ocr_recognition预测attention模型报错

prettyocean85 commented 4 years ago

问题描述：使用models\PaddleCV\ocr_recognition训练attention模型（crnn_ctc模型使用正常），并在infer.py中重新保存为固化模型，使用paddle predictor预测报错，错误信息为： File "D:\pythonAPI\Python37\call_paddle_OCR.py", line 108, in test_image predictor.zero_copy_run() paddle.fluid.core_avx.EnforceNotMet:

C++ Call Stacks (More useful to developers):

Windows not support stack backtrace yet.

Python Call Stacks (More useful to users):

File "D:\pythonAPI\Python37\Lib\site-packages\paddle\fluid\framework.py", line 2525, in append_op attrs=kwargs.get("attrs", None)) File "D:\pythonAPI\Python37\Lib\site-packages\paddle\fluid\layer_helper.py", line 43, in append_op return self.main_program.current_block().append_op(*args, **kwargs) File "D:\pythonAPI\Python37\Lib\site-packages\paddle\fluid\layers\sequence_lod.py", line 723, in sequence_expand attrs={'ref_level': ref_level}) File "D:\LSS\OCR\ocr_recognition\attention_model.py", line 294, in attention_infer pre_state_expanded = fluid.layers.sequence_expand(pre_state, pre_score) File "D:\LSS\OCR\ocr_recognition\infer.py", line 63, in inference ids = infer(images, num_classes, use_cudnn=True if args.use_gpu else False) File "D:\LSS\OCR\ocr_recognition\infer.py", line 196, in main inference(args) File "D:\LSS\OCR\ocr_recognition\infer.py", line 200, in main()

Error Message Summary:

Error: Level number of Input(Y)'s lod should be greater than 0. [Hint: Expected y_lod.size() > 0UL, but received y_lod.size():0 <= 0UL:0.] at (D:\1.7.0\paddle\paddle\fluid\operators\sequence_ops\sequence_expand_op.cc:57) [operator < sequence_expand > error]

版本、环境信息： 1）PaddlePaddle版本：1.7.0 2）系统环境：win10，Python3.7 -预测代码 data = preprocess_image(image) data=[data] inputs = np.concatenate( list(map(lambda x: x[0][np.newaxis, :], data)), axis=0).astype("float32") input1=np.array([[0]], dtype='int64') input2=np.array([[1.]], dtype='float32') input_names = predictor.get_input_names() input_tensor0 = predictor.get_input_tensor(input_names[0]) input_tensor0.reshape(inputs.shape) input_tensor0.copy_from_cpu(inputs) if INFER_TYPE=="attention" :##'crnn_ctc' or 'attention' input_tensor1 = predictor.get_input_tensor(input_names[1]) input_tensor1.reshape(input1.shape) input_tensor1.copy_from_cpu(input1) input_tensor2 = predictor.get_input_tensor(input_names[2]) input_tensor2.reshape(input2.shape) input_tensor2.copy_from_cpu(input2) predictor.zero_copy_run()
模型保存代码： if args.save_inference:#保存 print("save freeze_model in :" + args.freeze_model_path) if args.model == "crnn_ctc": feed_var_names = ['pixel'] else: feed_var_names = ['pixel', 'init_ids', 'init_scores'] print(feed_var_names) test_program = fluid.default_main_program().clone(for_test=True) target_vars = ids fluid.io.save_inference_model(dirname=args.freeze_model_path, feeded_var_names=feed_var_names, target_vars=target_vars, main_program=test_program, executor=exe, model_filename='model', params_filename='params')

yongqiangma commented 4 years ago

模型不保存，用同样的输入进行预测，会报错吗？

prettyocean85 commented 4 years ago

@yongqiangma 直接用infer.py预测都不报错

jiweibo commented 4 years ago

有可能是save_inference_model的时候发生了错误裁剪，等https://github.com/PaddlePaddle/Paddle/pull/23306 合入后再测试下吧

jiweibo commented 4 years ago

@prettyocean85 在ocr_recognition上使用inference_model测试了一下，发现没有报错。

再结合你的报错信息发现是输入的问题。 Error: Level number of Input(Y)'s lod should be greater than 0. [Hint: Expected y_lod.size() > 0UL, but received y_lod.size():0 <= 0UL:0.] at

你可以参考下python是如何处理输入数据的。

最简单的方法测试，修改infer.py使得其加载inference_model完成预测。

prettyocean85 commented 4 years ago

@jiweibo 修改infer.py可以完成预测，但是报错是使用零拷贝方式，数据处理是参考了python来做的，和训练用的数据没什么区别不知道为什么报错。预测代码如下（crnn_ctc模型使用正常）：

jiweibo commented 4 years ago

@prettyocean85 看报错信息是输入tensor没有设置lod。

lod的概念请参考：https://www.paddlepaddle.org.cn/documentation/docs/zh/beginners_guide/basic_concept/lod_tensor.html#lod

另外请查看ocr_recognition中，对输入数据的处理，lod是如何设置的。

prettyocean85 commented 4 years ago

@jiweibo 添加set_lod后问题已解决，感谢

PaddlePaddle / Paddle