Created by: Aurelius84

1.背景

目前Paddle框架支持动态图组网训练，为了更易于部署，可以将动态图转为静态图，更方便基于图做进一步的优化和上线。动态图下的if/else语法为python原生语法，转静态图时，需要解析判断、并适应转为paddle的控制流API fluid.layers.cond。

2.功能支持

2.1 支持简单if/else语句转写

动态图代码：

@dygraph_to_static_output
def dyfunc_with_ifElse(x_v):
    if (x_v).numpy()[0] > 5:
        x_v = x_v - 1
    else:
        x_v = x_v + 1
    return x_v

转换后的静态图代码：

def true_fn(x_v):
    x_v = x_v - 1
    return ()

def false_fn(x_v):
    x_v = x_v + 1
    return ()

def dyfunc_with_ifElse2(x_v):
    x_v = fluid.layers.cond((x_v[0]) > 5), lambda : true_fn(x_v), lambda : false_fn(x_v))
    return x_v

2.2 支持嵌套的if/else

动态图代码：

@dygraph_to_static_output
def dyfunc_with_ifElse(x_v):
    batch_size = 10
    shape = [batch_size, 16]
    bias = fluid.layers.fill_constant([shape[-1]], dtype='float32', value=1)
    if fluid.layers.mean(x_v).numpy()[0] < 0:
        x_v = x_v + bias
        w = fluid.layers.fill_constant([shape[-1]], dtype='float32', value=10)
        new_var = True
        # 嵌套的if，且缺省了else
        if x_v.numpy()[0] < 10:
            tmp = x_v * w
            x_v = fluid.layers.relu(tmp)
    else:
        new_var = False
        x_v = x_v - bias
    return x_v

转换后的静态图代码：

# 对应最内层的if
def true_fn_0(w, x_v):
    tmp = x_v * w
    x_v = fluid.layers.relu(tmp)
    return x_v

# 对应最内层缺省else，函数内部没有实际处理逻辑
# 但由于layers.cond对返回值嵌套结构需一致，故需要返回与true_fn一致的变量
def false_fn_0(x_v):
    return x_v

# 对应最外层的if
def true_fn_1(True, bias, shape, x_v):
    x_v = x_v + bias
    w = fluid.layers.fill_constant([shape[-1]], dtype='float32', value=10)
    new_var = True
    x_v = fluid.layers.cond((x_v[0] < 10), lambda : true_fn_0(w, x_v), lambda : false_fn_0(x_v))
    return (new_var, x_v)

# 对应最外层的else
def false_fn_1(False, bias, x_v):
    new_var = False
    x_v = x_v - bias
    return (new_var, x_v)

def dyfunc_with_ifElse(x_v):
    batch_size = 10
    shape = [batch_size, 16]
    bias = fluid.layers.fill_constant([shape[-1]], dtype='float32', value=1)
    (new_var, x_v) = fluid.layers.cond((fluid.layers.mean(x_v)[0] < 0), lambda : true_fn_1(True, bias, shape, x_v), lambda : false_fn_1(False, bias, x_v))
    return x_v

2.3 支持控制流分支函数输入和输出参数自动解析

见2.2中构建true_fn和false_fn时对输入参数和输出参数的解析

3. What's Next?

• 分支函数名的友好管理方案
• AST转回可调用python函数的功能实现
• 各子逻辑函数的单元测试