VictorMylle/Kohya-ss-sd-scripts
1
0
Fork 1
mirror of https://github.com/kohya-ss/sd-scripts.git synced 2026-04-18 01:30:02 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update train_util.py for add DAdaptAdan and DAdaptSGD

Browse Source
This commit is contained in:
青龍聖者@bdsqlsz
2023-04-26 16:15:33 +08:00
committed by GitHub
parent ac4935bf79
commit 0c9c90a87e
1 changed files with 58 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

58
library/train_util.py
View File

@@ -2563,7 +2563,65 @@ def get_optimizer(args, trainable_params):
optimizer_class = dadaptation.DAdaptAdam
optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs)
elif optimizer_type == "DAdaptAdan".lower():
try:
import dadaptation
except ImportError:
raise ImportError("No dadaptation / dadaptation がインストールされていないようです")
print(f"use D-Adaptation Adan optimizer | {optimizer_kwargs}")
actual_lr = lr
lr_count = 1
if type(trainable_params) == list and type(trainable_params[0]) == dict:
lrs = set()
actual_lr = trainable_params[0].get("lr", actual_lr)
for group in trainable_params:
lrs.add(group.get("lr", actual_lr))
lr_count = len(lrs)
if actual_lr <= 0.1:
print(
f"learning rate is too low. If using dadaptation, set learning rate around 1.0 / 学習率が低すぎるようです。1.0前後の値を指定してください: lr={actual_lr}"
)
print("recommend option: lr=1.0 / 推奨は1.0です")
if lr_count > 1:
print(
f"when multiple learning rates are specified with dadaptation (e.g. for Text Encoder and U-Net), only the first one will take effect / D-Adaptationで複数の学習率を指定した場合（Text EncoderとU-Netなど）、最初の学習率のみが有効になります: lr={actual_lr}"
)
optimizer_class = dadaptation.DAdaptAdan
optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs)
elif optimizer_type == "DAdaptSGD".lower():
try:
import dadaptation
except ImportError:
raise ImportError("No dadaptation / dadaptation がインストールされていないようです")
print(f"use D-Adaptation SGD optimizer | {optimizer_kwargs}")
actual_lr = lr
lr_count = 1
if type(trainable_params) == list and type(trainable_params[0]) == dict:
lrs = set()
actual_lr = trainable_params[0].get("lr", actual_lr)
for group in trainable_params:
lrs.add(group.get("lr", actual_lr))
lr_count = len(lrs)
if actual_lr <= 0.1:
print(
f"learning rate is too low. If using dadaptation, set learning rate around 1.0 / 学習率が低すぎるようです。1.0前後の値を指定してください: lr={actual_lr}"
)
print("recommend option: lr=1.0 / 推奨は1.0です")
if lr_count > 1:
print(
f"when multiple learning rates are specified with dadaptation (e.g. for Text Encoder and U-Net), only the first one will take effect / D-Adaptationで複数の学習率を指定した場合（Text EncoderとU-Netなど）、最初の学習率のみが有効になります: lr={actual_lr}"
)
optimizer_class = dadaptation.DAdaptSGD
optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs)
elif optimizer_type == "Adafactor".lower():
# 引数を確認して適宜補正する
if "relative_step" not in optimizer_kwargs: