add center loss cod and cfg #4165
Conversation
|
Thanks for your contribution! |
| epoch_num: 800 | ||
| log_smooth_window: 20 | ||
| print_batch_step: 10 | ||
| save_model_dir: ./output/rec_mobile_pp-OCRv2_center_loss |
There was a problem hiding this comment.
建议将center改成enhanced_ctc， center并不准确。 arxiv论文上也是这么对外宣传的
There was a problem hiding this comment.
好的， 已修改
| infer_mode: false | ||
| use_space_char: true | ||
| distributed: true | ||
| save_res_path: ./output/rec/predicts_mobile_pp-OCRv2_center_loss.txt |
| algorithm: CRNN | ||
| Transform: | ||
| Backbone: | ||
| name: MobileNetV1Enhance |
There was a problem hiding this comment.
是不是已经确定叫LCNet了，如果确定了，可以直接写LCNet.
There was a problem hiding this comment.
这块由于pretrain已经开源，目前修改风险可能比较高，暂时感觉修改不太合适
| num_classes: 6625 | ||
| feat_dim: 96 | ||
| init_center: false | ||
| center_file_path: |
|
|
||
| feats_reshape = paddle.reshape( | ||
| features, [-1, features.shape[-1]]).astype("float64") | ||
| label = paddle.argmax(predicts, axis=2) |
There was a problem hiding this comment.
这里的标签，是不是应该使用batch也就是gt进行计算，predicts是预测的标签？
There was a problem hiding this comment.
OCR识别里面，是没有显式的对齐标签的；所以训练的时候gt没法拿来使用
There was a problem hiding this comment.
嗯，用gt是没有对齐的有明显的问题，但是会不会存在用predicts因为不是正确的标签，计算出来的centerloss引入比较大的偏差，从而导致学习出的“center”质量不是很高？
There was a problem hiding this comment.
如果center不做初始化的话，确实会影响效果；PPOCR-V2的方案，会先基于单纯的ctcloss，训练一个模型；然后基于训练数据提取center来做初始化。 另外，centerloss的权重一般也不会设置得很大
|
No description provided.