The FCE-Net-ctw1500 log file is confusing.

[williamking5]

Hi, I'm checking the log file of FCE-Net training process, but I am confused of the config file of FCE-net-ctw1500.

According to your official log json in your documentation https://download.openmmlab.com/mmocr/textdet/fcenet/20210511_181328.log.json, samples_per_gpu=6, GPU 0, 1 and there are about 160 iterations each epoch. Hence the dataset size should be approximately 6 * 2 * 160 = 1920. But as we know, the size of training part of ctw-1500 dataset is 1000.

The number of samples is 6 per GPU:

There are 2 GPU for training:

There are around 160 iterations each epoch:

I use the same config file to train on my own, finding the size is correct:

There are 2 GPU for training:

There are around 80 iterations each epoch:

The number of samples is 6 per GPU

Do you have any idea?

[gaotongxiao]

We only used 1 GPU for training. GPU 0,1: GeForce RTX 2080 Ti only lists the GPUs available on the machine, but for the actual GPU(s) that we have used please refer to gpu_ids. In this example, gpu_ids = [1], which means that we have used the GPU indexed as 1.

[williamking5]

@gaotongxiao Thank you for explanation!