open-mmlab / mmocr

OpenMMLab Text Detection, Recognition and Understanding Toolbox
https://mmocr.readthedocs.io/en/dev-1.x/
Apache License 2.0
4.27k stars 743 forks source link

请问fcenet是如何达到20fps的 #916

Closed 6098669 closed 2 years ago

6098669 commented 2 years ago

为啥我测试的只有5.8fps。 如果说减小输入图片尺寸,那不是性能也会下降么。 为啥我测的5.8fps和您之前报道的20fps相差这么大呀 -5baf53f8fff9dc9f

mm-assistant[bot] commented 2 years ago

Please use English or English & Chinese for issues so that we could have broader discussion.

Mountchicken commented 2 years ago

5.8 task/s is not a good reflection of FPS. This is not an average time estimate, but an instantaneous estimate, meaning that the current image inference uses 1/5.8 s. It may just be that the current image inference is slow, and you can observe what this speed is roughly when inferring these 500 images. Or you can count the total inference time and divide it by the number of images to calculate the average FPS

xinke-wang commented 2 years ago

Usually, the inference speed FPS only counts the network forward time, but this task/s also includes other data processing time.

ming-eng commented 2 years ago

请问这个fcenet需要训练多久

xinke-wang commented 2 years ago

请问这个fcenet需要训练多久

You may check the FCENet training log here to estimate the training time.

ming-eng commented 2 years ago

当训练fcenet使用默认的ic15配置文件 单卡训练的过程中 损失函数突然为nan 请问您遇到了吗

gaotongxiao commented 2 years ago

It happens usually because the learning rate is too large for your batch size. Since the discussion is becoming a bit off-topic, I'm closing this issue now. Please create a new issue for any follow-up discussion.