Перенос обучения с ResNet50 - хорошие результаты при валидации, но плохие результаты в производстве

Я обучаю модель бинарной классификации примерно с 8000 изображений, примерно по 4000 на класс, я переношу обучение на resnet50, замораживаю все слои и получаю результаты: val_loss: 0,0340 - val_acc: 0,9890

Но когда я провожу тест на модели, я получаю почти случайные результаты с очень высокой вероятностью ...

Я не понимаю, какой в ​​этом смысл? В конце концов, модель не изучала изображения проверки, так как же существует такой большой разрыв между результатами на графике и результатами в производстве?

model.summary():
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
resnet50 (Model)             (None, 2048)              23587712  
_________________________________________________________________
dense (Dense)                (None, 2)                 4098      
=================================================================
Total params: 23,591,810
Trainable params: 4,098
Non-trainable params: 23,587,712
_________________________________________________________________

(Плотный - активация softmax)

Вот весь код:

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
%matplotlib inline 
import cv2
import os
from tensorflow.python.keras.applications import ResNet50
from tensorflow.python.keras.models import Sequential
from tensorflow.python.keras.layers import Dense
from tensorflow.python.keras import optimizers
from keras.applications.resnet50 import preprocess_input
from keras.preprocessing.image import ImageDataGenerator
from tensorflow.python.keras.callbacks import EarlyStopping, ModelCheckpoint

NUM_CLASSES = 2

CHANNELS = 3

IMAGE_RESIZE = 224
RESNET50_POOLING_AVERAGE = 'avg'
DENSE_LAYER_ACTIVATION = 'softmax'
OBJECTIVE_FUNCTION = 'categorical_crossentropy'

LOSS_METRICS = ['accuracy']

NUM_EPOCHS = 10
EARLY_STOP_PATIENCE = 3

STEPS_PER_EPOCH_TRAINING = 10
STEPS_PER_EPOCH_VALIDATION = 10

BATCH_SIZE_TRAINING = 100
BATCH_SIZE_VALIDATION = 100

# Using 1 to easily manage mapping between test_generator & prediction for submission preparation
BATCH_SIZE_TESTING = 1


model = Sequential()

model.add(ResNet50(include_top = False, pooling = RESNET50_POOLING_AVERAGE, weights = 'imagenet'))

model.add(Dense(NUM_CLASSES, activation = DENSE_LAYER_ACTIVATION))

model.layers[0].trainable = False

sgd = optimizers.SGD(lr = 0.01, decay = 1e-6, momentum = 0.9, nesterov = True)
model.compile(optimizer = sgd, loss = OBJECTIVE_FUNCTION, metrics = LOSS_METRICS)

image_size = IMAGE_RESIZE

data_generator = ImageDataGenerator(preprocessing_function=preprocess_input)


train_generator = data_generator.flow_from_directory(
        './datasets/Helmet_verification_v2/train',
        target_size=(image_size, image_size),
        batch_size=BATCH_SIZE_TRAINING,
        class_mode='categorical')

validation_generator = data_generator.flow_from_directory(
        './datasets/Helmet_verification_v2/validation',
        target_size=(image_size, image_size),
        batch_size=BATCH_SIZE_VALIDATION,
        class_mode='categorical')

output_dir = "./models/working"
output_file = output_dir + "/best.hdf5"
if not os.path.exists(output_dir):
    print("create folder: {}".format(output_dir))
    os.makedirs(output_dir)

cb_early_stopper = EarlyStopping(monitor = 'val_loss', patience = EARLY_STOP_PATIENCE)
cb_checkpointer = ModelCheckpoint(filepath = output_file, monitor = 'val_loss', save_best_only = True, mode = 'auto')

fit_history = model.fit_generator(
        train_generator,
        steps_per_epoch=STEPS_PER_EPOCH_TRAINING,
        epochs = NUM_EPOCHS,
        validation_data=validation_generator,
        validation_steps=STEPS_PER_EPOCH_VALIDATION,
        callbacks=[cb_checkpointer, cb_early_stopper]
)
model.load_weights(output_file)

Тестовое задание:

test_generator = data_generator.flow_from_directory(
directory = './datasets/Helmet_verification_v2/test',
target_size = (image_size, image_size),
batch_size = BATCH_SIZE_TESTING,
class_mode = None,
shuffle = False,
seed = 123
)
test_generator.reset()

pred = model.predict_generator(test_generator, steps = len(test_generator), verbose = 1)

predicted_class_indices = np.argmax(pred, axis = 1)

некоторые примеры данных: