.tflite 模型（从 keras .h5 模型转换而来）总是以相同的概率预测相同的类别

Question

我已经训练了一个模型（使用 keras）来计算举起的手指数量。该模型运行良好（测试图像的准确率约为 99%）。 现在，我正在尝试通过将保存的模型（.h5 文件）转换为 .tflite 文件来将该模型部署到边缘。

使用 tf.lite.TFLiteConverter.from_keras_model_file()，它会转换并给我一个带有此错误的 .tflite 文件：

tensorflow/core/grappler/grappler_item_builder.cc:637] Init node conv2d/kernel/Assign doesn't exist in graph

当我加载这个 tflite 文件并尝试对相同的输入图像进行预测时，它总是预测“零”，这是第一类，概率 = 0.003922。其余类始终为 0.00 在从 Tensorflow 存储库的 Android 图像分类示例应用程序中加载我的 tflite 模型时，我得到了相同的结果。

为什么这个 tflite 模型不能按预期工作？我在转换过程中是否遗漏了某些内容或使用了 TFlite 不支持的操作？请帮忙！

到目前为止，我已经尝试在不同版本的 Tensorflow 上进行转换；
- 1.12
- 1.14
- tf-nightly-gpu 1.14

所有结果都相同。

My .h5 model and a test image, if you would like to try it yourself!

这是我用来转换模型的代码：

import tensorflow as tf

converter = tf.lite.TFLiteConverter.from_keras_model_file("fingers_latest.h5")
tflite_model = converter.convert()
open("fingers_latest.tflite", "wb").write(tflite_model)

我的 keras 模型：

nbatch = 64
IMG_SIZE = 256


def load_data():
    print("Batch size = ", nbatch, "\n")

    train_datagen = ImageDataGenerator(rescale=1. / 255, rotation_range=12., width_shift_range=0.2,
                                       height_shift_range=0.2,
                                       zoom_range=0.15, shear_range=0.2, horizontal_flip=False)

    test_datagen = ImageDataGenerator(rescale=1. / 255)

    train_gen = train_datagen.flow_from_directory('./datasets/fingers_white/train/', target_size=(IMG_SIZE, IMG_SIZE),
                                                  color_mode='rgb',
                                                  batch_size=nbatch, shuffle=True,
                                                  classes=['ZERO', 'ONE', 'TWO', 'THREE', 'FOUR', 'FIVE'],
                                                  class_mode='categorical')

    test_gen = test_datagen.flow_from_directory('./datasets/fingers_white/test/', target_size=(IMG_SIZE, IMG_SIZE),
                                                color_mode='rgb',
                                                batch_size=nbatch,
                                                classes=['ZERO', 'ONE', 'TWO', 'THREE', 'FOUR', 'FIVE'],
                                                class_mode='categorical')

    return train_gen, test_gen


def train_model(train_gen, test_gen):
    model = Sequential()
    model.add(Conv2D(32, (3, 3), activation='relu', input_shape=(IMG_SIZE, IMG_SIZE, 3)))
    model.add(Dropout(0.2))
    model.add(MaxPooling2D((3, 3)))
    model.add(Conv2D(64, (3, 3), activation='relu'))
    model.add(Conv2D(64, (3, 3), activation='relu'))
    model.add(Dropout(0.4))
    model.add(MaxPooling2D((3, 3)))
    model.add(Conv2D(128, (3, 3), activation='relu'))
    model.add(Dropout(0.6))
    model.add(MaxPooling2D((3, 3)))
    model.add(Conv2D(32, (3, 3), activation='relu'))
    model.add(Flatten())
    model.add(Dense(128, activation='relu'))
    model.add(Dropout(0.2))
    model.add(Dense(6, activation='softmax'))

    model.summary()

    model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['acc'])

    STEP_SIZE_TRAIN = train_gen.n // train_gen.batch_size
    STEP_SIZE_TEST = test_gen.n // test_gen.batch_size
    print(STEP_SIZE_TEST, STEP_SIZE_TRAIN)

    model.fit_generator(train_gen, steps_per_epoch=STEP_SIZE_TRAIN, epochs=5, validation_data=test_gen,
                        validation_steps=STEP_SIZE_TEST, use_multiprocessing=True, workers=6)

我用来加载和运行 .tflite 模型的代码：

import tensorflow as tf
import tkinter as tk
from tkinter import filedialog
import PIL
from PIL import Image
import numpy as np
import time

# DEF. PARAMETERS
img_row, img_column = 224, 224
num_channel = 3
num_batch = 1
input_mean = 127.5
input_std = 127.5
floating_model = False

path_1 = r"./models/mobilenet_v2_1.0_224.tflite"
labels_path = "./models/labels_mobilenet.txt"

def load_labels(filename):
    my_labels = []
    input_file = open(filename, 'r')
    for l in input_file:
        my_labels.append(l.strip())
    return my_labels

interpreter = tf.lite.Interpreter(path_1)
interpreter.allocate_tensors()

# obtaining the input-output shapes and types
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
print(input_details, '\n', output_details)

# file selection window for input selection
root = tk.Tk()
root.withdraw()
file_path = filedialog.askopenfilename()
input_img = Image.open(file_path)
input_img = input_img.resize((img_row, img_column))
input_img = np.expand_dims(input_img, axis=0)

input_img = (np.float32(input_img) - input_mean) / input_std

interpreter.set_tensor(input_details[0]['index'], input_img)

# running inference
interpreter.invoke()

output_data = interpreter.get_tensor(output_details[0]['index'])
results = np.squeeze(output_data)

top_k = results.argsort()[-5:][::-1]
labels = load_labels(labels_path)
for i in top_k:
    print('{0:08.6f}'.format(float(results[i] / 255.0)) + ":", labels[i])

Answer 1

在您的训练代码中，您将图像标准化为[0..1] 的范围，该范围由以下几行指定：

train_datagen = ImageDataGenerator(rescale=1. / 255, ...)
test_datagen = ImageDataGenerator(rescale=1. / 255)

转换后的tflite模型的输入范围是[-1..1]，由以下几行指定：

input_mean = 127.5
input_std = 127.5
input_img = (np.float32(input_img) - input_mean) / input_std

因此，您可以尝试将以上行替换为：

input_mean = 0.
input_std = 255.