mistralk/srcnn-rnn

srcnn with rnn tensorflow implementation

srcnn-rnn

• SRCNN with Vanilla RNN tensorflow implementation
• Hanyang University ITE4053 Deep Learning Methods and Applications assignment #2

Requirements

• Python 3.6
• TensorFlow 1.12
• NumPy
• imageio
• Pillow

Assignment detail

• 32x32x1 input
• 32x32x1 output
• 3x3 kernel
• number of epoch=5000
• mini batch size=128
• Recurrent structure (Vanilla RNN)

Trained models

Model A

• Downscaling interpolation to half size (preprocess): AREA
• ADAM optimizer setting: learning rate=0.003, β1=0.9, β2=0.999, ε=1E−08

Model B

• Downscaling interpolation (preprocess): NEAREST_NEIGHBOR
• ADAM optimizer setting: learning rate=0.0001, β1=0.9, β2=0.999, ε=1E−08

Model C

• Downscaling interpolation (preprocess): BICUBIC
• ADAM optimizer setting: learning rate=0.0001, β1=0.9, β2=0.999, ε=1E−08

Result (Model A)

Low-resolution (BICUBIC)

SRCNN-RNN result

Metrics (but something wrong...)

References

• Original paper
  • Image Super-Resolution Using Deep Convolutional Networks
  • Deep Learning for Single Image Super-Resolution: A Brief Review
• My program architecture is inspired by some good tutorials and examples
  • https://github.com/cs230-stanford/cs230-code-examples
  • https://cs230-stanford.github.io/tensorflow-input-data.html
  • https://www.tensorflow.org/guide/datasets
  • https://docs.python-guide.org/writing/structure/

Problems and solutions

Python 3.7, TensorFlow 1.13에서 tf.concat 사용시 에러 발생

AbortedError (see above for traceback): Operation received an exception:Status: 5, message: could not create a view primitive descriptor, in file tensorflow/core/kernels/mkl_slice_op.cc:435
         [[node gradients/concat_2_grad/Slice_1 (defined at C:\srcnn-rnn\framework\model.py:70) ]]

• 모델 구축도 다 잘 되는데 sess.run할때 뜬금없이 tf.concat 관련하여 에러가 발생
• 이미 보고된 버그이지만 비교적 사례가 적은데, 인텔 mkl에 관계된 것으로 보임 tensorflow/tensorflow#17494
• 그런데 또 노트북에서는 잘 돌아감을 확인. AMD CPU 문제?
• 텐서플로우 GPU 버전에서는 확인해보지 않음
• 웹상에 공개된 (아마도) 유일한 해결책은 Python 3.6, Tensorflow 1.12로 버전을 낮추는 것 https://blog.csdn.net/Bing_bing_bing_/article/details/90211188

TensorBoard 이미지 슬라이더 촘촘하게 만들기

tensorboard --logdir=logs/ --samples_per_plugin images=100