Always get 'Could not find a face that downscales correctly within epsilon' after running run.py about pulse HOT 7 OPEN

The above change adds a new argument eps to run.py such that if the L2 loss goes below eps, the program returns. If at the end of steps steps, the L2 loss is not less than eps the run fails with that error. Thus another fix is to choose a value for eps that is less than your L2 loss at the end.

After 100 steps I had an L2 loss of 0.006 (6e-3) so I set my eps to 8e-3 with the following colab code, and it worked. The step after this is to lower epsilon and increase steps to get a better result. The best option would be reverting though as we generally just want to see the best result achieved.

seed = 100 #@param {type:"integer"}
epsilon = 8e-3
noise_type = 'trainable'  #@param ['zero', 'fixed', 'trainable']
optimizer = 'adam'  #@param ['sgd', 'adam','sgdm', 'adamax']
learning_rate = 0.4 #@param {type:"slider", min:0, max:1, step:0.05}
learning_rate_schedule = 'linear1cycledrop'  #@param ['fixed', 'linear1cycle', 'linear1cycledrop']
steps = 750 #@param {type:"slider", min:100, max:1000, step:50}
clear_output()

seed = abs(seed)
print('Estimated Runtime: {}s.\n'.format(round(0.23*steps)+6))
!python run.py \
  -seed $seed \
  -noise_type $noise_type \
  -opt_name $optimizer \
  -learning_rate $learning_rate \
  -steps $steps \
  -lr_schedule $learning_rate_schedule \
  -eps $epsilon

from pulse.

Increasing the number of steps would solve the issue.

from pulse.

Not sure if this is the recommended way to fix this, but I had the same problem with a 32x32 image. I upscaled it to 1024x1024 (bicubic, so the result is blurry), and it seemed to work for me.

from pulse.

I got the same error with all the images I tried. But only one of them worked using @udf 's method. Thanks anyway! :)

from pulse.

It seems to have been introduced by this commit ece9211
I ran a test on Face-Depixelizer with it reverted, it worked.

from pulse.

Hi @udf , how did you upscaled your images? I have my input in 128x128, and I was looking to upscale it to 1024x1024. Can you explain a bit?

from pulse.

hi...can anyone please help me know how the downsampling loss is implemented. Is it l1 or l2 loss being used to compare the generated low-resolution and the LR image.
please any1 let me know

from pulse.