forked from mrq/DL-Art-School
Notes for lucidrains converter.
This commit is contained in:
parent
b4720ea377
commit
c717765bcb
|
@ -1,3 +1,7 @@
|
|||
# Note this is an attempt to conform the lucidrains stylegan implementation to the official reference spec so that
|
||||
# I could use pretrained weights from NVIDIA. It is not currently successful, but that may be due to the weight
|
||||
# converter and not the code changes here. Use at your own risk.
|
||||
|
||||
import math
|
||||
import multiprocessing
|
||||
from contextlib import contextmanager, ExitStack
|
||||
|
|
|
@ -4,6 +4,12 @@
|
|||
#
|
||||
# Also doesn't require you to install Tensorflow 1.15 or clone the nVidia repo.
|
||||
|
||||
# THIS DOES NOT CURRENTLY WORK.
|
||||
# It does transfer all weights from the stylegan model to the lucidrains one, but does not produce correct results.
|
||||
# The rosinality script this was stolen from has some "odd" intracacies that may be at cause for this: for example
|
||||
# weight "flipping" in the conv layers which I do not understand. It may also be because I botched some of the mods
|
||||
# required to make the lucidrains implementation conformant. I'll (maybe) get back to this some day.
|
||||
|
||||
import argparse
|
||||
import os
|
||||
import sys
|
||||
|
|
Loading…
Reference in New Issue
Block a user