问题
I am studying tensorflow lite. I downloaded the ResNet frozen graph ResNet_V2_101 from https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/g3doc/models.md#image-classification-float-models .
And then I followed https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/tutorials/post_training_quant.ipynb to convert this frozen graph to both Lite model and quantized lite model.
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)
Then I followed https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/tools/accuracy/ilsvrc to evalute its accuracy using ImageNet Validation Dataset (50000 images) on my desktop.
However, when I run
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
and checked the output accuracy_output.txt. The accuracy is very poor. I can capture some results among the 50000 images.
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
However, according to https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/tutorials/post_training_quant.ipynb, the top-1 accuracy can reach 76.8 but my attempt even cannot reach 1 in the end. Why this happens? Where I did wrong? Thanks!
回答1:
Please also check your category labels. In case wrong category labels have been used the results would be as you described.
回答2:
Check your model path, in your Python code, it is resnet_v2_101_299_frozen_quantize.tflite but you used a different one ResNet_V2_101.tflite in command line
来源：https://stackoverflow.com/questions/53734494/tensorflow-lite-resnet-example-model-gave-very-poor-result-during-validation-wi