建议exporter.py中translate_graph函数采用deepcopy(model)避免deepcopy出现Dynamic-to-Static信息

[GuoQuanhao]

在使用add_graph后使用deepcopy出现Dynamic-to-Static

import paddle
from copy import deepcopy

from visualdl import LogWriter

net = paddle.nn.Linear(10, 10)
with LogWriter(logdir="./log/graph_test/") as writer:
    writer.add_graph(
        model=net,
        input_spec=[paddle.static.InputSpec([-1, 10], 'float32')],
        verbose=True)

B = deepcopy(net)

Tue Mar 21 17:23:47 Dynamic-to-Static INFO: (Level -1) Not recommend to deepcopy 'Linear' decorated with @to_static, it has side effect that will rollback into original state before @to_static. Please deepcopy 'Linear' before applying @to_static.
-------------------------Graph Summary--------------------------
total operators: 4	total layers:1
--------------------  --------------------  --------------------  
Name                  Type                  Count                 
--------------------  --------------------  --------------------  
feed                  operator              1                     
matmul_v2             operator              1                     
elementwise_add       operator              1                     
fetch                 operator              1                     
Linear                layer                 1                     
----------------------------------------------------------------

发现在exporter.py中translate_graph中采用model = paddle.jit.to_static(model, input_spec)，如果使用deepcopy(model)，就不会出现警告

import paddle
from copy import deepcopy

from visualdl import LogWriter

net = paddle.nn.Linear(10, 10)
with LogWriter(logdir="./log/graph_test/") as writer:
    writer.add_graph(
        model=deepcopy(net), # 采用deepcopy
        input_spec=[paddle.static.InputSpec([-1, 10], 'float32')],
        verbose=True)

B = deepcopy(net)

-------------------------Graph Summary--------------------------
total operators: 4	total layers:1
--------------------  --------------------  --------------------  
Name                  Type                  Count                 
--------------------  --------------------  --------------------  
feed                  operator              1                     
matmul_v2             operator              1                     
elementwise_add       operator              1                     
fetch                 operator              1                     
Linear                layer                 1                     
----------------------------------------------------------------

建议translate_graph使用deepcopy(model)

def translate_graph(model, input_spec, verbose=True):
    model = deepcopy(model)
    import paddle
    with tempfile.TemporaryDirectory() as tmp:
        model._full_name = '{}[{}]'.format(model.__class__.__name__, "model")
        create_opname_scope(model)
        model = paddle.jit.to_static(model, input_spec)
        paddle.jit.save(model, os.path.join(tmp, 'temp'))
        model_data = open(os.path.join(tmp, 'temp.pdmodel'), 'rb').read()
        result = analyse_model(model_data)
    if verbose:
        print_model(result)
    result = json.dumps(result, indent=2)
    return result

[rainyfly]

请问出现需要deepcopy的场景一般是什么呢