search

holyseven / PSPNet-TF-Reproduce

Training PSPNet in Tensorflow. Reproduce the performance from the paper.
MIT License
125 stars 30 forks source link

Does changing image size and GPU number affect precision? #18

Closed zdluffy closed 5 years ago

zdluffy commented 5 years ago

Hello! I am new in semantic segmention and deep learning, so I I might ask a stupid a question. Due to the limit of my GPU resource, I trained resnet-50 on Cityscapes with the following command. python ./train.py --batch_size 4 --gpu_num 1 --consider_dilated 1 --weight_decay_rate 0.0001 --weight_decay_rate2 0.001 --random_rotate 0 --database 'Cityscapes' --train_image_size 512 --test_image_size 512 only changed the train_image_size and test_image_size from 816 to 512, and the gpu_num is 1 not 4. Then I got 69.93 mIoU scores on the val dataset, which is lower than yours. I wonder if changing the image size and GPU number will affect the performance?

holyseven commented 5 years ago

Yes.

train_image_size is the size of input images during training, which may give more information for segmentation if it is larger. batch_size * gpu_num is the total batch size of each step, which makes the batch normalization more stable if it is larger. So both are quite important.

zdluffy commented 5 years ago

Got it! Thank you very much!