fujiki-1emon/illusorydnn

illusorydnn's Issues

Results for exp3

• illusory motion generation for illusion images for which human perceives strong illusory motion.

another illusory motion generation

• illusory motion generation for another illusory motion image [link].
  • the starting points for optical flow vectors in this image is represented as blue points.
    

reproduce again

• start from this repo
  • requirements are the same (before reproduced w/ chainer==5.4.0)

$ git clone https://github.com/eijwat/Predictive_Coding_DNN_TB

• generate training images

$ python generate_image.py Training_data.mp4 -d trainset -r 0

$ pip install cupy-cuda90
$ export LD_LIBRARY_PATH="/usr/local/cuda/lib64:$LD_LIBRARY_PATH"
$ pip --no-cache-dir install -I chainer==5.3.0
$ pip --no-cache-dir install -I cupy-cuda90

ImportError: Something is wrong with the numpy installation. While importing we detected an older version of numpy in ['/usr/local/lib/python2.7/dist-packages/numpy']. One method of fixing this is to repeatedly uninstall numpy until none is found, then reinstall this version.
$ rm -rf /usr/local/lib/python2.7/dist-packages/numpy*
$ pip install numpy

$ pip install --upgrade pip
$ pip --no-cache-dir install -I numpy
ERROR: chainer 5.3.0 has requirement cupy-cuda90<6.0.0,>=5.3.0, but you'll have cupy-cuda90 6.6.0 which is incompatible.
$ pip --no-cache-dir install -I cupy-cuda90
$ pip --no-cache-dir install -I chainer==5.3.0
RuntimeError: CUDA environment is not correctly set up
(see https://github.com/chainer/chainer#installation).cannot import name _is_fusing

LD_LIBRARY_PATH="/usr/local/cuda/lib64:$LD_LIBRARY_PATH" pip --no-cache-dir install -IU chainer==5.3.0
CUDA_VISIBLE_DEVICES=6 python PredNet/main.py -i trainset/train_list.txt -g 0

note - about the current reproduction

• Tesla V100
• chainer==5.4.0
• cupy-cuda101==5.4.0
• used trainset
  • logs/20191212120406
  • logs/20191212125800
  • logs/20191212125922
  • logs/20191212131222
  • logs/20191212131227
  • logs/20191212131439
  • logs/20191212131804
  • Predictive_Coding_DNN_TB/models

# for Predictive_Coding_DNN_TB/models
# snakes/illusionSingleL.list is list of 21 still images
CUDA_VISIBLE_DEVICES=3 python2 PredNet/main.py --test -g 0 --initmodel models/400000.model --images snakes/illusionSingleL.list --input_len 19 --ext 2

• ./logs/20191213055816 grayscale input
• ./logs/20191215013528.1of10 chainer==5.3.0, cupy==5.4.0?

is "illusory motion" reproduced in DNN really an illusory motion?

"illusory motion" in color images

• Red optical flow vectors corresponds to the illusory motion perceived in human with yellow points as their starting points.
• top to bottom: non-illusion, crockwise-illusory motion, counter-crockwise-illusory motion
• left to right: original images, 2x "illusory motion" visualized images

"illusory motion" in grayscale images