paddle 1.8 ernie生成文本分类模型后，使用paddle inference tensorRT库报错

[lonelydancer]

docker环境
docker pull hub.baidubce.com/paddlepaddle/paddle:1.8.0-gpu-cuda10.0-cudnn7-trt6
python2.7 / ernie v2.4.0
操作步骤：
运行 finetune_classifier_dygraph.py,得到model后，使用paddle_inference_demo中的例子做推理加速。

错误信息：

Error Message Summary:

Error: Operator skip_layernorm has not been registered
[Hint: op_info_ptr should not be null.] at (/paddle /trt_refine_int8/paddle/fluid/framework/op_info.h:140)

相关代码&模型:
链接: https://pan.baidu.com/s/1xVIyc2ntboVq_f7hse4uyg 提取码: gudn 复制这段内容后打开百度网盘手机App，操作更方便哦

[lonelydancer]

用2.0.5版本跑,python3.6
https://github.com/lonelydancer/algorithm/blob/master/ernie_test.py
报错
I0422 09:49:28.191212 11579 tensorrt_subgraph_pass.cc:347] Prepare TRT engine (Optimize model structure, Select OP kernel etc). This process may cost a lot of time.
Traceback (most recent call last):
File "test_classifier.py", line 77, in
pred = create_predictor(args)
File "test_classifier.py", line 36, in create_predictor
predictor = create_paddle_predictor(config)
ValueError: (InvalidArgument) The Bias's size(768) should be equal to the first dim(-1) of the Input.
[Hint: Expected Y_t->dims()[0] == dims_x.d[1], but received Y_t->dims()[0]:768 != dims_x.d[1]:-1.] (at /paddle/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc:115)

[OleNet]

非 TRT 的情况下， 能够正常运行吗？

[lonelydancer]

@OleNet 可以

[shangzhizhou]

用2.0.5版本跑,python3.6
https://github.com/lonelydancer/algorithm/blob/master/ernie_test.py
报错
I0422 09:49:28.191212 11579 tensorrt_subgraph_pass.cc:347] Prepare TRT engine (Optimize model structure, Select OP kernel etc). This process may cost a lot of time.
Traceback (most recent call last):
File "test_classifier.py", line 77, in
pred = create_predictor(args)
File "test_classifier.py", line 36, in create_predictor
predictor = create_paddle_predictor(config)
ValueError: (InvalidArgument) The Bias's size(768) should be equal to the first dim(-1) of the Input.
[Hint: Expected Y_t->dims()[0] == dims_x.d[1], but received Y_t->dims()[0]:768 != dims_x.d[1]:-1.] (at /paddle/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc:115)

请问2.0.5是什么版本？paddle好像没有发布这个版本，

[lonelydancer]

@shangzhizhou我typo了，应该是 paddlepaddle_gpu-2.0.2

[shangzhizhou]

好的，我们尽快复现一下问题

[fengxiaoshuai]

您好，这边根据提供的模型和脚本进行了测试，目前发现脚本存在一些问题，主要是设置trt子图的动态输入理解有误，下面是修改后可以运行的代码，参考代码链接

[lonelydancer]

您好，这边根据提供的模型和脚本进行了测试，目前发现脚本存在一些问题，主要是设置trt子图的动态输入理解有误，下面是修改后可以运行的代码，参考代码链接

十分感谢。
我看了修改的几处diff，
输入的数据是
in1 = np.ones((1, 128)).astype(np.int64)
而min_input_shape = [1, 1, 768]
为什么输入的shape发生了变化呢？哪里有更详尽的文档。

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Feel free to reopen it. Thank you for your contributions.