Очень низкая точность проверки, но высокая точность обучения ⇐ Python

[Anonymous]

Я скопировал пример кода прямо с официального сайта Keras и отредактировал его, чтобы создать модель машинного обучения.
Я использую Google Colab для своего кода.
Ссылка: https ://keras.io/examples/vision/image_classification_from_scratch/

Код: Выделить всё

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers

from google.colab import drive
drive.mount("/content/gdrive")
image_size = (50, 50)
batch_size = 400

import random
num = random.randint(1, 400)
#random seed
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
"/content/gdrive/My Drive/pest/train",
seed=num,
image_size=image_size,
batch_size=batch_size,
)
val_ds = tf.keras.preprocessing.image_dataset_from_directory(
"/content/gdrive/My Drive/pest/test",
seed=num,
image_size=image_size,
batch_size=batch_size,
)
#tried data augmentation
data_augmentation = keras.Sequential(
[
layers.RandomFlip("horizontal"),
layers.RandomRotation(0.1),
]
)

def make_model(input_shape, num_classes):
inputs = keras.Input(shape=input_shape)
# Image augmentation block
x = data_augmentation(inputs)

# Entry block
x = layers.Rescaling(1.0 / 255)(x)
x = layers.Conv2D(32, 3, strides=2, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.Activation("relu")(x)

x = layers.Conv2D(64, 3, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.Activation("relu")(x)

previous_block_activation = x  # Set aside residual

for size in [128, 256, 512, 728]:
x = layers.Activation("relu")(x)
x = layers.SeparableConv2D(size, 3, padding="same")(x)
x = layers.BatchNormalization()(x)

x = layers.Activation("relu")(x)
x = layers.SeparableConv2D(size, 3, padding="same")(x)
x = layers.BatchNormalization()(x)

x = layers.MaxPooling2D(3, strides=2, padding="same")(x)

# Project residual
residual = layers.Conv2D(size, 1, strides=2, padding="same")(
previous_block_activation
)
x = layers.add([x, residual])  # Add back residual
previous_block_activation = x  # Set aside next residual

x = layers.SeparableConv2D(1024, 3, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.Activation("relu")(x)

x = layers.GlobalAveragePooling2D()(x)
if num_classes == 2:
activation = "sigmoid"
units = 1
else:
activation = "softmax"
units = num_classes

x = layers.Dropout(0.5)(x)
outputs = layers.Dense(units, activation=activation)(x)
return keras.Model(inputs, outputs)

model = make_model(input_shape=image_size + (3,), num_classes=2)
keras.utils.plot_model(model, show_shapes=True)

epochs = 50

callbacks = [
keras.callbacks.ModelCheckpoint("save_at_{epoch}.h5"),
]
model.compile(
optimizer=keras.optimizers.Adam(1e-3),
loss="binary_crossentropy",
metrics=["accuracy"],
)
model.fit(
train_ds, epochs=epochs, callbacks=callbacks, validation_data=val_ds,
#should have automatic shuffling I think
)

Однако, когда я запускаю его, результат

Код: Выделить всё

Epoch 1/50
2/2 [==============================] - 71s 14s/step - loss: 0.6260 - accuracy: 0.6050 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 2/50
2/2 [==============================] - 2s 507ms/step - loss: 0.2689 - accuracy: 0.8867 - val_loss: 0.6932 - val_accuracy: 0.5000
Epoch 3/50
2/2 [==============================] - 2s 536ms/step - loss: 0.1241 - accuracy: 0.9483 - val_loss: 0.6932 - val_accuracy: 0.5000
Epoch 4/50
2/2 [==============================] - 2s 506ms/step - loss: 0.0697 - accuracy: 0.9750 - val_loss: 0.6934 - val_accuracy: 0.5000
Epoch 5/50
2/2 [==============================] - 2s 525ms/step - loss: 0.0479 - accuracy: 0.9867 - val_loss: 0.6936 - val_accuracy: 0.5000
Epoch 6/50
2/2 [==============================] - 2s 534ms/step - loss: 0.0359 - accuracy: 0.9867 - val_loss: 0.6940 - val_accuracy: 0.5000
Epoch 7/50
2/2 [==============================] - 2s 509ms/step - loss: 0.0145 - accuracy:  0.9983 - val_loss: 0.6946 - val_accuracy: 0.5000
Epoch 8/50
2/2 [==============================] - 2s 545ms/step - loss: 0.0124 - accuracy: 0.9967 - val_loss: 0.6954 - val_accuracy: 0.5000
Epoch 9/50
2/2 [==============================] - 2s 544ms/step - loss: 0.0092 - accuracy: 0.9967 - val_loss: 0.6964 - val_accuracy: 0.5000
Epoch 10/50
2/2 [==============================] - 2s 512ms/step - loss: 0.0060 - accuracy: 0.9967 - val_loss: 0.6980 - val_accuracy: 0.5000
Epoch 11/50
2/2 [==============================] - 2s 535ms/step - loss: 0.0036 - accuracy: 0.9983 - val_loss: 0.6998 - val_accuracy: 0.5000
Epoch 12/50
2/2 [==============================] - 2s 503ms/step - loss: 0.0085 - accuracy: 0.9983 - val_loss: 0.7020 - val_accuracy: 0.5000
Epoch 13/50
2/2 [==============================] - 2s 665ms/step - loss: 0.0040 - accuracy: 1.0000 - val_loss: 0.7046 - val_accuracy: 0.5000
Epoch 14/50
2/2 [==============================] - 2s 516ms/step - loss: 0.0017 - accuracy: 1.0000 - val_loss: 0.7078 - val_accuracy: 0.5000
Epoch 15/50
2/2 [==============================] - 2s 520ms/step - loss: 0.0023 - accuracy: 0.9983 - val_loss: 0.7115 - val_accuracy: 0.5000
Epoch 16/50
2/2 [==============================] - 2s 500ms/step - loss: 8.5606e-04 - accuracy: 1.0000 - val_loss: 0.7157 - val_accuracy: 0.5000
Epoch 17/50
2/2 [==============================] - 2s 524ms/step - loss: 0.0018 - accuracy: 1.0000 - val_loss: 0.7205 - val_accuracy: 0.5000
Epoch 18/50
2/2 [==============================] - 2s 499ms/step - loss: 9.0626e-04 - accuracy: 1.0000 - val_loss: 0.7258 - val_accuracy: 0.5000
Epoch 19/50
2/2 [==============================] - 2s 510ms/step - loss: 0.0014 - accuracy: 1.0000 - val_loss: 0.7313 - val_accuracy: 0.5000
Epoch 20/50
2/2 [==============================] - 2s 711ms/step - loss: 0.0013 - accuracy: 1.0000 - val_loss: 0.7371 - val_accuracy: 0.5000
Epoch 21/50
2/2 [==============================] - 2s 511ms/step - loss: 9.9904e-04 - accuracy: 1.0000 - val_loss: 0.7431 - val_accuracy: 0.5000
Epoch 22/50
2/2 [==============================] - 2s 540ms/step - loss: 0.0019 - accuracy: 1.0000 - val_loss: 0.7489 - val_accuracy: 0.5000
Epoch 23/50
2/2 [==============================] - 2s 513ms/step - loss: 4.9861e-04 - accuracy: 1.0000 - val_loss: 0.7553 - val_accuracy: 0.5000
Epoch 24/50
2/2 [==============================] - 2s 542ms/step - loss: 6.6248e-04 - accuracy: 1.0000 - val_loss: 0.7622 - val_accuracy: 0.5000
Epoch 25/50
2/2 [==============================] - 2s 510ms/step - loss: 7.7911e-04 - accuracy: 1.0000 - val_loss: 0.7699 - val_accuracy: 0.5000
Epoch 26/50
2/2 [==============================] - 2s 502ms/step - loss: 3.3703e-04 - accuracy: 1.0000 - val_loss: 0.7781 - val_accuracy: 0.5000
Epoch 27/50
2/2 [==============================] - 2s 539ms/step - loss: 3.7860e-04 - accuracy: 1.0000 - val_loss: 0.7870 - val_accuracy: 0.5000
Epoch 28/50
2/2 [==============================] - 2s 507ms/step - loss: 2.4852e-04 - accuracy: 1.0000 - val_loss: 0.7962 - val_accuracy: 0.5000
Epoch 29/50
2/2 [==============================] - 2s 512ms/step - loss: 1.7709e-04 - accuracy: 1.0000 - val_loss: 0.8058 - val_accuracy: 0.5000
Epoch 30/50
2/2 [==============================] - 2s 538ms/step - loss: 1.6884e-04 - accuracy: 1.0000 - val_loss: 0.8161 - val_accuracy: 0.5000
Epoch 31/50
2/2 [==============================] - 2s 521ms/step - loss: 2.0884e-04 - accuracy: 1.0000 - val_loss: 0.8266 - val_accuracy: 0.5000
Epoch 32/50
2/2 [==============================] - 2s 543ms/step - loss: 1.8691e-04 - accuracy: 1.0000 - val_loss: 0.8375 - val_accuracy: 0.5000
Epoch 33/50
2/2 [==============================] - 2s 520ms/step - loss: 1.7296e-04 - accuracy: 1.0000 - val_loss: 0.8487 - val_accuracy: 0.5000
Epoch 34/50
2/2 [==============================] - 2s 516ms/step - loss: 4.5739e-04 - accuracy: 1.0000 - val_loss: 0.8601 - val_accuracy: 0.5000
Epoch 35/50
2/2 [==============================] - 2s 530ms/step - loss: 9.6831e-05 - accuracy: 1.0000 - val_loss: 0.8720 - val_accuracy: 0.5000
Epoch 36/50
2/2 [==============================] - 2s 553ms/step - loss: 1.2694e-04 - accuracy: 1.0000 - val_loss: 0.8847 - val_accuracy: 0.5000
Epoch 37/50
2/2 [==============================] - 2s 514ms/step - loss: 8.6252e-05 - accuracy: 1.0000 - val_loss: 0.8977 - val_accuracy: 0.5000
Epoch 38/50
2/2 [==============================] - 2s 520ms/step - loss: 2.6762e-04 - accuracy: 1.0000 - val_loss: 0.9115 - val_accuracy: 0.5000
Epoch 39/50
2/2 [==============================] - 2s 542ms/step - loss: 8.1350e-05 - accuracy: 1.0000 - val_loss: 0.9258 - val_accuracy:  0.5000
Epoch 40/50
2/2 [==============================] - 2s 506ms/step - loss: 8.0961e-05 - accuracy: 1.0000 - val_loss: 0.9405 - val_accuracy: 0.5000
Epoch 41/50
2/2 [==============================] - 2s 526ms/step - loss: 6.6102e-05 - accuracy: 1.0000 - val_loss: 0.9555 - val_accuracy: 0.5000
Epoch 42/50
2/2 [==============================] - 2s 549ms/step - loss: 1.1529e-04 - accuracy: 1.0000 - val_loss: 0.9707 - val_accuracy: 0.5000
Epoch 43/50
2/2 [==============================] - 2s 528ms/step - loss: 6.1373e-05 - accuracy: 1.0000 - val_loss: 0.9864 - val_accuracy: 0.5000
Epoch 44/50
2/2 [==============================] - 2s 516ms/step - loss: 7.2809e-05 - accuracy: 1.0000 - val_loss: 1.0025 - val_accuracy: 0.5000
Epoch 45/50
2/2 [==============================] - 2s 513ms/step - loss: 5.9504e-05 - accuracy: 1.0000 - val_loss: 1.0191 - val_accuracy: 0.5000
Epoch 46/50
2/2 [==============================] - 2s 515ms/step - loss: 6.1622e-05 - accuracy: 1.0000 - val_loss: 1.0361 - val_accuracy: 0.5000
Epoch 47/50
2/2 [==============================] - 2s 525ms/step - loss: 7.7296e-05 - accuracy: 1.0000 - val_loss: 1.0534 - val_accuracy: 0.5000
Epoch 48/50
2/2 [==============================] - 2s 512ms/step - loss: 4.5088e-05 - accuracy: 1.0000 - val_loss: 1.0711 - val_accuracy: 0.5000
Epoch 49/50
2/2 [==============================] - 2s 532ms/step - loss: 1.1449e-04 - accuracy: 1.0000 - val_loss: 1.0887 - val_accuracy: 0.5000
Epoch 50/50
2/2 [==============================] - 2s 516ms/step - loss: 6.0932e-05 - accuracy: 1.0000 - val_loss: 1.1071 - val_accuracy: 0.5000

Поскольку у меня 2 класса, мой учитель сказал, что точность проверки 0,5 означает, что она полностью случайна.
Мои изображения имеют формат изображений .jpg размером 50x50 на Google Диске. Может ли это быть проблемой, поскольку мой текущий размер изображения — 50x50? Но когда я бегу

Код: Выделить всё

import matplotlib.pyplot as plt

plt.figure(figsize=(10, 10))
for images, labels in train_ds.take(1):
for i in range(9):
ax = plt.subplot(3, 3, i + 1)
plt.imshow(images[i].numpy().astype("uint8"))
plt.title(int(labels[i]))
plt.axis("off")

Изображения правильные, так как все изображение показано и четко.
Я попробовал изменить начальное число на случайное число. Код поставляется с увеличением данных, и model.fit() должен автоматически перетасовывать изображения (если я правильно понял онлайн-сайты).
Мой учитель тоже не знает, в чем дело. Есть решения?
Это набор данных
https://www.kaggle.com/datasets/simranv ... st-dataset
Я использовал только два набора данных: тлю и совку.

Подробнее здесь: https://stackoverflow.com/questions/734 ... g-accuracy