[Question]: 对paddleocr-vl进行LORA微调时报错 Warning: Error processing example, skipping. Error: 'SiglipImageProcessor' object has no attribute 'get_smarted_resize'

[fatjsh]

请提出你的问题

对paddleocr-vl进行LORA微调时候出现 Warning: Error processing example, skipping. Error: 'SiglipImageProcessor' object has no attribute 'get_smarted_resize'的报错，我使用的各个软件版本如下：
paddlecodec 0.1.0
paddleformers 1.0.0.post20260224 /home/PaddleFormers
paddleocr 3.3.2
paddlepaddle-gpu 3.3.0
paddlex 3.3.11
use_triton_in_paddle 0.1.0

[forBlank]

我们后续对一些模型配置进行了调整，请从 huggingface 拉取最新的权重文件后再次尝试；或者请再提供模型下载源和 transformers 版本，以便于我们复现该问题

[fatjsh]

你好，以下是一些具体的环境配置情况。

模型下载源：
export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download PaddlePaddle/PaddleOCR-VL --local-dir PaddlePaddle/PaddleOCR-VL --local-dir-use-symlinks False

transformers版本：
transformers 4.55.4

显卡及配置情况：
显卡：NVIDIA GeForce RTX 5090 D
显卡驱动：NVIDIA-SMI 570.133.07
CUDA Version: 12.8

PaddleFormers的部署：

1.拉取镜像
docker run
-it
--gpus all
--network host
--user root
ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-vl:latest-nvidia-gpu-sm120-offline
/bin/bash

2.升级paddlepaddle3.3.0
python -m pip install paddlepaddle-gpu==3.3.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu129/ （我的CUDA Version是 12.8，是否和这个有关?）

3.安装 PaddleFormers
git clone https://github.com/PaddlePaddle/PaddleFormers.git
cd PaddleFormers
python -m pip install -e .

4.训练
paddleformers-cli train examples/best_practices/PaddleOCR-VL/paddleocr-vl_lora_16k_config.yaml model_name_or_path=./PaddlePaddle/PaddleOCR-VL train_dataset_path=./ocr_vl_sft-train_Bengali.jsonl eval_dataset_path=./ocr_vl_sft-test_Bengali.jsonl pre_alloc_memory=16

部分错误信息如下：

[2026-02-26 06:13:19,144] [ INFO] - loading configuration file ./PaddlePaddle/PaddleOCR-VL/preprocessor_config.json
[2026-02-26 06:13:19,144] [ WARNING] - The requested image processor SiglipImageProcessor does not have a fast version. Detected use_fast=True but the slow version will be loaded instead (use_fast=False).
[2026-02-26 06:13:19,149] [ INFO] - Using download source: huggingface
[2026-02-26 06:13:20,025] [ WARNING] - suffix is not specified, using eos token as suffix.
[2026-02-26 06:13:20,518] [ INFO] - Mark only lora and trainable_module as trainable.
[2026-02-26 06:13:20,532] [ DEBUG] - Frozen parameters: 9.06e+08 || Trainable parameters:7.49e+06 || Total parameters:9.13e+08|| Trainable:0.82%
[2026-02-26 06:13:20,533] [ INFO] - Setting max_seq_len to None using PaddleFormers Model.
[2026-02-26 06:13:20,533] [ INFO] - Start to estimate max training steps...
[2026-02-26 06:13:20,533] [ WARNING] - The results between sampling and non-sampling methods may differ.
Warning: Error processing example, skipping. Error: 'SiglipImageProcessor' object has no attribute 'get_smarted_resize'
Warning: Error processing example, skipping. Error: 'SiglipImageProcessor' object has no attribute 'get_smarted_resize'
Warning: Error processing example, skipping. Error: 'SiglipImageProcessor' object has no attribute 'get_smarted_resize'
Warning: Error processing example, skipping. Error: 'SiglipImageProcessor' object has no attribute 'get_smarted_resize'
Warning: Error processing example, skipping. Error: 'SiglipImageProcessor' object has no attribute 'get_smarted_resize'
Warning: Error processing example, skipping. Error: 'SiglipImageProcessor' object has no attribute 'get_smarted_resize'

相关源代码：
是paddleformers/datasets/template/mm_plugin.py中的_preprocess_image调用了get_smarted_resize方法，_preprocess_image的代码如下：

392 @OverRide
393 def _preprocess_image(self, image, **kwargs):
394
395 width, height = image.size
396 image_max_pixels = kwargs["image_max_pixels"]
397 image_min_pixels = kwargs["image_min_pixels"]
398 image_processor = kwargs["image_processor"]
399 #print("i am in _preprocess_image , i am ok")
400 # pre-resize before augmentation
401 resized_height, resized_width = image_processor.get_smarted_resize(
402 height,
403 width,
404 min_pixels=image_min_pixels,
405 max_pixels=image_max_pixels,
406 )[0]
407
408 image = image.resize((resized_width, resized_height))
409
410 if image and hasattr(self, "image_augmentation"):
411 image = self.image_augmentation(image)
412
413 return image

[forBlank]

请提供一下权重中 preprocessor_config.json 的内容

[forBlank]

transformers版本：
transformers 4.55.4

或者尝试将 transformers 升级到 5.0 版本

[fatjsh]

好的，我升级试一下。

preprocessor_config.json ：
root@user-Super-Server:/home/PaddleFormers/PaddlePaddle/PaddleOCR-VL# cat preprocessor_config.json
{
"auto_map": {
"AutoImageProcessor": "image_processing.SiglipImageProcessor",
"AutoProcessor": "processing_paddleocr_vl.PaddleOCRVLProcessor"
},
"do_convert_rgb": true,
"do_normalize": true,
"do_rescale": true,
"do_resize": true,
"image_mean": [
0.5,
0.5,
0.5
],
"image_processor_type": "SiglipImageProcessor",
"image_std": [
0.5,
0.5,
0.5
],
"max_pixels": 2822400,
"merge_size": 2,
"min_pixels": 147384,
"patch_size": 14,
"processor_class": "PaddleOCRVLProcessor",
"resample": 3,
"rescale_factor": 0.00392156862745098,
"size": {
"max_pixels": 2822400,
"min_pixels": 147384
},
"temporal_patch_size": 1
}

[forBlank]

"image_processor_type": "SiglipImageProcessor",

这里不太对，为了避免错误加载，我们已经更新成 PaddleOCRVLImageProcessor，请从 huggingface 拉取最新的权重文件以保证加载正确的 ImageProcessor

[fatjsh]

好的，谢谢，huggingface这个链接一直打不开，有没有其他的地址可以获取更新。

[forBlank]

模型下载源：
export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download PaddlePaddle/PaddleOCR-VL --local-dir PaddlePaddle/PaddleOCR-VL --local-dir-use-symlinks False

可以这样再下载一次，权重中 preprocessor_config.json 的 image_processor_type 应该是 PaddleOCRVLImageProcessor

[fatjsh]

好的，谢谢～