Slightly lower validation accuracy in Pytorch 1.0.0

[yzou2]

Hi @speedinghzl ,

I used your pytorch deeplab v2 implementation, same setting, with pytorch version of 1.0.0. The validation mIoU of VOC_scenes_20000.pth is 71.1. Is this degradation due to randomness? Could you give me any comments?
Here is the parameter I used:

BATCH_SIZE = 10
DATA_DIRECTORY = './dataset/voc12'
DATA_LIST_PATH = './dataset/list/train_aug.txt'
IGNORE_LABEL = 255
INPUT_SIZE = '321,321'
LEARNING_RATE = 2.5e-4
MOMENTUM = 0.9
NUM_CLASSES = 21
NUM_STEPS = 20000
POWER = 0.9
RANDOM_SEED = 1234
RESTORE_FROM = './dataset/MS_DeepLab_resnet_pretrained_COCO_init.pth'
SAVE_NUM_IMAGES = 2
SAVE_PRED_EVERY = 1000
SNAPSHOT_DIR = './snapshots/'
WEIGHT_DECAY = 0.0005

Thank you for your help!

[speedinghzl]

Maybe you should set align_corners=True for Upsample for Pytroch > 0.4.

[yzou2]

Maybe you should set align_corners=True for Upsample for Pytroch > 0.4.

Thanks for your prompt reply. Actually the evaluate.py I used, which is your up-to-dated implementation, already set align_corners=True. So I don't think this is the reason for the performance degradation. Do you have any other ideas? Really appreciate your help.

[speedinghzl]

Do you evaluate the VOC_scenes_20000.pth I provided?

[yzou2]

Do you evaluate the VOC_scenes_20000.pth I provided?

I just evaluated the provided VOC_scenes_20000.pth. It can only get 70.4 which is slightly worse than the model fine-tuned in my computer. Any idea why this happen?

[WilliamLwj]

I retrained the whole model and get 73.58 IoU at 20k iter, slightly lower but understandable.

[speedinghzl]

@WilliamLwj Thanks for your feedback and your produced results. I'm not sure whether the performance gap is caused by the different Pytorch versions (0.2 vs 1.0).