python - 如何在 Keras CNN 模型中预测单个图像？

Question

我正在按照本指南学习使用 CNN 进行图像分类，并将此代码实现到我的数据集中:

https://www.tensorflow.org/tutorials/images/classification

train_image_generator = ImageDataGenerator(rescale=1. / 255)  # Generator for our training data
validation_image_generator = ImageDataGenerator(rescale=1. / 255)  # Generator for our validation data

train_data_gen = train_image_generator.flow_from_directory(batch_size=batch_size,
                                                           directory=train_img_folder,
                                                           shuffle=True,
                                                           target_size=(IMG_HEIGHT, IMG_WIDTH),
                                                           class_mode='categorical',
                                                           color_mode='grayscale')

val_data_gen = validation_image_generator.flow_from_directory(batch_size=batch_size,
                                                              directory=valid_img_folder,
                                                              target_size=(IMG_HEIGHT, IMG_WIDTH),
                                                              class_mode='categorical',
                                                              color_mode='grayscale'
                                                              )

model = Sequential([
    Conv2D(16, 3, padding='same', activation='relu', input_shape=(IMG_HEIGHT, IMG_WIDTH, 1)),
    MaxPooling2D(),
    Conv2D(32, 3, padding='same', activation='relu'),
    MaxPooling2D(),
    Conv2D(64, 3, padding='same', activation='relu'),
    MaxPooling2D(),
    Flatten(),
    Dense(512, activation='relu'),
    Dense(3, activation='softmax')
])

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

history = model.fit_generator(
    train_data_gen,
    steps_per_epoch=total_train_value // batch_size,
    epochs=epochs,
    validation_data=val_data_gen,
    validation_steps=total_valid_value // batch_size
)

# Single prediction
img = []
temp = np.array(Image.open('path/to/pic.jpg').resize((256, 256), Image.ANTIALIAS))
temp.shape = temp.shape + (1,) # now its (256, 256, 1)
img.append(temp)
test = np.array(img) # (1, 1024, 1024, 1)
prediction = model.predict(test) 

当我尝试 Predict_generator 函数时:

test_datagen = ImageDataGenerator(rescale=1 / 255.)

test_generator = test_datagen.flow_from_directory('test_images/',
                                                  classes=['0', '1', '2'],
                                                  color_mode='grayscale',
                                                  shuffle=True,
                                                  # use same size as in training
                                                  target_size=(256, 256))

preds = model.predict_generator(test_generator, steps=4) # I dont know what is steps doing. I put there because of error. 

我的第一个问题是:我可以获得训练和验证的准确性，但我想获得单张图片的预测结果。我怎样才能做到这一点？示例:

foo = model.predict(path/to/pic.jpg)
# foo returns 0-> 0.70 | 1-> 0.30

添加:当我尝试像这样使用 model.predict 时，出现此错误:

ValueError: Error when checking input: expected conv2d_1_input to have 4 dimensions, but got array with shape (1024, 1024)

或转换为 2d(以及 3d)np.array 仍然相同。

我的第二个问题是:有没有办法在不完成 %100 的情况下进行预测？我的意思是，如果我们有 2 个类别(猫和狗)并测试月球图片，我希望得到这样的结果:

%15 cat | %10 dog

不是

%50 cat | %50 dog

我尝试将垃圾类别设置为以下更改。当我在 history = model.fit_generator 行运行该命令时，出现以下错误:

ValueError: Error when checking target: expected dense_2 to have shape (3,) but got array with shape (2,) 

Answer 1

First question :I can get training and validation accuracy but I want to get single picture's prediction result. How can I do that?

正如您在 the doc 中看到的那样，您完全可以使用 model.predict(x)，只要您的 x 是:
- Numpy 数组(或类似数组)或数组列表(如果模型有多个输入)。
- 如果模型具有命名输入，则将输入名称映射到相应的数组/张量的字典。
- 生成器或 keras.utils.Sequence 返回(输入、目标)或(输入、目标、样本权重)。

您只需编写读取 .jpg 图像的代码并将其提供给模型即可。

Second question : is there any way to predict without complete %100? I mean If we have 2 classes (cat and dog) and test moon picture i want to get results like that:

您可以创建第三类“垃圾”，为此，您需要将网络的最后一层更改为:

Dense(3, activation='softmax')

并将损失更改为categorical_crossentropy

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

并将class_mode更改为categorical而不是binary。

在这种情况下，你将拥有狗:15%，猫:10%，垃圾:75%

编辑 conv2D 错误:

ValueError: Error when checking input: expected conv2d_1_input to have 4 dimensions, but got array with shape (1024, 1024)

你有:

Conv2D(16, 3, padding='same', activation='relu', input_shape=(IMG_HEIGHT, IMG_WIDTH, 3)),

这意味着图像是(高度、宽度、 channel )。
如the doc中所示，因为这是 input_layer，所以您需要提供具有以下形状的 4D 格式:(samples, rows, cols, Channels)。如果您只想提供一张图像，则需要一个形状为 (1, rows, cols, Channels) 的数组。