inference result with model trained with train.py is not good and model size is almost twise bigger then pretrained model

[kiitosu]

[kiitosu]

I tried to train PSPNet101 with cityscapes dataset. I did below. ・download dataset ・edit cityscapes_trail_list.txt gtFine_labelTrainIds.png to gtFine_labeIds.png ・edit train.py DATA_DIRECTORY to suit my image data ・change batch size 2 to 1 ・python train.py

After 15hours I finished the training . Inference with inference.py is so poor.

Traind data I created is File size of model.ckpt-60000.data-00000-of-00001 is about 502M byte. File size of model.ckpt-60000.meta is about 2.7M byte.

Pretraind data is File size of model.ckpt-0.data-00000-of-00001 size is 254M . File size of model.ckpt-0.meta size is 254M .

I think the problem is the model trained is not proper because of the model.ckpt-60000.data-00000-of-00001 size is twice bigger than pretrained file model.ckpt-0.data-00000-of-00001.

Or the problem is model.ckpt-60000.meta size is too small compared wit pretrained file model.ckpt-0.meta size is 254M.

How could I solve this problem?

[mistaker-likf]

My problem is same with you. Have you solve the issue?

[kiitosu]

I'm sorry. No I haven't.