ch_PP-OCRv4_rec_svtr_large.yml训练导出的模型，使用predict_rec.py预测宽度比较大的图片时出现(InvalidArgument) Broadcast dimension mismatch

[LMR2018]

使用ch_PP-OCRv4_rec_svtr_large.yml训练的OCR识别模型，训练正常，
使用python tools/eval.py -c configs/rec/PP-OCRv4/ch_PP-OCRv4_rec_svtr_large.yml也是正常的，
用python tools/export_model.py -c configs/rec/PP-OCRv4/ch_PP-OCRv4_rec_svtr_large.yml 也是能成功导出模型的
用export_model.py导出的模型，使用python tools/infer/predict_rec.py预测宽度不太长的单行图片也是能正常预测的，
但是predict_rec.py预测宽度比较大的图片时出现：(InvalidArgument) Broadcast dimension mismatch错误
即使是训练、验证用的宽度比较大的图片也是出现这个错误，这个怎么解决？
训练配置：image_shape: [3, 48, 320]， max_text_length: &max_text_length 50

`[2024/05/25 14:48:36] ppocr INFO: Traceback (most recent call last):
File "tools/infer/predict_rec.py", line 728, in main
rec_res, _ = text_recognizer(img_list)
File "tools/infer/predict_rec.py", line 675, in call
self.predictor.run()
ValueError: (InvalidArgument) Broadcast dimension mismatch. Operands could not be broadcast together with the shape of X = [1, 2148, 192] and the shape of Y = [1, 960, 192]. Received [2148] in X is not equal to [960] 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/phi/kernels/funcs/common_shape.h:86)
[operator < elementwise_add > error]

[2024/05/25 14:48:36] ppocr INFO: (InvalidArgument) Broadcast dimension mismatch. Operands could not be broadcast together with the shape of X = [1, 2148, 192] and the shape of Y = [1, 960, 192]. Received [2148] in X is not equal to [960] 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/phi/kernels/funcs/common_shape.h:86)
[operator < elementwise_add > error]`

[LMR2018]

[GreatV]

直接resize 成 [3, 48, 320] 可以infer吗

[LMR2018]

直接resize 成 [3, 48, 320] 可以infer吗

一开始就进行直接resize成（320， 48），不行，后面官方又做了别的处理了，最后图片不是（320，48，3）

[LMR2018]

还发现一个问题，上面28032、904118的原始图片出错，但是用过58587，579102，583271，359190的却能正常，所以在猜想不是宽度太长的问题，用ch_PP-OCRv4_rec.yml配置训练的模型就没有这个问题

[nissansz]

ch_PP-OCRv4_rec.yml

ch_PP-OCRv4_rec.yml 这个配置我怎么训练不了，batch size = 1都爆显存。
ch_PP-OCRv4_rec.yml 这个配置没问题，可以训练