Почему моя модель в tensorflow.js не предсказывает правильно, а в Python — нет? [Передача обучения Resnet-50] ⇐ Python

[Anonymous]

Я проверил около 20 изображений. Те же для Python и те же для tensorflow.js с очень похожей предварительной обработкой. Он продолжает давать мне 0-й класс.
Вот мой код обучения Python:
import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.keras.preprocessing import image
from tensorflow.keras.applications.resnet50 import preprocess_input
import numpy as np

tf.random.set_seed(123)

img_height, img_width = 224, 224
batch_size = 32
data_dir = '/content/drive/MyDrive/AI2/Processed'

# random horizontal flips, random rotation, color jitter, normalize
data_augmentation = tf.keras.Sequential([
layers.experimental.preprocessing.RandomFlip('horizontal'),
layers.experimental.preprocessing.RandomRotation(0.2),
layers.experimental.preprocessing.RandomZoom(0.2),
layers.experimental.preprocessing.RandomContrast(0.2),
layers.experimental.preprocessing.RandomTranslation(height_factor=(-0.1, 0.1), width_factor=(-0.1, 0.1)),
])

def one_hot_encode(image, label):
label = tf.one_hot(label, 5)
return image, label

def main():
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
data_dir,
validation_split=0.2,
subset="training",
seed=123,
image_size=(img_height, img_width),
)

val_ds = tf.keras.preprocessing.image_dataset_from_directory(
data_dir,
validation_split=0.2,
subset="validation",
seed=123,
image_size=(img_height, img_width),
)

class_names = train_ds.class_names
print(class_names)

class_count = [0] * len(class_names)
for images, labels in train_ds.unbatch().as_numpy_iterator():
class_count[int(labels)] += 1

train_ds = train_ds.map(lambda x, y: (data_augmentation(x, training=True), y))
train_ds = train_ds.map(one_hot_encode).prefetch(buffer_size=tf.data.AUTOTUNE)

val_ds = val_ds.map(one_hot_encode).prefetch(buffer_size=tf.data.AUTOTUNE)

base_model = tf.keras.applications.ResNet50(
include_top=False,
weights='imagenet',
input_shape=(img_height, img_width, 3)
)

# Freeze all layers in the base model
base_model.trainable = False

# Create the model architecture
inputs = tf.keras.Input(shape=(img_height, img_width, 3))
x = base_model(inputs, training=False)
x = layers.GlobalAveragePooling2D()(x)
x = layers.Dense(1024, activation='relu')(x)
outputs = layers.Dense(len(class_names), activation='softmax')(x)

model = tf.keras.Model(inputs, outputs)

# Freeze all layers except the last one
# for layer in model.layers[:-1]:
# layer.trainable = False

# Unfreeze the last layer in the model
model.layers[-1].trainable = True

# Compile the model
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
loss='categorical_crossentropy',
metrics=['accuracy'])

# Train the model
class_weights = {i: 1. / count for i, count in enumerate(class_count)}
history = model.fit(
train_ds,
validation_data=val_ds,
epochs=10,
class_weight=class_weights
)

# Evaluate on validation set
loss, accuracy = model.evaluate(val_ds)
print(f"Validation accuracy: {accuracy * 100}%")

# Print the model summary
model.summary()

model.save('/content/drive/MyDrive/AI2/sampleModel.h5')

Затем я делаю следующее:
tensorflowjs_converter --input_format keras /Users/hb/Desktop/tfModel3/sampleModel.h5 /Users/hb/Desktop/tfModel3/sampleModelNew

Вот мой код Python для прогнозирования:
model = tf.keras.models.load_model('/content/drive/MyDrive/AI2/sampleModel.h5')

# Show the model architecture
model.summary()

img_height = 224
img_width = 224

# Load the image
img_path = '/content/drive/MyDrive/AI2/RealTest/bmw.jpg' # replace with your image path
img = image.load_img(img_path, target_size=(img_height, img_width))
img_array = image.img_to_array(img)

# Preprocess the image using the same preprocessing function used for ImageNet
img_array = preprocess_input(img_array)

img_array = np.expand_dims(img_array, axis=0)

# Make a prediction
predictions = model.predict(img_array)
predicted_class = np.argmax(predictions[0])

# Print the class name
print('Predicted class:', predicted_class)

И после всего этого у меня есть код React, использующий tensorflowjs, который не выдает ошибок для каждого изображения, которое я ему предоставляю. Модель успешно импортирована! Он дает прогноз без ошибок. Но каждый раз это неверный прогноз!:
const { state, dispatch } = useMyContext(); // Use the custom hook to access state and dispatch
const [model, setModel] = React.useState(null);

useEffect(() => {
async function loadModel() {
const model = await tf.loadLayersModel('/models/model.json');
setModel(model);
console.log('model loaded')
console.log('model', model);

model.summary();
}

loadModel()
}, []);
useEffect(() => {
if (!state.imageFiles) return;
if (state.imageFiles.length === 0) return;

const newImgFunc = async () => {
const img = new Image();
const imageFile = state.imageFiles[0];
img.src = await URL.createObjectURL(imageFile);
img.onload = async () => {
const class_names = ['Audi_A3_2021_Processed', 'BMW_1-Series_2016_Processed', 'Ford_Puma_2021_Processed', 'Nissan_Juke_2023_Processed', 'Nissan_Qashqai_2021_Processed'];

// Match preprocess_input from Keras for ResNet50
const mean = tf.tensor1d([123.68, 116.779, 103.939]);
const std = tf.tensor1d([58.393, 57.12, 57.375]); // This line can be removed if not used

let tensor = tf.browser.fromPixels(img)
.resizeBilinear([224, 224]) // Using bilinear, as its typically more accurate
.toFloat()
.sub(mean) // Subtract mean
.div(tf.scalar(255)); // Scale to [0,1]

const prediction = await model.predict(tensor.expandDims());
const predictions = prediction.dataSync();
const predictedClassIndex = predictions.indexOf(Math.max(...predictions));
console.log('Predicted Class:', class_names[predictedClassIndex]);
};
}

newImgFunc();
}, [state]);
const onDrop = useCallback((acceptedFiles: any[]) => {
acceptedFiles.forEach((file: Blob) => {
const reader = new FileReader()

reader.onabort = () => console.log('file reading was aborted')
reader.onerror = () => console.log('file reading has failed')
reader.onload = () => {
const binaryStr = reader.result
console.log(binaryStr)

const file = new File(
[binaryStr as BlobPart],
'imageName'

);

dispatch({ type: 'SET_IMAGE_FILES', payload: [file] });
}
reader.readAsArrayBuffer(file)
})

}, [])


Подробнее здесь: https://stackoverflow.com/questions/784 ... -python-it