rec_pphgnetv2.py 对大于320的文本长度，比如640时候 F.adaptive_avg_pool2d(x, [1, 40])强制输出（1，40）大小不会损失特征？

[cqray1990]

🔎 Search before asking

• I have searched the PaddleOCR Docs and found no similar bug report.
• I have searched the PaddleOCR Issues and found no similar bug report.
• I have searched the PaddleOCR Discussions and found no similar bug report.

🐛 Bug (问题描述)

因为320对文长文本太短了，想增加长度，是否可以直接增加到1024之类的

        def forward(self, x):
            x = self.stem(x)
            out = []
            for i, stage in enumerate(self.stages):
                x = stage(x)
                if self.det and i in self.out_indices:
                    out.append(x)
            if self.det:
                return out
    
            if self.text_rec:
                if self.training:
                    x = F.adaptive_avg_pool2d(x, [1, 40])
                else:
                    x = F.avg_pool2d(x, [3, 2])
            return x

🏃‍♂️ Environment (运行环境)

ubantu 20

🌰 Minimal Reproducible Example (最小可复现问题的Demo)

paddleocrv5

@changdazhou @jzhang533

[FocusMode-coder]

Hey! I may be able to help with this bug.

I build custom fixes for Python projects and automation systems.

If you'd like help resolving this issue quickly feel free to reach out.

Happy to take a look 👍

[ICEVIL]

确实会有特征损失的，在我的项目上就存在这个问题，最开始以320的长度进行训练，发现模型对于长文本识别效果不好（一句话中总是存在部分字符漏识别），后来以640长度进行训练，又发现精度对不齐。最终也是发现了这个固定的（1， 40），问题，然后将这个代码改了一下就OK了：
width_dim = int(x.shape[-1] / 2)
x = F.adaptive_avg_pool2d(x, [1, width_dim ])