search

Jongchan / tensorflow-vdsr

A tensorflow implementation of "Accurate Image Super-Resolution Using Very Deep Convolutional Networks", CVPR 16'
255 stars 125 forks source link

Test and train errors #17

Open haihuanzxx opened 7 years ago

haihuanzxx commented 7 years ago

I test your uploaded model ''VDSR_adam_epoch_016.ckpt-57188'' by 'TEST.py', but the following error occurs "NotFoundErrors: Unsucessful TensorSliceReader construcator: Failed to find any matching files for VDSR_adam_epoch_016.ckpt-57188'" I retrain the network by”VDSR.py" and test the all the ckpts, the best psnr is 29.10 which id much samller than your results(29.67). my VDSR_adam_epoch_016.ckpt-57188 is 29.08 my tensorflow version is 1.0. I am confused about the above questions.

haihuanzxx commented 7 years ago

supplement： the enlargement times is 3, the test dataset is Set14

Jongchan commented 7 years ago

Sorry I got here so late. Do you still have this problem? Training results are varying from time to time, so my result is the best model out of many trials.

haihuanzxx commented 7 years ago

the result of my best model（implementation according to your codes） has so great distance in comparsion with yours

xiaobaiyang-programer commented 7 years ago

@haihuanzxx I alse meet the same problem to test my model by 'TEST.py', what do you do to fix the problem?

Jongchan commented 7 years ago

There are 2 suggestions. 1) I suspect that data augmentation may be different. I used the uploaded data augmentation code written in matlab. Note that bicubic interpolation algorithm may be different between libraries, and Matlab's bicubic interpolation worked better than python PIL's bicubic interpolation.

2) If you used the code that I've uploaded, I can only suspect that the training result is not stable. It can change from time to time... Know that the best model may not be exactly at epoch 16. I can only say that I got my results using the code I provided in the current repository, without finetuning or extra engineering.

2017develper commented 7 years ago

hello , please i want to add a layer in network can you help me please

Jongchan commented 7 years ago

@2017develper Hi, you can refer to MODEL.py for model definition. When you train the newly defined model (of your own) from scratch, use python VDSR.py to start.

wangWilson commented 6 years ago

@haihuanzxx After retraining, the error that" Failed to find any matching files for VDSR_adam_epoch_016.ckpt-57188'"has been resolved?

DawyD commented 6 years ago

@wangWilson Hi, the problem you are having comes from the fact, that the API of TensorFlow changes over time. The code was written for some older version of TensorFlow (<1.0). For example, if you have trained your own model on TensorFlow 1.3, it is stored in checkpoint folder in a different way than it is expected by TEST.py. A dirty and simple way to overcome this issue is to modify the lines 79 and 80 of TEST.py to the following:

model_list = glob.glob("./checkpoints/VDSR_adam_epoch_*.data-00000-of-00001")
model_list = sorted([fn[:-20] for fn in model_list])