PaddlePaddle在优化方法处报错 EnforceNotMet: Enforce failed

• 关键字：rank，优化方法，损失函数

• 问题描述：执行定义损失函数代码后，再执行优化方法就报错，提示执行失败，dy_dims.size():1 != rank:2。

• 报错信息：

/usr/local/lib/python3.5/dist-packages/paddle/fluid/optimizer.py in minimize(self, loss, startup_program, parameter_list, no_grad_set)
    253         """
    254         params_grads = append_backward(loss, parameter_list, no_grad_set,
--> 255                                        [error_clip_callback])
    256 
    257         params_grads = sorted(params_grads, key=lambda x: x[0].name)

/usr/local/lib/python3.5/dist-packages/paddle/fluid/backward.py in append_backward(loss, parameter_list, no_grad_set, callbacks)
    588     _rename_grad_(root_block, fwd_op_num, grad_to_var, {})
    589 
--> 590     _append_backward_vars_(root_block, fwd_op_num, grad_to_var, grad_info_map)
    591 
    592     program.current_block_idx = current_block_idx

/usr/local/lib/python3.5/dist-packages/paddle/fluid/backward.py in _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map)
    424         # infer_shape and infer_type
    425         op_desc.infer_var_type(block.desc)
--> 426         op_desc.infer_shape(block.desc)
    427         # ncclInit dones't need to set data_type
    428         if op_desc.type() == 'ncclInit':

EnforceNotMet: Enforce failed. Expected dy_dims.size() == rank, but received dy_dims.size():1 != rank:2.
Input(Y@Grad) and Input(X) should have the same rank. at [/paddle/paddle/fluid/operators/cross_entropy_op.cc:82]
PaddlePaddle Call Stacks: 

• 问题复现：定义一个交叉熵损失函数，直接使用这个损失函数传给优化方法，再执行到这一行代码就出现这个问题。错误代码如下：

cost = fluid.layers.cross_entropy(input=model, label=label)

optimizer = fluid.optimizer.AdamOptimizer(learning_rate=0.001)
opts = optimizer.minimize(cost)

• 问题解决：训练是一个Batch进行训练的，所以计算的损失值也是计算一个Batch的损失值。优化方法参数使用的是一个平均的损失函数，所以不能直接使用损失函数，还需要对损失函数求平均值。正确代码如下：

cost = fluid.layers.cross_entropy(input=model, label=label)
avg_cost = fluid.layers.mean(cost)

optimizer = fluid.optimizer.AdamOptimizer(learning_rate=0.001)
opts = optimizer.minimize(avg_cost)

• 问题拓展：如果在训练的时候，fetch_list参数使用的是cost，而不是avg_cost的话，训练输出的也会是一个Batch的损失值。所以在训练的时候，fetch_list参数的值最好使用avg_cost，输出的是平均损失值，从而更方便观察训练情况。