paddle 动转静示例代码

dynamic_to_static.py

# 1. print the paddle version
from paddle import __version__
print("Paddle Version: " + __version__)

from paddlenlp.transformers import AutoModel
import paddle
# 2. load the dynamic model
input_ids = paddle.randint(10, 20, shape=[1, 20], dtype='int64')

model = AutoModel.from_pretrained("albert-base-v1")
dynamic_output = model(input_ids)[0]

# 3. save & load static model
inputs = [paddle.static.InputSpec(shape=[None, None], dtype="int64")]
model = paddle.jit.to_static(model, input_spec=inputs)
path = "sss/static_model"
paddle.jit.save(model, path)

model = paddle.jit.load(path)
static_output = model(input_ids)[0]

# 4. compare the dynamic output and static output
print(dynamic_output[:, 1:4, 1:4]) 

print("=======")
print(static_output[:, 1:4, 1:4]) 

assert paddle.allclose(dynamic_output[:, 1:4, 1:4], static_output[:, 1:4, 1:4], atol=1e-4)

Oh, good catch, using eval can output the same logit. @gongel thx.