表格识别 训练时候，进行eval 时候遇到bug

[liuzhipengchd]

PaddleOCR/ppocr/modeling/heads/table_att_head.py

Line 370 in 7a3c580

structure_ids = paddle.zeros(

[Hint: Expected data_type == phi::DataType::FLOAT16 || data_type == phi::DataType::BFLOAT16 == true, but received data_type == phi::DataType::FLOAT16 || data_type == phi::DataType::BFLOAT16:0 != true:1.] (at ../paddle/fluid/imperative/amp_auto_cast.cc:190)

在评估时候，structure_ids 的数据类型和pre_chars 不一致，导致bug

[GreatV]

运行环境是什么 paddle版本是多少 paddleocr 版本是多少

[liuzhipengchd]

运行环境是什么 paddle版本是多少 paddleocr 版本是多少

paddleocr 是2.8 ，paddlepaddle-gpu == 0.0.0.post112

[GreatV]

用的是啥gpu呀，amp模式训练的吗，不用amp可以运行吗

[liuzhipengchd]

用的是啥gpu呀，amp模式训练的吗，不用amp可以运行吗

4090，就是正常模式。把那个数据格式改成统一的int32就可以执行。。不是这个bug吗？

[GreatV]

好的 @liuzhipengchd ，谢谢反馈，你能提一个PR来修复它吗？

[GreatV]

4090，就是正常模式。把那个数据格式改成统一的int32就可以执行。。不是这个bug吗？

因为这里看上去是启用amp，导致的bug。

[liuzhipengchd]

4090，就是正常模式。把那个数据格式改成统一的int32就可以执行。。不是这个bug吗？

因为这里看上去是启用amp，导致的bug。

好的。。大佬，我想问下。1、训练表格识别模型，那个效果好 SLANet_lcnetv2.yml 和 SLANet_ch.yml （如果我选择scale 2.5的呢），2、SLANet_lcnetv2 中 我想选择 PPLCNetV2_large的话，应该怎么配置

[GreatV]

我想问下。1、训练表格识别模型，那个效果好 SLANet_lcnetv2.yml 和 SLANet_ch.yml （如果我选择scale 2.5的呢），2、SLANet_lcnetv2 中 我想选择 PPLCNetV2_large的话，应该怎么配置

应该是 SLANet_lcnetv2.yml 更好一点，具体的得请教一下 @invictuszhao

[GreatV]

@liuzhipengchd 无法复现你的问题

 python3 tools/train.py -c configs/table/SLANet.yml -o Train.loader.batch_size_per_card=16 Eval.loader.batch_size_per_card=16

[2024/07/06 07:43:04] ppocr INFO: epoch: [1/100], global_step: 1000, lr: 0.001000, acc: 0.000000, loss: 0.083020, structure_loss: 0.055820, loc_loss: 0.026588, avg_reader_cost: 0.00075 s, avg_batch_cost: 0.28300 s, avg_samples: 16.0, ips: 56.53801 samples/s, eta: 10 days, 9:10:27, max_mem_reserved: 8722 MB, max_mem_allocated: 8043 MB
eval model:: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 570/570 [01:45<00:00,  5.41it/s]
[2024/07/06 07:44:49] ppocr INFO: cur metric, acc: 0.026001097089851773, fps: 94.59463124211499
[2024/07/06 07:44:49] ppocr INFO: save best model is to ./output/SLANet/best_accuracy
[2024/07/06 07:44:49] ppocr INFO: best metric, acc: 0.026001097089851773, is_float16: False, fps: 94.59463124211499, best_epoch: 1
[2024/07/06 07:44:55] ppocr INFO: epoch: [1/100], global_step: 1020, lr: 0.001000, acc: 0.000000, loss: 0.078646, structure_loss: 0.050355, loc_loss: 0.032018, avg_reader_cost: 0.00104 s, avg_batch_cost: 0.27847 s, avg_samples: 16.0, ips: 57.45583 samples/s, eta: 10 days, 9:01:56, max_mem_reserved: 8722 MB, max_mem_allocated: 8043 MB

python3 tools/train.py -c configs/table/SLANet.yml -o Train.loader.batch_size_per_card=16 Eval.loader.batch_size_per_card=16 Global.use_amp=True

[2024/07/06 07:59:55] ppocr INFO: epoch: [1/100], global_step: 1000, lr: 0.001000, acc: 0.000000, loss: 0.084489, structure_loss: 0.057010, loc_loss: 0.028551, avg_reader_cost: 0.00082 s, avg_batch_cost: 0.26592 s, avg_samples: 16.0, ips: 60.16862 samples/s, eta: 9 days, 18:05:22, max_mem_reserved: 4516 MB, max_mem_allocated: 4079 MB
eval model:: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 570/570 [01:59<00:00,  4.78it/s]
[2024/07/06 08:01:55] ppocr INFO: cur metric, acc: 0.005485463521065774, fps: 82.6024714859491
[2024/07/06 08:01:55] ppocr INFO: save best model is to ./output/SLANet/best_accuracy
[2024/07/06 08:01:55] ppocr INFO: best metric, acc: 0.005485463521065774, is_float16: False, fps: 82.6024714859491, best_epoch: 1
[2024/07/06 08:02:00] ppocr INFO: epoch: [1/100], global_step: 1020, lr: 0.001000, acc: 0.000000, loss: 0.078040, structure_loss: 0.051105, loc_loss: 0.030460, avg_reader_cost: 0.00108 s, avg_batch_cost: 0.26121 s, avg_samples: 16.0, ips: 61.25244 samples/s, eta: 9 days, 17:56:58, max_mem_reserved: 4691 MB, max_mem_allocated: 4079 MB

[liuzhipengchd]

@liuzhipengchd 无法复现你的问题

 python3 tools/train.py -c configs/table/SLANet.yml -o Train.loader.batch_size_per_card=16 Eval.loader.batch_size_per_card=16

[2024/07/06 07:43:04] ppocr INFO: epoch: [1/100], global_step: 1000, lr: 0.001000, acc: 0.000000, loss: 0.083020, structure_loss: 0.055820, loc_loss: 0.026588, avg_reader_cost: 0.00075 s, avg_batch_cost: 0.28300 s, avg_samples: 16.0, ips: 56.53801 samples/s, eta: 10 days, 9:10:27, max_mem_reserved: 8722 MB, max_mem_allocated: 8043 MB
eval model:: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 570/570 [01:45<00:00,  5.41it/s]
[2024/07/06 07:44:49] ppocr INFO: cur metric, acc: 0.026001097089851773, fps: 94.59463124211499
[2024/07/06 07:44:49] ppocr INFO: save best model is to ./output/SLANet/best_accuracy
[2024/07/06 07:44:49] ppocr INFO: best metric, acc: 0.026001097089851773, is_float16: False, fps: 94.59463124211499, best_epoch: 1
[2024/07/06 07:44:55] ppocr INFO: epoch: [1/100], global_step: 1020, lr: 0.001000, acc: 0.000000, loss: 0.078646, structure_loss: 0.050355, loc_loss: 0.032018, avg_reader_cost: 0.00104 s, avg_batch_cost: 0.27847 s, avg_samples: 16.0, ips: 57.45583 samples/s, eta: 10 days, 9:01:56, max_mem_reserved: 8722 MB, max_mem_allocated: 8043 MB

python3 tools/train.py -c configs/table/SLANet.yml -o Train.loader.batch_size_per_card=16 Eval.loader.batch_size_per_card=16 Global.use_amp=True

[2024/07/06 07:59:55] ppocr INFO: epoch: [1/100], global_step: 1000, lr: 0.001000, acc: 0.000000, loss: 0.084489, structure_loss: 0.057010, loc_loss: 0.028551, avg_reader_cost: 0.00082 s, avg_batch_cost: 0.26592 s, avg_samples: 16.0, ips: 60.16862 samples/s, eta: 9 days, 18:05:22, max_mem_reserved: 4516 MB, max_mem_allocated: 4079 MB
eval model:: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 570/570 [01:59<00:00,  4.78it/s]
[2024/07/06 08:01:55] ppocr INFO: cur metric, acc: 0.005485463521065774, fps: 82.6024714859491
[2024/07/06 08:01:55] ppocr INFO: save best model is to ./output/SLANet/best_accuracy
[2024/07/06 08:01:55] ppocr INFO: best metric, acc: 0.005485463521065774, is_float16: False, fps: 82.6024714859491, best_epoch: 1
[2024/07/06 08:02:00] ppocr INFO: epoch: [1/100], global_step: 1020, lr: 0.001000, acc: 0.000000, loss: 0.078040, structure_loss: 0.051105, loc_loss: 0.030460, avg_reader_cost: 0.00108 s, avg_batch_cost: 0.26121 s, avg_samples: 16.0, ips: 61.25244 samples/s, eta: 9 days, 17:56:58, max_mem_reserved: 4691 MB, max_mem_allocated: 4079 MB

我这里可以复现问题
保持类型一样就可以了。（我把类型保持一样，会影响训练效果吗）

[GreatV]

@liuzhipengchd 不会影响效果，这个已经修复了。

• unifying data types in the SLAHead #13276