Fix Using "isinstance" in Loop, test=develop

[zhhsplendid]

PR types

Bug fixes

PR changes

APIs

Describe

Fix a bug that used in PaddleGAN model which used isinstance in a for loop (the code is located at https://github.com/PaddlePaddle/PaddleGAN/blob/32c59d5539de9182b8307d41ba7aad7de76f1c28/ppgan/models/generators/deoldify.py#L35)

The reason of the bug is that the type of function isinstance is treated as a variable in loop_vars. For example, before this bug fix, if you print(model.forward.code) for the test_sequential_layer unittest of this PR, you will get code like below. Note that AddAttrLayer is a type but here our code treated it as a variable.

def forward(self, x):
    res = x
    AddAttrLayer = paddle.jit.dy2static.data_layer_not_check(name=
        'AddAttrLayer', shape=[-1], dtype='float32')
    __for_loop_var_index_0 = 0
    __for_loop_var_len_0 = paddle.jit.dy2static.convert_len(self.layers)

    def for_loop_condition_0(AddAttrLayer, __for_loop_var_index_0, res, x):
        return __for_loop_var_index_0 + 1 <= __for_loop_var_len_0

    def for_loop_body_0(AddAttrLayer, __for_loop_var_index_0, res, x):

        def true_fn_1(__for_loop_var_index_0, self, x):
            self.layers[__for_loop_var_index_0].attr = x
            return

        def false_fn_1():
            return
        paddle.jit.dy2static.convert_ifelse(isinstance(self.layers[
            __for_loop_var_index_0], AddAttrLayer), true_fn_1, false_fn_1,
            (__for_loop_var_index_0, self, x), (), ())
        res = paddle.jit.dy2static.convert_call(self.layers[
            __for_loop_var_index_0])(res)
        __for_loop_var_index_0 += 1
        return AddAttrLayer, __for_loop_var_index_0, res, x
    [AddAttrLayer, __for_loop_var_index_0, res, x
        ] = paddle.jit.dy2static.convert_while_loop(for_loop_condition_0,
        for_loop_body_0, [AddAttrLayer, __for_loop_var_index_0, res, x])
    return res

Solution: I recorded those variables in isinstance as types, then removed those type variables in loop_vars. Now the translated code for above example is:

def forward(self, x):
    res = x
    __for_loop_var_index_0 = 0
    __for_loop_var_len_0 = paddle.jit.dy2static.convert_len(self.layers)

    def for_loop_condition_0(res, __for_loop_var_index_0, x):
        return __for_loop_var_index_0 + 1 <= __for_loop_var_len_0

    def for_loop_body_0(res, __for_loop_var_index_0, x):

        def true_fn_1(__for_loop_var_index_0, self, x):
            self.layers[__for_loop_var_index_0].attr = x
            return

        def false_fn_1():
            return
        paddle.jit.dy2static.convert_ifelse(isinstance(self.layers[
            __for_loop_var_index_0], AddAttrLayer), true_fn_1, false_fn_1,
            (__for_loop_var_index_0, self, x), (), ())
        res = paddle.jit.dy2static.convert_call(self.layers[
            __for_loop_var_index_0])(res)
        __for_loop_var_index_0 += 1
        return res, __for_loop_var_index_0, x
    [res, __for_loop_var_index_0, x] = paddle.jit.dy2static.convert_while_loop(
        for_loop_condition_0, for_loop_body_0, [res, __for_loop_var_index_0, x]
        )
    return res

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[liym27]

LGTM

[Aurelius84]

LGTM