1

我想用联邦学习微调预训练模型，所以我这样做：

def create_keras_model():
    baseModel = tf.keras.models.load_model(path\to\model)
    headModel = baseModel.output
    model_output = tf.keras.layers.Dense(3)(headModel)
    model = tf.keras.Model(inputs=baseModel.input, outputs=model_output)
    for layer in baseModel.layers:
        layer.trainable = False
    return model

state = iterative_process.initialize()
keras_model = create_keras_model()
state = tff.learning.state_with_new_model_weights(
    state,
    trainable_weights=[v.numpy() for v in keras_model.trainable_weights],
    non_trainable_weights=[
        v.numpy() for v in keras_model.non_trainable_weights
    ])

evaluation = tff.learning.build_federated_evaluation(model_fn)

这是训练循环：

for round_num in range(1, NUM_ROUNDS):
    state, _ = iterative_process.next(state, train_data)
    test_metrics = evaluation(state.model, test_data)
    print(test_metrics))

问题是测试精度仍然保持不变，并且始终没有增加：

round  1, metrics=OrderedDict([('categorical_accuracy', 0.67105263), ('loss', 0.8680933)])
round  2, metrics=OrderedDict([('categorical_accuracy', 0.67105263), ('loss', 0.836558)])
round  3, metrics=OrderedDict([('categorical_accuracy', 0.67105263), ('loss', 0.82953715)])
round  4, metrics=OrderedDict([('categorical_accuracy', 0.67105263), ('loss', 0.82713753)])
round  5, metrics=OrderedDict([('categorical_accuracy', 0.67105263), ('loss', 0.82613766)])
round  6, metrics=OrderedDict([('categorical_accuracy', 0.67105263), ('loss', 0.8256878)])
round  7, metrics=OrderedDict([('categorical_accuracy', 0.67105263), ('loss', 0.82548285)])
round  8, metrics=OrderedDict([('categorical_accuracy', 0.67105263), ('loss', 0.825384)])
round  9, metrics=OrderedDict([('categorical_accuracy', 0.67105263), ('loss', 0.825332)])

我想了解原因，如果有其他方法可以做到这一点？知道我的数据集是一个具有 3 类的图像数据集。

4

0 回答 0