tensorflow - Tensorflow Lite：ResNet 示例模型在使用 ImageNet 进行验证期间给出了非常差的结果

Question

我正在学习 tensorflow lite。我从https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/g3doc/models.md#image-classification-float-models下载了 ResNet 冻结图ResNet_V2_101。

然后我按照https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/tutorials/post_training_quant.ipynb将此冻结图转换为 Lite 模型和量化的 lite 模型。

import tensorflow as tf
import pathlib
import sys
import tensorflow as tf
from tensorflow.python.saved_model import tag_constants
import time
graph_def_file = "resnet_saved_model/resnet_v2_101_299_frozen.pb"
input_arrays = ["input"]
output_arrays = ["output"]
converter = tf.lite.TocoConverter.from_frozen_graph(str(graph_def_file),input_arrays,output_arrays,input_shapes = {"input":[1,299,299,3]})
tflite_model = converter.convert()
open("saved_model/resnet_v2_101_299_frozen.tflite", "wb").write(tflite_model) 

converter.post_training_quantize = True
tflite_quantized_model = converter.convert()
open("saved_model/resnet_v2_101_299_frozen_quantize.tflite", "wb").write(tflite_quantized_model) 

然后我按照https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/tools/accuracy/ilsvrc在我的桌面上使用 ImageNet 验证数据集（50000 张图像）评估其准确性。

但是，当我跑步时

bazel run -c opt   --cxxopt='--std=c++11'   --   //tensorflow/lite/tools/accuracy/ilsvrc:imagenet_accuracy_eval   --model_file="/home/kathy/saved_model/ResNet_V2_101.tflite"   --ground_truth_images_path="/media/kathy/Documents/val_imgs"   --ground_truth_labels="/home/kathy/workspace/tensorflow/tensorflow/lite/tools/accuracy/ilsvrc/VALIDATION_LABELS.txt"   --model_output_labels="/home/kathy/workspace/tensorflow/tensorflow/lite/tools/accuracy/ilsvrc/resnet_output_labels.txt"   --output_file_path="/tmp/accuracy_output.txt" --num_images=0

并检查了输出accuracy_output.txt。准确性非常差。我可以在 50000 张图像中捕获一些结果。

Top 1, Top 2, Top 3, Top 4, Top 5, Top 6, Top 7, Top 8, Top 9, Top 10
0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000
0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000
0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000
0.000, 0.000, 0.000, 25.000, 25.000, 25.000, 25.000, 25.000, 25.000, 25.000
0.000, 0.000, 0.000, 20.000, 20.000, 20.000, 20.000, 20.000, 20.000, 20.000
0.000, 0.000, 0.000, 16.667, 16.667, 16.667, 16.667, 16.667, 16.667, 16.667
0.000, 0.000, 0.000, 14.286, 14.286, 14.286, 14.286, 14.286, 14.286, 14.286
0.000, 0.000, 0.000, 12.500, 12.500, 12.500, 12.500, 12.500, 12.500, 12.500
0.000, 0.000, 0.000, 11.111, 11.111, 11.111, 11.111, 11.111, 11.111, 11.111
0.000, 0.000, 0.000, 10.000, 10.000, 10.000, 10.000, 10.000, 10.000, 10.000
0.000, 0.000, 0.000, 9.091, 9.091, 9.091, 9.091, 9.091, 9.091, 9.091
0.000, 0.000, 0.000, 8.333, 8.333, 8.333, 8.333, 8.333, 8.333, 8.333
0.000, 0.000, 0.000, 7.692, 7.692, 7.692, 7.692, 7.692, 7.692, 7.692
0.000, 0.000, 0.000, 7.143, 7.143, 7.143, 7.143, 7.143, 7.143, 7.143
0.000, 0.000, 0.000, 6.667, 6.667, 6.667, 6.667, 6.667, 6.667, 6.667
0.000, 0.000, 0.000, 6.250, 6.250, 6.250, 6.250, 6.250, 6.250, 6.250
0.000, 0.000, 0.000, 5.882, 5.882, 5.882, 5.882, 5.882, 5.882, 5.882
0.000, 0.000, 0.000, 5.556, 5.556, 5.556, 5.556, 5.556, 5.556, 5.556
0.000, 0.000, 0.000, 5.263, 5.263, 5.263, 5.263, 5.263, 5.263, 5.263
0.000, 0.000, 0.000, 5.000, 5.000, 5.000, 5.000, 5.000, 5.000, 5.000
0.000, 0.000, 0.000, 4.762, 4.762, 4.762, 4.762, 4.762, 4.762, 4.762
0.000, 0.000, 0.000, 4.545, 4.545, 4.545, 4.545, 4.545, 4.545, 4.545
0.000, 0.000, 0.000, 4.348, 4.348, 4.348, 4.348, 4.348, 4.348, 4.348
0.000, 0.000, 0.000, 4.167, 4.167, 4.167, 4.167, 4.167, 4.167, 4.167
0.000, 0.000, 0.000, 4.000, 4.000, 4.000, 4.000, 4.000, 4.000, 4.000
0.000, 0.000, 0.000, 3.846, 3.846, 3.846, 3.846, 3.846, 3.846, 3.846
0.000, 0.000, 0.000, 3.704, 3.704, 3.704, 3.704, 3.704, 3.704, 3.704
0.000, 0.000, 0.000, 3.571, 3.571, 3.571, 3.571, 3.571, 3.571, 3.571
0.000, 0.000, 0.000, 3.448, 3.448, 3.448, 3.448, 3.448, 3.448, 3.448
0.000, 0.000, 0.000, 3.333, 3.333, 3.333, 3.333, 3.333, 3.333, 3.333
0.000, 0.000, 0.000, 3.226, 3.226, 3.226, 3.226, 3.226, 3.226, 3.226
0.000, 0.000, 0.000, 3.125, 3.125, 3.125, 3.125, 3.125, 3.125, 3.125
0.000, 0.000, 0.000, 3.030, 3.030, 3.030, 3.030, 3.030, 3.030, 3.030
0.000, 0.000, 0.000, 2.941, 2.941, 2.941, 2.941, 2.941, 2.941, 2.941
0.000, 0.000, 0.000, 2.857, 2.857, 2.857, 2.857, 2.857, 2.857, 2.857
0.000, 0.000, 0.000, 2.778, 2.778, 2.778, 2.778, 2.778, 2.778, 2.778
0.000, 0.000, 0.000, 2.703, 2.703, 2.703, 2.703, 2.703, 2.703, 2.703
0.000, 0.000, 0.000, 2.632, 2.632, 2.632, 2.632, 2.632, 2.632, 2.632
0.000, 0.000, 0.000, 2.564, 2.564, 2.564, 2.564, 2.564, 2.564, 2.564
0.000, 0.000, 0.000, 2.500, 2.500, 2.500, 2.500, 2.500, 2.500, 2.500
0.000, 0.000, 0.000, 2.439, 2.439, 2.439, 2.439, 2.439, 2.439, 2.439
0.000, 0.000, 0.000, 2.381, 2.381, 2.381, 2.381, 2.381, 2.381, 2.381
0.000, 0.000, 0.000, 2.326, 2.326, 2.326, 2.326, 2.326, 2.326, 2.326
0.000, 0.000, 0.000, 2.273, 2.273, 2.273, 2.273, 2.273, 2.273, 2.273
0.000, 0.000, 0.000, 2.222, 2.222, 2.222, 2.222, 2.222, 2.222, 2.222
0.000, 0.000, 0.000, 2.174, 2.174, 2.174, 2.174, 2.174, 2.174, 2.174
0.000, 0.000, 0.000, 2.128, 2.128, 2.128, 2.128, 2.128, 2.128, 2.128
0.000, 0.000, 0.000, 2.083, 2.083, 2.083, 2.083, 2.083, 2.083, 2.083
0.000, 0.000, 0.000, 2.041, 2.041, 2.041, 2.041, 2.041, 2.041, 2.041
0.000, 0.000, 0.000, 2.000, 2.000, 2.000, 2.000, 2.000, 2.000, 2.000
0.000, 0.000, 0.000, 1.961, 1.961, 1.961, 1.961, 1.961, 1.961, 1.961
0.000, 0.000, 0.000, 1.923, 1.923, 1.923, 1.923, 1.923, 1.923, 1.923
0.000, 0.000, 0.000, 1.887, 1.887, 1.887, 1.887, 1.887, 1.887, 1.887

但是，根据https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/tutorials/post_training_quant.ipynb，top-1的准确率可以达到 76.8，但我的尝试最终甚至达不到 1。为什么会发生这种情况？我哪里做错了？谢谢！

标签： tensorflowtensorflow-literesnetimagenet

Answer 1

另请检查您的类别标签。如果使用了错误的类别标签，结果将如您所述。