迁移学习模型推理模式下无法学习的原因求助

我正在完成Kaggle上的「Yum or Yuck: Butterfly Mimics 2022」物种分类任务，由于数据集图像数量较少，决定采用迁移学习方案。模型构建代码如下：

inputs = tf.keras.layers.Input(shape=(224, 224, 3))
base_model = tf.keras.applications.resnet50.ResNet50(
    input_shape=(224,224,3),
    include_top=False,
    weights="imagenet")
for layer in base_model.layers:
    layer.trainable = False

x = base_model(inputs, training=False)
x = tf.keras.layers.GlobalAveragePooling2D()(x)
x = tf.keras.layers.Dropout(0.3)(x)
x = tf.keras.layers.Dense(1024, activation="relu")(x)
x = tf.keras.layers.Dropout(0.3)(x)
x = tf.keras.layers.Dense(512, activation="relu")(x)
x = tf.keras.layers.Dropout(0.3)(x)
x = tf.keras.layers.Dense(64, activation="relu")(x)
output = tf.keras.layers.Dense(6, activation="softmax")(x)
model = tf.keras.Model(inputs=inputs, outputs=output)

按照Keras迁移学习指南，我冻结了ResNet50的所有层，并在推理模式（training=False）下训练模型，但模型始终无法正常收敛，训练近200个epoch仍无明显学习迹象。训练配置及部分epoch结果如下：

model.compile(
    optimizer=tf.keras.optimizers.Adam(),
    loss="categorical_crossentropy",
    metrics="accuracy",
)
stop_early = tf.keras.callbacks.EarlyStopping(
    monitor='val_loss',
    min_delta=0.0001,
    patience=20,
    restore_best_weights=True
)
history = model.fit(train_generator,
          validation_data = val_generator,
          epochs = 200,
         callbacks=[stop_early])

22/22 [==============================] - 19s 442ms/step - loss: 1.9317 - accuracy: 0.1794 - val_loss: 1.8272 - val_accuracy: 0.1618
Epoch 2/200
22/22 [==============================] - 9s 398ms/step - loss: 1.8250 - accuracy: 0.1882 - val_loss: 1.7681 - val_accuracy: 0.2197
Epoch 3/200
22/22 [==============================] - 9s 402ms/step - loss: 1.7927 - accuracy: 0.2294 - val_loss: 1.7612 - val_accuracy: 0.2139
Epoch 4/200
22/22 [==============================] - 9s 424ms/step - loss: 1.7930 - accuracy: 0.2000 - val_loss: 1.7640 - val_accuracy: 0.2139
Epoch 5/200
22/22 [==============================] - 9s 391ms/step - loss: 1.7872 - accuracy: 0.2132 - val_loss: 1.7489 - val_accuracy: 0.3121
Epoch 6/200
22/22 [==============================] - 9s 389ms/step - loss: 1.7700 - accuracy: 0.2574 - val_loss: 1.7378 - val_accuracy: 0.2543
Epoch 7/200
22/22 [==============================] - 9s 396ms/step - loss: 1.7676 - accuracy: 0.2353 - val_loss: 1.7229 - val_accuracy: 0.3064
Epoch 8/200
22/22 [==============================] - 9s 427ms/step - loss: 1.7721 - accuracy: 0.2353 - val_loss: 1.7225 - val_accuracy: 0.2948
Epoch 9/200
22/22 [==============================] - 9s 399ms/step - loss: 1.7522 - accuracy: 0.2588 - val_loss: 1.7267 - val_accuracy: 0.2948
Epoch 10/200
22/22 [==============================] - 9s 395ms/step - loss: 1.7434 - accuracy: 0.2735 - val_loss: 1.7151 - val_accuracy: 0.2948
Epoch 11/200
22/22 [==============================] - 9s 391ms/step - loss: 1.7500 - accuracy: 0.2632 - val_loss: 1.7083 - val_accuracy: 0.3064
Epoch 12/200
22/22 [==============================] - 9s 425ms/step - loss: 1.7307 - accuracy: 0.2721 - val_loss: 1.6899 - val_accuracy: 0.3179
Epoch 13/200
22/22 [==============================] - 9s 407ms/step - loss: 1.7439 - accuracy: 0.2794 - val_loss: 1.7045 - val_accuracy: 0.2948
Epoch 14/200
22/22 [==============================] - 9s 404ms/step - loss: 1.7376 - accuracy: 0.2706 - val_loss: 1.7118 - val_accuracy: 0.2659
Epoch 15/200
22/22 [==============================] - 9s 419ms/step - loss: 1.7588 - accuracy: 0.2647 - val_loss: 1.6684 - val_accuracy: 0.3237
Epoch 16/200
22/22 [==============================] - 9s 394ms/step - loss: 1.7289 - accuracy: 0.2824 - val_loss: 1.6733 - val_accuracy: 0.3064
Epoch 17/200
22/22 [==============================] - 9s 387ms/step - loss: 1.7184 - accuracy: 0.2809 - val_loss: 1.7185 - val_accuracy: 0.2659
Epoch 18/200
22/22 [==============================] - 9s 408ms/step - loss: 1.7242 - accuracy: 0.2765 - val_loss: 1.6961 - val_accuracy: 0.2717
Epoch 19/200
22/22 [==============================] - 9s 424ms/step - loss: 1.7218 - accuracy: 0.2853 - val_loss: 1.6757 - val_accuracy: 0.3006
Epoch 20/200
22/22 [==============================] - 9s 396ms/step - loss: 1.7248 - accuracy: 0.2882 - val_loss: 1.6716 - val_accuracy: 0.3064
Epoch 21/200
22/22 [==============================] - 9s 401ms/step - loss: 1.7134 - accuracy: 0.2838 - val_loss: 1.6666 - val_accuracy: 0.2948
Epoch 22/200
22/22 [==============================] - 9s 393ms/step - loss: 1.7140 - accuracy: 0.2941 - val_loss: 1.6427 - val_accuracy: 0.3064

只有解冻ResNet50的层并关闭推理模式（training=True），模型才能正常学习。我用EfficientNet做了相同场景的测试，结果和ResNet50一致；但使用Xception时，冻结所有层并开启推理模式却能正常收敛。

尽管这三个模型都包含BatchNorm层，但表现差异巨大，我无法理解为何ResNet50和EfficientNet必须关闭推理模式才能正常学习，恳请各位提供分析思路。

补充训练结果图 #

• ResNet50训练结果：
  

• Xception训练结果：
  

内容的提问来源于stack exchange，提问作者DPM