search

hellochick / PSPNet-tensorflow

TensorFlow-based implementation of "Pyramid Scene Parsing Network".
326 stars 123 forks source link

inference result with model trained with train.py is not good and model size is almost twise bigger then pretrained model #62

Open kiitosu opened 5 years ago

kiitosu commented 5 years ago

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 commented 4 years ago

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

kiitosu commented 4 years ago

I'm sorry. No I haven't.