Fine Tune? about db HOT 13 CLOSED

@zobeirraisi Using --resume model_path in the training command.

from db.

Is the pretrained model "pre-trained-model-synthtext-resnet18" trained on synthtext dataset with dcn or normal conv2d?

from db.

@moooood It seems that I uploaded the wrong "pre-trained-model-synthtext-resnet18" which does not include the DCN.

from db.

@MhLiao 😎
in regards of F1, recall, precision, whats the difference between the dcn and normal conv2d version?
resnet 18 vs deformable resnet 18

from db.

@deepseek DCN usually achieves higher performance, especially for the long text lines, since DCN gives a more flexible receptive field for the backbone.

from db.

@MhLiao Could you provide the "pre-trained-model-synthtext-resnet18" model which trained with DCN?

from db.

@liangkaihuan I will provide the model but maybe I need to find it or re-train it.

from db.

The provided model ”pre-trained-model-synthtext-resnet18“ is trained with normal conv2d. Is that right？

from db.

The provided model ”pre-trained-model-synthtext-resnet18“ is trained with normal conv2d. Is that right？

Yes. I will fix it later.

from db.

Is the pretrained model "pre-trained-model-synthtext" trained on synthtext dataset with dcn or normal conv2d now ? @MhLiao

from db.

@shaohailin It was still pretrained with normal conv2d, you can print model.state_dict().keys().

from db.

can anyone provide pretrained with DCN?

from db.

您好，现在请问有DCN 预训练的pre-trained-model-synthtext-resnet18吗？

from db.