image, [height, width],
                                         align_corners=False)
        image = tf.squeeze(image, [0])
    image = tf.subtract(image, 0.5)
    image = tf.multiply(image, 2.0)
    return image

最后，對於我的情況，我不得不將處理后的張量轉換為numpy數組：

image = tf.Session().run(image)

因此，該圖像可以饋送到凍結模型

persistent_sess = tf.Session(graph=graph)  # , config=sess_config)

    input_node = graph.get_tensor_by_name('prefix/batch:0')
    output_node = graph.get_tensor_by_name('prefix/InceptionResnetV2/Logits/Predictions:0')

    predictions = persistent_sess.run(output_node, feed_dict={input_node: [image]})
    print(predictions)
    label_predicted = np.argmax(predictions[0])
    print(label_predicted)

Answer 2

我有類似的問題，使用冷凍模型時准確度降低了1.5％。 問題在於凍結模型的代碼中的保護程序對象。 您需要將移動平均值衰減作為參數傳遞給保護程序。 我使用Inception模型中的代碼，這就是我在凍結腳本中創建保護程序的方法：

variable_averages = tf.train.ExponentialMovingAverage(0.9997)
variables_to_restore = variable_averages.variables_to_restore()
saver = tf.train.Saver(variables_to_restore)

對我來說，它解決了這個問題。