python - 使用 tensorflow image_dataset_from_directory 时从数据集中获取标签

Question

我在 python (v3.8.3) 中使用 tensorflow (v2.4) + keras 编写了一个简单的 CNN。我正在尝试优化网络，我想要更多关于它无法预测的信息。我正在尝试添加一个混淆矩阵，我需要提供 tensorflow.math.confusion_matrix() 测试标签。
我的问题是我无法弄清楚如何从 tf.keras.preprocessing.image_dataset_from_directory() 创建的数据集对象访问标签
我的图像组织在以标签为名称的目录中。文档说该函数返回一个 tf.data.Dataset 对象。

If label_mode is None, it yields float32 tensors of shape (batch_size, image_size[0], image_size[1], num_channels), encoding

images (see below for rules regarding num_channels). Otherwise, it yields a tuple (images, labels), where images has shape (batch_size, image_size[0], image_size[1], num_channels), and labels follows the format described below.

这是代码:

import tensorflow as tf
from tensorflow.keras import layers
#import matplotlib.pyplot as plt
import numpy as np
import random

import PIL
import PIL.Image

import os
import pathlib

#load the IMAGES
dataDirectory = '/p/home/username/tensorflow/newBirds'

dataDirectory = pathlib.Path(dataDirectory)
imageCount = len(list(dataDirectory.glob('*/*.jpg')))
print('Image count: {0}\n'.format(imageCount))

#test display an image
# osprey = list(dataDirectory.glob('OSPREY/*'))
# ospreyImage = PIL.Image.open(str(osprey[random.randint(1,100)]))
# ospreyImage.show()

# nFlicker = list(dataDirectory.glob('NORTHERN FLICKER/*'))
# nFlickerImage = PIL.Image.open(str(nFlicker[random.randint(1,100)]))
# nFlickerImage.show()

#set parameters
batchSize = 32
height=224
width=224

(trainData, trainLabels) = tf.keras.preprocessing.image_dataset_from_directory(
    dataDirectory,
    labels='inferred',
    label_mode='categorical',
    validation_split=0.2,
    subset='training',
    seed=324893,
    image_size=(height,width),
    batch_size=batchSize)

testData = tf.keras.preprocessing.image_dataset_from_directory(
    dataDirectory,
    labels='inferred',
    label_mode='categorical',
    validation_split=0.2,
    subset='validation',
    seed=324893,
    image_size=(height,width),
    batch_size=batchSize)

#class names and sampling a few images
classes = trainData.class_names
testClasses = testData.class_names
#plt.figure(figsize=(10,10))
# for images, labels in trainData.take(1):
#     for i in range(9):
#         ax = plt.subplot(3, 3, i+1)
#         plt.imshow(images[i].numpy().astype("uint8"))
#         plt.title(classes[labels[i]])
#         plt.axis("off")
# plt.show()

#buffer to hold the data in memory for faster performance
autotune = tf.data.experimental.AUTOTUNE
trainData = trainData.cache().shuffle(1000).prefetch(buffer_size=autotune)
testData = testData.cache().prefetch(buffer_size=autotune)

#augment the dataset with zoomed and rotated images
#use convolutional layers to maintain spatial information about the images
#use max pool layers to reduce
#flatten and then apply a dense layer to predict classes
model = tf.keras.Sequential([
    #layers.experimental.preprocessing.RandomFlip('horizontal', input_shape=(height, width, 3)),
    #layers.experimental.preprocessing.RandomRotation(0.1),
    #layers.experimental.preprocessing.RandomZoom(0.1),
    layers.experimental.preprocessing.Rescaling(1./255, input_shape=(height, width, 3)),
    layers.Conv2D(16, 3, padding='same', activation='relu'),
    layers.MaxPooling2D(),
    layers.Conv2D(32, 3, padding='same', activation='relu'),
    layers.MaxPooling2D(),
    layers.Conv2D(64, 3, padding='same', activation='relu'),
    layers.MaxPooling2D(),
    layers.Conv2D(128, 3, padding='same', activation='relu'),
    layers.MaxPooling2D(),
    layers.Conv2D(256, 3, padding='same', activation='relu'),
    layers.MaxPooling2D(),
    # layers.Conv2D(512, 3, padding='same', activation='relu'),
    # layers.MaxPooling2D(),
    #layers.Conv2D(1024, 3, padding='same', activation='relu'),
    #layers.MaxPooling2D(),
    #dropout prevents overtraining by not allowing each node to see each datapoint
    #layers.Dropout(0.5),
    layers.Flatten(),
    layers.Dense(512, activation='relu'),
    layers.Dense(len(classes))
    ])

model.compile(optimizer='adam',
              loss=tf.keras.losses.CategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])
model.summary()
    
epochs=2
history = model.fit(
    trainData,
    validation_data=testData,
    epochs=epochs
    )

#create confusion matrix
predictions = model.predict_classes(testData)
confusionMatrix = tf.math.confusion_matrix(labels=testClasses, predictions=predictions).numpy()

我曾尝试使用 (foo, foo1) = tf.keras.preprocessing.image_dataset_from_directory(dataDirectory, etc)，但我得到
(trainData, trainLabels) = tf.keras.preprocessing.image_dataset_from_directory(
ValueError:解包的值太多(预期为 2)
如果我尝试作为一个变量返回，然后将其拆分为:

train = tf.keras.preprocessing.image_dataset_from_directory(
    dataDirectory,
    labels='inferred',
    label_mode='categorical',
    validation_split=0.2,
    subset='training',
    seed=324893,
    image_size=(height,width),
    batch_size=batchSize)
trainData = train[0]
trainLabels = train[1]

我得到 TypeError: 'BatchDataset' object is not subscriptable
我可以通过 testClasses = testData.class_names 访问标签，但我得到:

2020-11-03 14:15:14.643300: W tensorflow/core/framework/op_kernel.cc:1740] OP_REQUIRES failed at cast_op.cc:121 : Unimplemented: Cast string to int64 is not supported Traceback (most recent call last): File "birdFake.py", line 115, in confusionMatrix = tf.math.confusion_matrix(labels=testClasses, predictions=predictions).numpy() File "/p/home/username/miniconda3/lib/python3.8/site-packages/tensorflow/python/util/dispatch.py", line 201, in wrapper return target(*args, **kwargs) File "/p/home/username/miniconda3/lib/python3.8/site-packages/tensorflow/python/ops/confusion_matrix.py", line 159, in confusion_matrix labels = math_ops.cast(labels, dtypes.int64) File "/p/home/username/miniconda3/lib/python3.8/site-packages/tensorflow/python/util/dispatch.py", line 201, in wrapper return target(*args, **kwargs) File "/p/home/username/miniconda3/lib/python3.8/site-packages/tensorflow/python/ops/math_ops.py", line 966, in cast x = gen_math_ops.cast(x, base_type, name=name) File "/p/home/username/miniconda3/lib/python3.8/site-packages/tensorflow/python/ops/gen_math_ops.py", line 1827, in cast _ops.raise_from_not_ok_status(e, name) File "/p/home/username/miniconda3/lib/python3.8/site-packages/tensorflow/python/framework/ops.py", line 6862, in raise_from_not_ok_status six.raise_from(core._status_to_exception(e.code, message), None) File "", line 3, in raise_from tensorflow.python.framework.errors_impl.UnimplementedError: Cast string to int64 is not supported [Op:Cast]

我对任何将这些标签放入混淆矩阵的方法持开放态度。任何关于为什么我正在做的事情不起作用的想法也将不胜感激。
更新:我尝试了 Alexandre Catalano 提出的方法，但出现以下错误

Traceback (most recent call last): File "./birdFake.py", line 118, in labels = np.concatenate([labels, np.argmax(y.numpy(), axis=-1)]) File "<array_function internals>", line 5, in concatenate ValueError: all the input arrays must have same number of dimensions, but the array at index 0 has 1 dimension(s) and the array at index 1 has 0 dimension(s)

我打印了标签数组的第一个元素，它是零

Answer 1

如果我是你，我会遍历整个 testData，我会一路保存预测和标签，最后我会构建混淆矩阵。

testData = tf.keras.preprocessing.image_dataset_from_directory(
    dataDirectory,
    labels='inferred',
    label_mode='categorical',
    seed=324893,
    image_size=(height,width),
    batch_size=32)


predictions = np.array([])
labels =  np.array([])
for x, y in testData:
  predictions = np.concatenate([predictions, model.predict_classes(x)])
  labels = np.concatenate([labels, np.argmax(y.numpy(), axis=-1)])

tf.math.confusion_matrix(labels=labels, predictions=predictions).numpy()

结果是

Found 4 files belonging to 2 classes.
array([[2, 0],
       [2, 0]], dtype=int32)