zyxu1996

Your application is correct, missing keys displayed will not affect the experiment results.

How did you run the test code command? The result should not be related to the batch size. Maybe you should run the test code in the corresponding work_dir.

It seems the network params are random initialized.

The val batchsizes are used to accelerate the inferring process.

I noticed that your trained weights are saved in tmp_save dir, but the information in the test command is num1. Both the names should be same, otherwise the model can...

Have you trained the model? If you don't train the model, you will not get the trained weights in the work_dir, so the loaded trained weight are randomly initialized.

Different loaded weight will output different results.

Try this

CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --nproc_per_node=1 --master_port 29506 train.py