NTIRE 2020 Image Deblurring Challenge - Track 2 on Smartphone Forum

Go back to competition Back to thread list Post in this thread

> Is it ok to use tensorflow 1.14.0 to train the network and generate the .tflite model instead of using tensorflow 2.0 ?

Due to the limit of the devices, I have to use tensorflow 1.14.0 to train and generate the .tflite model´╝îbut the organizers recommend to use tensorflow 2.0. Is it necessary to use tensorflow 2.0 to train the network and generate the .tflite model? Is it ok to use tensorflow 1.14.0? Thank you very much!

Posted by: Schoolyellow @ Jan. 9, 2020, 2:17 p.m.

Hello! Thank you for your interest in our competition.

Although we recommend the TensorFlow 2.0, you can use any TensorFlow version if you can prepare the final .tflite model.

The primary reason for the TensorFlow 2.0 is because our submission guideline (a simple tutorial for making the .tflite model) will be based on the 2.0 version.

Posted by: sanghyun @ Jan. 14, 2020, 1:33 p.m.

Thank you very much! Another question, how to evaluate the final results? That is, will the final images be generated by the submitted .tflite model by the organizers? Or will the final images be generated by ourselves and submitted to the website? If the evaluation criteria is the latter, how to avoid the participants submitting the final images generated by large models, while the submitted .tflite model is small?

Posted by: Schoolyellow @ Jan. 21, 2020, 1:28 p.m.

Final images should be generated using the submitted .tflite model only. While the participants will submit their result images individually, we will double-check them for a fair evaluation.

Posted by: sanghyun @ Jan. 22, 2020, 2:03 a.m.
Post in this thread