Created by: jshower

我在使用paddle的fluid版本的例子时，发现存在的一个问题就是没有做类型的兼容性处理，比如说我定义一个全连接层，里面使用了NormalInitializer，scale设为1，如下所示。 fc1 = fluid.layers.fc(input=drop, size=4096, act=None, bias_attr=fluid.ParamAttr(initializer=NormalInitializer(loc=0.0, scale=1, seed=0))) 以修改Paddle/python/paddle/fluid/tests/book/test_image_classification.py 的89行为例，结果会出现：

Traceback (most recent call last):
  File "test_image_classification.py", line 252, in test_vgg_cpu
    main('vgg', use_cuda=False)
  File "test_image_classification.py", line 237, in main
    train(net_type, use_cuda, save_dirname, is_local)
  File "test_image_classification.py", line 105, in train
    net = vgg16_bn_drop(images)
  File "test_image_classification.py", line 89, in vgg16_bn_drop
    fc1 = fluid.layers.fc(input=drop, size=4096, act=None, bias_attr=fluid.ParamAttr(initializer=NormalInitializer(loc=0.0, scale=1, seed=0)))
  File "/usr/local/lib/python2.7/dist-packages/paddle/fluid/layers/nn.py", line 193, in fc
    pre_activation = helper.append_bias_op(pre_bias, dim_start=num_flatten_dims)
  File "/usr/local/lib/python2.7/dist-packages/paddle/fluid/layer_helper.py", line 385, in append_bias_op
    attr=bias_attr, shape=size, dtype=input_var.dtype, is_bias=True)
  File "/usr/local/lib/python2.7/dist-packages/paddle/fluid/layer_helper.py", line 312, in create_parameter
    dtype=dtype, shape=shape, **attr.to_kwargs(with_initializer=True))
  File "/usr/local/lib/python2.7/dist-packages/paddle/fluid/framework.py", line 810, in create_parameter
    kwargs['initializer'](param, self)
  File "/usr/local/lib/python2.7/dist-packages/paddle/fluid/initializer.py", line 239, in __call__
    "seed": self._seed
  File "/usr/local/lib/python2.7/dist-packages/paddle/fluid/framework.py", line 834, in prepend_op
    op = Operator(self, op_desc, *args, **kwargs)
  File "/usr/local/lib/python2.7/dist-packages/paddle/fluid/framework.py", line 484, in __init__
    self.desc.check_attrs()
EnforceNotMet: Cannot get attribute std by type f, its type is i at [/paddle/Paddle/paddle/fluid/framework/attribute.h:131]

而我如果将scale=1改成scale=1.0的话，这个问题就解决了。 类似的还有 fluid.optimizer.Momentum(momentum=0, learning_rate=1e-3） 会报类似的错误，将momentum改成0.0就不会报错。 这是一个比较明显的没有做类型兼容处理的问题。