应该是梯度爆炸了，可以试试减小学习率，加正则，使用BN层

batch_size设为1且存在batch_norm层是不是也可能出现这个问题？

您好，如果您在使用中遇到类似的问题，方便的话请您提供下详细代码，以便解决～感谢您的支持！

def prediction_program():
    hidden_size = 7
    image_shape =[3, 300, 300]
    img = fluid.layers.data(name='img', shape=image_shape, dtype='float32',lod_level=1)
    module11, module13, module14, module15, module16, module17 = mobile_net(img, scale=1.0)
    tmp0 = fluid.layers.fc(input=[module11,module13,module14,module15,module16,module17], size = 15,act='relu')
    input_forward_proj = fluid.layers.fc(input=tmp0,
                                         size=hidden_size * 4,
                                         bias_attr=False)
    forward, _ = fluid.layers.dynamic_lstm(
        input=input_forward_proj,
        size=hidden_size * 4)
    forward_last = fluid.layers.sequence_last_step(input=forward)
    prediction = fluid.layers.fc(input=forward_last, size = 1,act='relu')
    return  prediction
def train_program():
    prediction = prediction_program()    
    y = fluid.layers.data(name='y',shape = [1],dtype='float32',lod_level=1)
    gt1 = fluid.layers.abs(prediction-y)/y
    gt2 = fluid.layers.square_error_cost(y,prediction)
    gt3 = gt2/fluid.layers.square(y)
    loss1 = fluid.layers.mean(gt1)
    loss2 = fluid.layers.mean(gt2)
    loss3 = fluid.layers.mean(gt3)
    return [loss3,loss2,prediction,loss1]
def optimizer_program():
    return fluid.optimizer.Adam(learning_rate=0.01
                               # ,regularization=fluid.regularizer.L2DecayRegularizer(regularization_coeff=0.05)
                               )

module11-17来自https://github.com/PaddlePaddle/models/blob/develop/fluid/object_detection/mobilenet_ssd.py

谢谢，就是学习率的问题，改小了就好了

我也遇到这种情况了，不知道是不是接近0的数值太多了，不知道这种情况怎么处理

原来如此