为什么Tensorflow多类图像预测在加载模型时不起作用？

我目前正在尝试学习机器学习技术，并希望用tensorflow重新创建一个简单的图像识别算法。因此，我创建了两个Python文件：一个用于培训，一个用于预测

在Ubuntu 18.04上测试 使用的Python版本：3.7 使用的Numpy版本：1.18.1 使用的Tensorflow版本：1.14和2.1.0（以下输出来自版本1.14）

我的图像来自 这组照片由60个人的大约3000张剪过的脸组成

列车_model.py：

#!/usr/bin/env python

import concurrent.futures
import pandas as pd
import urllib
import pathlib
import hashlib
import os
import json
import tensorflow as tf
from tensorflow.keras.preprocessing.image import ImageDataGenerator

people = pd.read_csv("dev_people.txt")

image_generator = ImageDataGenerator(rescale=1./255, validation_split=0.2, rotation_range=45, zoom_range=0.2)

IMG_HEIGHT = 128
IMG_WIDTH = 128
LEARNING_RATE = 0.0001
BATCH_SIZE = 32
NUM_TRAIN = 100
STEPS_PER_EPOCH = round(NUM_TRAIN) // BATCH_SIZE
VAL_STEPS = 20
NUM_EPOCHS = 3

train_data = image_generator.flow_from_directory(batch_size=BATCH_SIZE,
    directory="persons-cropped",
    shuffle=True,
    target_size=(IMG_HEIGHT, IMG_WIDTH),
    class_mode="categorical",
    subset="training")

labels = train_data.class_indices
labels = {v: k for k, v in labels.items()}

with open("labels.json", "w") as labels_file:
    labels_file.write(json.dumps(labels))

validation_data = image_generator.flow_from_directory(batch_size=BATCH_SIZE,
    directory="persons-cropped",
    shuffle=True,
    target_size=(IMG_HEIGHT, IMG_WIDTH),
    class_mode="categorical",
    subset="validation")

base_model = tf.keras.applications.MobileNetV2(
    input_shape=(IMG_WIDTH, IMG_HEIGHT, 3),
    include_top=False,
    weights="imagenet"
)

base_model.trainable = False

maxpool_layer = tf.keras.layers.GlobalMaxPooling2D()
prediction_layer = tf.keras.layers.Dense(60, activation="sigmoid")
dropout_layer = tf.keras.layers.Dropout(0.2)

model = tf.keras.Sequential([
    base_model,
    maxpool_layer,
#   dropout_layer,
    prediction_layer
])

model.compile(optimizer=tf.keras.optimizers.Adam(lr=LEARNING_RATE),
    loss="binary_crossentropy",
    metrics=["accuracy"]
)

model.summary()

model.fit(
    train_data,
    epochs=NUM_EPOCHS,
    steps_per_epoch=None,
    validation_data=validation_data,
    validation_steps=None,
    use_multiprocessing=False,
    workers=6,
    verbose=2
)

model.save("model.h5")

输出：

Found 2431 images belonging to 60 classes.
Found 573 images belonging to 60 classes.
WARNING:tensorflow:From /usr/local/lib/python3.7/dist-packages/tensorflow/python/ops/init_ops.py:1251: calling VarianceScaling.__init__ (from tensorflow.python.ops.init_ops) with dtype is deprecated and will be removed in a future version.
Instructions for updating:
Call initializer instance with the dtype argument instead of passing it to the constructor
2020-01-25 22:23:40.036326: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 2599985000 Hz
2020-01-25 22:23:40.036657: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x6b81c60 executing computations on platform Host. Devices:
2020-01-25 22:23:40.036789: I tensorflow/compiler/xla/service/service.cc:175]   StreamExecutor device (0): <undefined>, <undefined>
2020-01-25 22:23:40.615771: W tensorflow/compiler/jit/mark_for_compilation_pass.cc:1412] (One-time warning): Not using XLA:CPU for cluster because envvar TF_XLA_FLAGS=--tf_xla_cpu_global_jit was not set.  If you want XLA:CPU, either set that envvar, or use experimental_jit_scope to enable XLA:CPU.  To confirm that XLA is active, pass --vmodule=xla_compilation_cache=1 (as a proper command-line flag, not via TF_XLA_FLAGS) or set the envvar XLA_FLAGS=--xla_hlo_profile.
WARNING:tensorflow:From /usr/local/lib/python3.7/dist-packages/tensorflow/python/ops/nn_impl.py:180: add_dispatch_support.<locals>.wrapper (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.
Instructions for updating:
Use tf.where in 2.0, which has the same broadcast rule as np.where
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
mobilenetv2_1.00_224 (Model) (None, 7, 7, 1280)        2257984
_________________________________________________________________
global_max_pooling2d (Global (None, 1280)              0
_________________________________________________________________
dense (Dense)                (None, 60)                76860
=================================================================
Total params: 2,334,844
Trainable params: 76,860
Non-trainable params: 2,257,984
_________________________________________________________________
Epoch 1/3
2020-01-25 22:23:55.995833: W tensorflow/core/framework/allocator.cc:107] Allocation of 154140672 exceeds 10% of system memory.
2020-01-25 22:23:56.730363: W tensorflow/core/framework/allocator.cc:107] Allocation of 156905472 exceeds 10% of system memory.
2020-01-25 22:24:02.782372: W tensorflow/core/framework/allocator.cc:107] Allocation of 154140672 exceeds 10% of system memory.
2020-01-25 22:24:03.531172: W tensorflow/core/framework/allocator.cc:107] Allocation of 156905472 exceeds 10% of system memory.
2020-01-25 22:24:09.474692: W tensorflow/core/framework/allocator.cc:107] Allocation of 154140672 exceeds 10% of system memory.
/usr/local/lib/python3.7/dist-packages/PIL/TiffImagePlugin.py:788: UserWarning: Corrupt EXIF data.  Expecting to read 4 bytes but only got 0.
  warnings.warn(str(msg))
76/76 - 602s - loss: 0.3851 - acc: 0.9097 - val_loss: 0.1812 - val_acc: 0.9495
Epoch 2/3
76/76 - 616s - loss: 0.1480 - acc: 0.9757 - val_loss: 0.1732 - val_acc: 0.9544
Epoch 3/3
76/76 - 627s - loss: 0.1452 - acc: 0.9760 - val_loss: 0.1767 - val_acc: 0.9516

输出：使用Zach Braff图像时：

{'0': 'Abhishek Bachan', '1': 'Alex Rodriguez', '2': 'Ali Landry', '3': 'Alyssa Milano', '4': 'Anderson Cooper', '5': 'Anna Paquin', '6': 'Audrey Tautou', '7': 'Barack Obama', '8': 'Ben Stiller', '9': 'Christina Ricci', '10': 'Clive Owen', '11': 'Cristiano Ronaldo', '12': 'Daniel Craig', '13': 'Danny Devito', '14': 'David Duchovny', '15': 'Denise Richards', '16': 'Diane Sawyer', '17': 'Donald Faison', '18': 'Ehud Olmert', '19': 'Faith Hill', '20': 'Famke Janssen', '21': 'Hugh Jackman', '22': 'Hugh Laurie', '23': 'James Spader', '24': 'Jared Leto', '25': 'Julia Roberts', '26': 'Julia Stiles', '27': 'Karl Rove', '28': 'Katherine Heigl', '29': 'Kevin Bacon', '30': 'Kiefer Sutherland', '31': 'Kim Basinger', '32': 'Mark Ruffalo', '33': 'Meg Ryan', '34': 'Michelle Trachtenberg', '35': 'Michelle Wie', '36': 'Mickey Rourke', '37': 'Miley Cyrus', '38': 'Milla Jovovich', '39': 'Nicole Richie', '40': 'Rachael Ray', '41': 'Robert Gates', '42': 'Ryan Seacrest', '43': 'Sania Mirza', '44': 'Sarah Chalke', '45': 'Sarah Palin', '46': 'Scarlett Johansson', '47': 'Seth Rogen', '48': 'Shahrukh Khan', '49': 'Shakira', '50': 'Stephen Colbert', '51': 'Stephen Fry', '52': 'Steve Carell', '53': 'Steve Martin', '54': 'Tracy Morgan', '55': 'Ty Pennington', '56': 'Viggo Mortensen', '57': 'Wilmer Valderrama', '58': 'Zac Efron', '59': 'Zach Braff'}
2020-01-25 22:58:05.582049: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 2599985000 Hz
2020-01-25 22:58:05.582514: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x618f910 executing computations on platform Host. Devices:
2020-01-25 22:58:05.582653: I tensorflow/compiler/xla/service/service.cc:175]   StreamExecutor device (0): <undefined>, <undefined>
2020-01-25 22:58:06.454565: W tensorflow/compiler/jit/mark_for_compilation_pass.cc:1412] (One-time warning): Not using XLA:CPU for cluster because envvar TF_XLA_FLAGS=--tf_xla_cpu_global_jit was not set.  If you want XLA:CPU, either set that envvar, or use experimental_jit_scope to enable XLA:CPU.  To confirm that XLA is active, pass --vmodule=xla_compilation_cache=1 (as a proper command-line flag, not via TF_XLA_FLAGS) or set the envvar XLA_FLAGS=--xla_hlo_profile.
1/1 [==============================] - 1s 999ms/sample
[[3.23683023e-04 6.47217035e-04 3.90201807e-04 2.69789696e-02
  2.17908323e-02 1.53781831e-01 4.79090214e-03 8.64863396e-04
  1.11432403e-01 8.87395382e-01 3.30170989e-03 2.17252970e-03
  1.78458661e-01 1.09243691e-02 1.47551298e-04 2.62927115e-02
  3.22049320e-01 2.69562006e-04 9.11523938e-01 2.44581699e-03
  7.65213370e-03 2.90286541e-03 1.01376325e-01 6.43432140e-05
  4.43832874e-02 3.94093990e-03 6.90050423e-02 7.47233629e-04
  1.05589628e-03 8.04662704e-07 3.76045704e-03 4.28827941e-01
  1.20029151e-02 1.77664489e-01 5.27173281e-04 2.45797634e-03
  5.89579344e-03 9.46103930e-01 2.79089808e-03 2.09265649e-02
  2.83238888e-02 4.86207008e-03 8.15459788e-02 1.30202770e-02
  1.50602162e-02 1.33922696e-03 1.24056339e-02 5.76970875e-02
  2.65627503e-02 5.18084109e-01 4.89562750e-04 3.15269828e-03
  4.88847494e-04 2.13665128e-01 1.40489936e-02 2.93705761e-02
  5.01989722e-02 1.21492555e-03 1.62564263e-01 2.91267484e-01]]
[37]
['Miley Cyrus']

{'0'：'Abhishek Bachan'，'1'：'Alex Rodriguez'，'2'：'Ali Landry'，'3'：'Alyssa Milano'，'4'：'Anderson Cooper'，'5'：'Anna Paquin'，'6'：'Audrey Tautou'，'7'：'Barack Obama'，'8'：'Ben Stiller'，'9'：'Christina Ricci'，'Clive Owen 11'：'Cristiano Ronaldo'，'12'：'Daniel Craig'，'13'：'Danny Devito'，'14'：'David Duchovny'，'15'：'Denise理查兹、16岁、黛安·索耶、17岁、唐纳德·费森、18岁、埃胡德·奥尔默特、19岁、费思·希尔、20岁、法姆·扬森、21岁、休·杰克曼、22岁、休·劳里、23岁、詹姆斯·斯派德、24岁、杰瑞德·莱托、25岁、茱莉亚·罗伯茨、26岁、茱莉亚·斯蒂尔斯、27岁、卡尔·罗夫、28岁、凯瑟琳·海格尔、29岁、凯文·培根、30岁、基弗·萨瑟兰、31岁金·贝辛格，32岁：马克·鲁法洛，33岁：梅格·瑞安，34岁：米歇尔·特雷斯登堡，35岁：米歇尔·维，36岁：米奇·洛克，37岁：麦莉·塞勒斯，38岁：米拉·乔沃维奇，39岁：妮可·里奇，40岁：瑞秋·雷，41岁：罗伯特·盖茨，42岁：瑞安·西克雷斯特，43岁：萨尼亚·米尔扎，44岁：莎拉·乔克，45岁：莎拉·佩林，46岁：斯佳丽·约翰纳森、47、48、49、夏奇拉、50、斯蒂芬·科尔伯特、51、斯蒂芬·弗莱、52、史蒂夫·卡雷尔、53、史蒂夫·马丁、54、特雷西·摩根、55、蒂·彭宁顿、56、维戈·莫滕森、57、威尔默·瓦尔德拉马、58、扎克·埃夫隆、59、扎克·布拉夫
2020-01-25 22:58:05.582049:I tensorflow/core/platform/profile_-utils/cpu_-utils.cc:94]cpu频率：2599985000 Hz
20201-01-2522:58: 5.582514：I TysFrace/编译器/ XLA/Service / Service .CC：168）XLA服务0x618F910在平台主机上执行计算。
2020-01-25 22:58:05.582653:I tensorflow/compiler/xla/service/service.cc:175]流执行器设备（0）：，
2020-01-25 22:58:06.454565:W tensorflow/compiler/jit/mark_for_compilement_pass.cc:1412]（一次警告）不使用XLA:CopyCopy:EnvavaTfxxLaFLAG= = TfxxLaCpUGLULALLJJIT没有设置。如果您想要XLA:CPU，要么设置EnVaR，要么使用ExtalTall JITIOSCAPE启用XLA:CPU。确认XLA是活动的，通过-VMask= XLAYCOMPILITCIONSOCH＝1（作为一个适当的命令行标志，而不是通过TFXxLAFLAGS）或者设置envvar XLA\u FLAGS=--XLA\u hlo\u配置文件。
1/1[====================================]-1s 999ms/样本
[3.23683023e-04 6.47217035e-04 3.90201807e-04 2.69789696e-02
2.17908323e-02 1.53781831e-01 4.79090214e-03 8.64863396e-04
1.11432403e-01 8.87395382e-01 3.30170989e-03 2.17252970e-03
1.78458661e-01 1.09243691e-02 1.47551298e-04 2.62927115e-02
3.22049320e-01 2.69562006e-04 9.11523938e-01 2.44581699e-03
7.65213370e-03 2.90286541e-03 1.01376325e-01 6.43432140e-05
4.43832874e-02 3.94093990e-03 6.90050423e-02 7.47233629e-04
1.05589628e-03 8.04662704e-07 3.76045704e-03 4.28827941e-01
1.20029151e-02 1.77664489e-01 5.27173281e-04 2.45797634e-03
5.89579344e-03 9.46103930e-01 2.79089808e-03 2.09265649e-02
2.83238888e-02 4.86207008e-03 8.15459788e-02 1.30202770e-02
1.50602162e-02 1.33922696e-03 1.24056339e-02 5.76970875e-02
2.65627503e-02 5.18084109e-01 4.89562750e-04 3.15269828e-03
4.88847494e-04 2.13665128e-01 1.40489936e-02 2.93705761e-02
5.01989722e-02 1.21492555e-03 1.62564263e-01 2.91267484e-01]]
[37]
[“麦莉·塞勒斯”]

预测算法一直都是错误的。如果我使用其他Zach Braff图像，当然同一张图片的输出保持不变，但在10个测试案例中，它从来不是Zach Braff，而是一个不同的人。（不仅是Miley Cyrus，还有Shakira，Steve Carell，…）

这个模式对于我在这里用作输入的任何类都会出现

---

我在互联网上找不到任何有用的建议，尝试了我能想象到的所有参数组合。我还使用了两个版本的Tensorflow，确保所有库都是最新的。

嘿，我相信你得到了奇怪的预测，因为在编译模型时，你的数据分布有60类人具有一个设置为二进制交叉熵的损失函数

 model.compile(optimizer=tf.keras.optimizers.Adam(lr=LEARNING_RATE),
 loss="categorical_crossentropy", # here
 metrics=["accuracy"]
)

二进制交叉熵用于确定最多2个类。您需要做的是将损失函数更改为分类交叉熵

 model.compile(optimizer=tf.keras.optimizers.Adam(lr=LEARNING_RATE),
 loss="categorical_crossentropy", # here
 metrics=["accuracy"]
)

---

嘿，我相信你得到了奇怪的预测，因为你的数据分布有60类人，而你的模型是用一个设置为二进制交叉熵的损失函数编译的

 model.compile(optimizer=tf.keras.optimizers.Adam(lr=LEARNING_RATE),
 loss="categorical_crossentropy", # here
 metrics=["accuracy"]
)

二进制交叉熵用于确定最多2个类。您需要做的是将损失函数更改为分类交叉熵

 model.compile(optimizer=tf.keras.optimizers.Adam(lr=LEARNING_RATE),
 loss="categorical_crossentropy", # here
 metrics=["accuracy"]
)