tf.keras model.predict 每次提供不同的值

Question

每次我跑：

y_true = np.argmax(tf.concat([y for x, y in train_ds], axis=0), axis=1)
y_pred = np.argmax(model.predict(train_ds), axis=1)

confusion_matrix(y_true, y_pred)

每次的结果都与我的理解不同： y_pred = np.argmax(model.predict(train_ds), axis=1) 每次都不一样。

澄清：我运行单元 1（训练）一次。和单元格 2（推理）几次。

为什么？

代码： 单元格 1 (jupyter)

import numpy as np
import tensorflow as tf
from tensorflow.keras.layers import Input, Conv2D, experimental
from tensorflow.keras.layers import MaxPool2D, Flatten, Dense 
from tensorflow.keras import Model
from tensorflow.keras.losses import categorical_crossentropy
from sklearn.metrics import accuracy_score


image_size = (100, 100)
batch_size = 32

train_ds = tf.keras.preprocessing.image_dataset_from_directory(
    directory,
    label_mode='categorical',
    validation_split=0.2,
    subset="training",
    seed=1337,
    color_mode="grayscale",
    image_size=image_size,
    batch_size=batch_size,
)


inputs = Input(shape =(100,100,1))
x = experimental.preprocessing.Rescaling(1./255)(inputs)
x = Conv2D (filters =4, kernel_size =3, padding ='same', activation='relu')(x)
x = Conv2D (filters =4, kernel_size =3, padding ='same', activation='relu')(x)
x = MaxPool2D(pool_size =2, strides =2, padding ='same')(x)

x = Conv2D (filters =8, kernel_size =3, padding ='same', activation='relu')(x)
x = Conv2D (filters =8, kernel_size =3, padding ='same', activation='relu')(x)
x = MaxPool2D(pool_size =2, strides =2, padding ='same')(x)

x = Flatten()(x)
x = Dense(units = 4, activation ='relu')(x)
x = Dense(units = 4, activation ='relu')(x)
output = Dense(units = 5, activation ='softmax')(x)

model = Model (inputs=inputs, outputs =output)

model.compile(
    optimizer=tf.keras.optimizers.Adam(1e-3),
    loss=categorical_crossentropy,
    metrics=["accuracy"])

model.fit(train_ds, epochs=5)

细胞 2：

print (Accuracy:)
y_pred = np.argmax(model.predict(train_ds), axis=1)
print (accuracy_score(y_true, y_pred))

y_pred = np.argmax(model.predict(train_ds), axis=1)
print (accuracy_score(y_true, y_pred))

输出：

118/118 [===============================] - 7 秒 57 毫秒/步 - 损失：0.1888 - 准确度： 0.9398

准确度：

0.593

0.586

Answer 1

据我目前的理解，上述原因是：

tf.keras.preprocessing.image_dataset_from_directory

虽然它的实例是：

type(train_ds)

tensorflow.python.data.ops.dataset_ops.BatchDataset

复制：

首次运行：

[x for x, y in train_ds]

输出：

[<tf.Tensor: shape=(32, 100, 100, 1), dtype=float32, numpy=  array([[[[157.],
          [155.],
          [159.],

第二次运行：

[x for x, y in train_ds]

输出：

[<tf.Tensor: shape=(32, 100, 100, 1), dtype=float32, numpy=  array([[[[ 34.],
          [ 36.],
          [ 39.],
          ...,

可能的解决方案

imgs, y_true = [], []
for img, label in train_ds:
    imgs.append(img)
    y_true.append(label)

imgs = tf.concat(imgs, axis=0)    
y_true = np.argmax(tf.concat(y_true, axis=0), axis=1)

y_pred = np.argmax(model.predict(imgs), axis=1)
print (accuracy_score(y_true, y_pred))

y_pred = np.argmax(model.predict(imgs), axis=1)
print (accuracy_score(y_true, y_pred))

输出

0.944044764
0.944044764

有没有更好的解决方案？

更新 2：

在验证数据集的情况下可能更合适（这里的train_ds 只是例如添加一个参数Shuffle=False）

    train_ds = tf.keras.preprocessing.image_dataset_from_directory(
        directory,
        label_mode='categorical',
        validation_split=0.2,
        subset="training",
        seed=1337,
        color_mode="grayscale",
        image_size=image_size,
        batch_size=batch_size,
        Shuffle=False

    )

更新 3：

如果您的 test images 在单独的文件夹中，这可能是最好的选择。

path = 'your path to test folder'


test_generator = ImageDataGenerator().flow_from_directory(
    directory=path,
    class_mode='categorical',
    shuffle=False,
    batch_size=32,
    target_size=(512, 512)
)

test_generator.reset()

这比选项 1 更好，因为它可以在不适合内存 (RAM) 的数据集上工作。

Answer 2

您确定不会在每次运行代码时再次训练模型吗？如果模型的参数相同，则相同输入的预测结果每次都应该相同。