Bunch of corrections

[szopeno]

mv ../shape_predictor_68_face_landmarks.dat.bz2 ../ffhq.pkl ../model_ir_se50.pth ../faces_w_encoder.pt checkpoints/

FIRST correction required to work

mv data_preprcoess/ data_preprocess # (!!)

SETTING PYTHONPATH required to work

export PYTHONPATH=$PYTHONPATH:pwd

this fails, dlib model is in bz2 archive, must be bunzipped

python data_preprocess/0_align_face.py
--images_dir test_album
--save_dir test
--trash_dir trash
--landmarks_model checkpoints/shape_predictor_68_face_landmarks.dat.bz2
--id_model checkpoints/model_ir_se50.pth

Traceback (most recent call last):
File "PATH-EDITED/mystyle/data_preprocess/0_align_face.py", line 242, in
main()
File "PATH-EDITED/mystyle/data_preprocess/0_align_face.py", line 219, in main
lnds_predictor = dlib.shape_predictor(str(args.landmarks_model))
RuntimeError: Unexpected version found while deserializing dlib::shape_predictor.

cd checkpoints
bunzip2 shape_predictor_68_face_landmarks.dat.bz2 # (!)
cd ..
python data_preprocess/0_align_face.py
--images_dir test_album
--save_dir test
--trash_dir trash
--landmarks_model checkpoints/shape_predictor_68_face_landmarks.dat
--id_model checkpoints/model_ir_se50.pth

OK - now it fails a bit further, but it's because I have n physical GPU no board. I managed to run arcnet/resnet withut GPU before, so I will try later to check whether I'll be able to do it here. We'll see.

[szopeno]

Corrected it to work with CPU only. Damn it's slow without GPU :-) I have a question:

what should be the input image resolution for superresolution? It seems that images have to be 1024x1024; With images with smaller resolution, I am getting errors like this (path edited)

Traceback (most recent call last):
File "project.py", line 120, in
main()
File "project.py", line 116, in main
project(args)
File "project.py", line 60, in project
sample = alpha_projector.reconstruct(sample)
File "mystyle/reconstruct/base_projector.py", line 66, in reconstruct
recon_loss = self.reconstruction_loss(synth, target, mask=sample.mask)
File "mystyle/reconstruct/base_projector.py", line 52, in reconstruction_loss
return super(BaseProjector, self).reconstruction_loss(degraded_synth, target)
File "mystyle/reconstruct/base_reconstructor.py", line 50, in reconstruction_loss
dist = self.lpips(synth, target)
File "HOME.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
return forward_call(*input, **kwargs)
File "HOME/.local/lib/python3.6/site-packages/lpips/lpips.py", line 124, in forward
diffs[kk] = (feats0[kk]-feats1[kk])**2 = (feats0[kk]-feats1[kk])**2
RuntimeError: The size of tensor a (255) must match the size of tensor b (127) at non-singleton dimension 3

[luchaoqi]

Corrected it to work with CPU only. Damn it's slow without GPU :-) I have a question:

what should be the input image resolution for superresolution? It seems that images have to be 1024x1024; With images with smaller resolution, I am getting errors like this (path edited)

Traceback (most recent call last): File "project.py", line 120, in main() File "project.py", line 116, in main project(args) File "project.py", line 60, in project sample = alpha_projector.reconstruct(sample) File "mystyle/reconstruct/base_projector.py", line 66, in reconstruct recon_loss = self.reconstruction_loss(synth, target, mask=sample.mask) File "mystyle/reconstruct/base_projector.py", line 52, in reconstruction_loss return super(BaseProjector, self).reconstruction_loss(degraded_synth, target) File "mystyle/reconstruct/base_reconstructor.py", line 50, in reconstruction_loss dist = self.lpips(synth, target) File "HOME.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl return forward_call(*input, **kwargs) File "HOME/.local/lib/python3.6/site-packages/lpips/lpips.py", line 124, in forward diffs[kk] = (feats0[kk]-feats1[kk])**2 = (feats0[kk]-feats1[kk])**2 RuntimeError: The size of tensor a (255) must match the size of tensor b (127) at non-singleton dimension 3

Same here, not sure if there are any requirements about image resolution