Я изучаю tenorflow lite.Я скачал замороженный график ResNet ResNet_V2_101 из https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/g3doc/models.md#image-classification-float-models.
И затем я последовал за https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/tutorials/post_training_quant.ipynb, чтобы преобразовать этот замороженный граф и в модель Lite, и в квантованную модель Lite.
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)
Затем я последовал https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/tools/accuracy/ilsvrc, чтобы оценитьего точность с использованием набора данных валидации ImageNet (50000 изображений) на моем рабочем столе.
Однако, когда я запускаю
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
и проверяю вывод accuracy_output.txt.Точность очень плохая.Я могу захватить некоторые результаты среди 50000 изображений.
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
Однако, согласно https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/tutorials/post_training_quant.ipynb, точность топ-1 может достигать 76,8, но моя попытка даже не может достичь 1 в конце.Почему это происходит?Где я сделал не так?Спасибо!