Created by: oraoto

我在实现DeepDream，需要计算某个特征图对输入图像的梯度，所以使用了fluid.backward.calc_gradient，发现两个问题：

1. batch_norm无法计算梯度

x = fluid.layers.create_parameter(name="x0", shape=[1, 3, 5, 5], dtype='float32')

x2 = x * 2.0
    
x3 = fluid.layers.batch_norm(x2)

y = fluid.layers.mean(x3)

g = fluid.backward.calc_gradient(y, x) # 出错

错误信息：

EnforceNotMet                             Traceback (most recent call last)
<ipython-input-58-ce613e3efa1a> in <module>()
     11     y = fluid.layers.mean(x3)
     12 
---> 13     g = fluid.backward.calc_gradient(y, x)

/home/lyp/projects/paddle/.env/lib/python2.7/site-packages/paddle/fluid/backward.pyc in calc_gradient(targets, inputs, target_gradients, no_grad_set)
    729     _rename_grad_(block, fwd_op_num, grad_to_var, target_grad_map)
    730 
--> 731     _append_backward_vars_(block, fwd_op_num, grad_to_var, grad_info_map)
    732     prog.sync_with_cpp()
    733 

/home/lyp/projects/paddle/.env/lib/python2.7/site-packages/paddle/fluid/backward.pyc in _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map)
    390         # infer_shape and infer_type
    391         op_desc.infer_var_type(block.desc)
--> 392         op_desc.infer_shape(block.desc)
    393         # ncclInit dones't need to set data_type
    394         if op_desc.type() == 'ncclInit':

EnforceNotMet:  at [/paddle/paddle/fluid/operators/batch_norm_op.cc:325]
PaddlePaddle Call Stacks: 
0       0x7f26d1ee4376p paddle::platform::EnforceNotMet::EnforceNotMet(std::__exception_ptr::exception_ptr, char const*, int) + 486
1       0x7f26d247efb8p paddle::operators::BatchNormGradOp::InferShape(paddle::framework::InferShapeContext*) const + 2888
2       0x7f26d1f64691p paddle::framework::OpDesc::InferShape(paddle::framework::BlockDesc const&) const + 881
3       0x7f26d1f37915p pybind11::cpp_function::initialize<pybind11::cpp_function::initialize<void, paddle::framework::OpDesc, paddle::framework::BlockDesc const&, pybind11::name, pybind11::is_method, pybind11::sibling>(void (paddle::framework::OpDesc::*)(paddle::framework::BlockDesc const&) const, pybind11::name const&, pybind11::is_method const&, pybind11::sibling const&)::{lambda(paddle::framework::OpDesc const*, paddle::framework::BlockDesc const&)#1}, void, paddle::framework::OpDesc const*, paddle::framework::BlockDesc const&, pybind11::name, pybind11::is_method, pybind11::sibling>(pybind11::cpp_function::initialize<void, paddle::framework::OpDesc, paddle::framework::BlockDesc const&, pybind11::name, pybind11::is_method, pybind11::sibling>(void (paddle::framework::OpDesc::*)(paddle::framework::BlockDesc const&) const, pybind11::name const&, pybind11::is_method const&, pybind11::sibling const&)::{lambda(paddle::framework::OpDesc const*, paddle::framework::BlockDesc const&)#1}&&, void (*)(paddle::framework::OpDesc const*, paddle::framework::BlockDesc const&), pybind11::name const&, pybind11::is_method const&, pybind11::sibling const&)::{lambda(pybind11::detail::function_call&)#3}::_FUN(pybind11::detail::function_call) + 213
4       0x7f26d1ef346cp pybind11::cpp_function::dispatcher(_object*, _object*, _object*) + 2540
5       0x7f271bc071a8p PyEval_EvalFrameEx + 25384

2.fluid.layers.slice某些情况会无法计算梯度

没有slice时能正常求梯度：

x = fluid.layers.create_parameter(name="x1", shape=[1, 3, 5, 5], dtype='float32')

x2 = x * 2.0
    
y = fluid.layers.mean(x2)

g = fluid.backward.calc_gradient(y, x)

加了一层slice，无法求梯度（返回None）：

x = fluid.layers.create_parameter(name="x2", shape=[1, 3, 5, 5], dtype='float32') # (1, 3, 5, 5)

x2 = x * 2.0
    
s = fluid.layers.slice(x2, axes=[0, 1], starts=[0, 0], ends=[1, 2]) # (1, 2, 5, 5)
    
y = fluid.layers.mean(s) # (1, )

g = fluid.backward.calc_gradient(y, x2)  # g is None

再加一层Conv2D，g不为None，但是梯度都是0：

x = fluid.layers.create_parameter(name="x3", shape=[1, 3, 5, 5], dtype='float32')

x2 = x * 2.0
    
conv = fluid.layers.conv2d(x2, num_filters=3, filter_size=3, padding=1, act='relu') # (1, 3, 5, 5)
    
s = fluid.layers.slice(conv, axes=[0, 1], starts=[0, 0], ends=[1, 2]) # (1, 2, 5, 5)
   
y = fluid.layers.mean(s)

g = fluid.backward.calc_gradient(y, x)