The training effect is good, but the test effect is poor.

The image generated in the training process has a good effect, but it is also poor to test the effect with the training set.My orders are： train:

python train.py --name InstanceData --instance_feat --checkpoints_dir ./CheckPoint --batchSize 2 --label_nc 7

python precompute_feature_maps.py --name InstanceData --label_nc 7 --batchSize 4 --checkpoints_dir ./CheckPoint --instance_feat

python train.py --name InstanceDataNew --netG local --ngf 32 --num_D 3 --load_pretrain CheckPoint/InstanceData/ --niter 20 --niter_decay 20 --niter_fix_global 10 --instance_feat --load_features --batchSize 4 --label_nc 7 --checkpoints_dir ./CheckPoint

test:

python encode_features.py --name InstanceDataNew --netG local --ngf 32 --checkpoints_dir ./CheckPoint --label_nc 7 --instance_feat

python test.py --name InstanceDataNew --netG local --ngf 32 --instance_feat --checkpoints_dir ./CheckPoint --label_nc 7