ch_PP-OCRv3_rec_distillation.yml 这个配置文件，怎么改成用resnet 34? 原配置是mobilenet的，可以训练， 但是不知道怎么改成resnet 34? 自己改的报错

[nissansz]

请提供下述完整信息以便快速定位问题/Please provide the following information to quickly locate the problem

• 系统环境/System Environment：win10
• 版本号/Version：Paddle： PaddleOCR：2.5 问题相关组件/Related components：
• 运行指令/Command Code：
• 完整报错/Complete Error Message：

ch_PP-OCRv3_rec_distillation.yml 这个配置文件，怎么改成用resnet 34? 原配置是mobilenet的，可以训练， 但是不知道怎么改成resnet 34? 自己改的报错

是不是用了distillation yml，就不用另外设置combined loss? combined loss需要的center loss, train_center.pkl, 文件导出是1kb，是导出失败了吧？

[andyjiang1116]

报错信息是什么呢？

[nissansz]

好像要改loss之类的，不是只改backbone

[nissansz]

Traceback (most recent call last): File "../input/paddle251/PaddleOCR-release-2.5/tools/train.py", line 191, in main(config, device, logger, vdl_writer) File "../input/paddle251/PaddleOCR-release-2.5/tools/train.py", line 166, in main eval_class, pre_best_model_dict, logger, vdl_writer, scaler) File "/kaggle/input/paddle251/PaddleOCR-release-2.5/tools/program.py", line 268, in train loss = loss_class(preds, batch) File "/opt/conda/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py", line 930, in call return self._dygraph_call_func(*inputs, kwargs) File "/opt/conda/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py", line 915, in _dygraph_call_func outputs = self.forward(inputs, kwargs) File "/kaggle/input/paddle251/PaddleOCR-release-2.5/ppocr/losses/combined_loss.py", line 55, in forward loss = loss_func(input, batch, kargs) File "/opt/conda/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py", line 930, in call return self._dygraph_call_func(inputs, kwargs) File "/opt/conda/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py", line 915, in _dygraph_call_func outputs = self.forward(*inputs, *kwargs) File "/kaggle/input/paddle251/PaddleOCR-release-2.5/ppocr/losses/distillation_loss.py", line 107, in forward out2[self.dis_head]) File "/kaggle/input/paddle251/PaddleOCR-release-2.5/ppocr/losses/basic_loss.py", line 116, in forward self._kldiv(log_out1, out2) + self._kldiv(log_out2, out1)) / 2.0 File "/kaggle/input/paddle251/PaddleOCR-release-2.5/ppocr/losses/basic_loss.py", line 102, in _kldiv loss = target (paddle.log(target + eps) - x) File "/opt/conda/lib/python3.7/site-packages/paddle/fluid/dygraph/math_op_patch.py", line 299, in impl return math_op(self, other_var, 'axis', axis) ValueError: (InvalidArgument) Broadcast dimension mismatch. Operands could not be broadcast together with the shape of X = [16, 80, 63466] and the shape of Y = [16, 40, 63466]. Received [80] in X is not equal to [40] in Y at i:1. [Hint: Expected x_dims_array[i] == y_dims_array[i] || x_dims_array[i] <= 1 || y_dims_array[i] <= 1 == true, but received x_dims_array[i] == y_dims_array[i] || x_dims_array[i] <= 1 || y_dims_array[i] <= 1:0 != true:1.] (at /paddle/paddle/phi/kernels/funcs/common_shape.h:84) [operator < elementwise_sub > error]

[andyjiang1116]

这个是因为backbone输出的特征维度和原始的不一致，原始应该是40，resnet34是80，可以在backbone最后加一层pool层保持一致

[nissansz]

不会改。可以帮忙改一下，发给能跑通的yml文件试试吗？

[andyjiang1116]

这个要改源码，这边替换backbone的初衷是啥呢

[nissansz]

因为rec_r34_vd_none_bilstm_ctc.yml 这个配置文件训练速度比mobilenet快很多。 而Distillation好像又能提高准确率，所以想试试rec_r34_Distillation

另外VD还能改善R34准确率？提高1-2%？

[andyjiang1116]

V3模型提供的是轻量化模型，直接更改结构不保证能提升效果，如果需要大模型，可以试下svtr模型https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.6/doc/doc_ch/algorithm_rec_svtr.md

[nissansz]

这种用哪个yml配置文件可以训练？也可以像distillation那样互相监督，以改善准确率？

[nissansz]

如果在原rec_r34_vd_none_bilstm_ctc.yml 基础上用combinedloss, 也能达到相互监督改善准确率的效果吗？ 自己生成了train_center.pkl，加入centerloss,好像没法用来训练

[andyjiang1116]

distillation更多是用来提升小模型精度的，你要是只追求高精度，就试试大模型svtr吧

[nissansz]

svtr生成的onnx模型也能和resnet34一样用？ 训练速度慢很多？

[andyjiang1116]

可以试下哈，大模型会慢一些，但是精度高，参考上面的文档

[nissansz]

好的。谢谢。我试试。 后面会有更新的源码，支持resnet34 distillation吗？

[nissansz]

这个rec_svtr_large_10local_11global_stn_ch.yml 不需要做distillation? combinedloss?

[andyjiang1116]

推荐使用默认官方配置就好，自定义配置需要自己来修改适配哈

[nissansz]

官方配置rec_r34_vd_none_bilstm_ctc.yml ，不改源码， 有什么办法实现监督训练吗？

[nissansz]

用rec_svtr_large_10local_11global_stn_ch.yml 这个没法训练，我之前使用Simple_Dataset.py 改了配置文件，就报错了

[2023/05/11 09:22:25] ppocr INFO: train with paddle 2.3.2 and device Place(gpu:0) Traceback (most recent call last): File "../input/paddle251/PaddleOCR-release-2.5/tools/train.py", line 191, in main(config, device, logger, vdl_writer) File "../input/paddle251/PaddleOCR-release-2.5/tools/train.py", line 52, in main train_dataloader = build_dataloader(config, 'Train', device, logger) File "/kaggle/input/paddle251/PaddleOCR-release-2.5/ppocr/data/init.py", line 61, in build_dataloader 'DataSet only support {}'.format(support_dict)) AssertionError: DataSet only support ['SimpleDataSet', 'LMDBDataSet', 'PGDataSet', 'PubTabDataSet']

[nissansz]

可以帮忙修改源码，发一个resnet34 distillation的配置文件 和 支持该配置文件的py文件吗？

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed in 7 days if no further activity occurs. Thank you for your contributions.