我正在研究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 。
然后我跟随https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/tutorials/post_training_quant.ipynb将冻结的图转换为Lite模型和量化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)
然后我跟随https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/tools/accuracy/ilsvrc使用台式机上的ImageNet验证数据集（50000张图像）评估其准确性。
但是，当我跑步
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
并检查输出accuracy_output.txt。准确性很差。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，前1位的准确度可以达到76.8，但我的尝试甚至最终都无法达到1。 为什么会这样？我哪里做错了？谢谢！
答案 0 :(得分：0)
还请检查类别标签。如果使用了错误的类别标签，结果将与您描述的一样。
答案 1 :(得分：0)
在您的Python代码中检查模型路径，它是resnet_v2_101_299_frozen_quantize.tflite，但是您在命令行中使用了另一个ResNet_V2_101.tflite