困扰我两天了,mobilenet问题,谁能帮我解答下,谢谢了

你好,这个问题困扰我两天了,为什么我用官方maixpy的mobilenet训练脚本训练模型会出现:
====>…] - ETA: 6s - loss: 1.4539 - accuracy: 0.5451
274/283 [============================>.] - ETA: 5s - loss: 1.4517 - accuracy: 0.5458
275/283 [============================>.] - ETA: 5s - loss: 1.4502 - accuracy: 0.5461
276/283 [============================>.] - ETA: 4s - loss: 1.4497 - accuracy: 0.5460
277/283 [============================>.] - ETA: 3s - loss: 1.4489 - accuracy: 0.5461
278/283 [============================>.] - ETA: 3s - loss: 1.4465 - accuracy: 0.5467
279/283 [============================>.] - ETA: 2s - loss: 1.4455 - accuracy: 0.5469
280/283 [============================>.] - ETA: 1s - loss: 1.4437 - accuracy: 0.5476
281/283 [============================>.] - ETA: 1s - loss: 1.4408 - accuracy: 0.5485
282/283 [============================>.] - ETA: 0s - loss: 1.4398 - accuracy: 0.5488
283/283 [==============================] - 205s 723ms/step - loss: 1.4385 - accuracy: 0.5490 - val_loss: 3.4533 - val_accuracy: 0.0804

val_acc严重不正常,但是经过尝试去掉

for layer in model.layers[:86]:
layer.trainable=False
for layer in model.layers[86:]:
layer.trainable=True
后训练正常,模型测试也正常,但是转为kmodel并烧录后出现不管输入什么都输出同一个标签,而且所有标签的概率不管怎么输入变动范围都很小(0.0几)